DRAFT – Not for Distribution

Executive Summary

Background

A key priority of the United States Patent and Trademark Office (USPTO) is to maintain

United States leadership in innovation, especially in emerging technologies, including

artificial intelligence (AI). To further this goal, the USPTO has been actively engaging with

the innovation community and experts in AI to promote the understanding and reliability of

intellectual property (IP) rights in relation to AI technology. Additionally, the USPTO is

working to ensure that appropriate IP incentives are in place to encourage further

innovation in and around this critical area. To this end, in January 2019, the USPTO held an

AI IP policy conference, one of the first of its kind. The conference featured IP specialists from

around the world and included panel discussions on patents, trade secrets, copyrights,

trademarks, IP enforcement, global perspectives, and the economics of IP protection of AI.

Building on the momentum of those

discussions, on August 27, 2019, the

USPTO issued a request for comments

(RFC) on patenting AI inventions. The

RFC sought feedback from our

stakeholders on a variety of patent policy

issues, such as AI’s impact on

inventorship and ownership, eligibility,

disclosure, and the level of ordinary skill

in the art. The comment period closed on

November 8, 2019. The USPTO received

99 comments from a wide range of

stakeholders, including individuals,

associations, corporations, and foreign IP

offices. (See Table 1.)

The full recordings of the conference may be viewed at: https://www.uspto.gov/about-us/events/artificial-

intelligence-intellectual-property-policy-considerations.

Category of Responses to Aug. 27,

2019 RFC (Patents)

No. of

submissions

Foreign patent offices

Bar associations

Trade associations/Advocacy groups

Companies

Academia

Law firms (submitted as firm)

Practitioners (other than firm or

academia submissions)

Individuals (not in other categories)

Total

Table 1

On October 30, 2019, the USPTO issued

a second RFC related to the impact of AI

on other IP policy areas, including

copyrights, trademarks, database

protections, and trade secret law. The

comment period for the second RFC

closed on January 10, 2020. The USPTO

received 98 comments from a wide

range of stakeholders, including

individuals, associations, and

corporations. (See Table 2.)

Report and Public Comment Themes

Following the conclusion of the comment periods, a team of experts assembled from across

the USPTO to examine the responses and generate the following report. The report is divided

into two parts. Part I focuses on the first RFC solicitation dedicated to patenting of artificial

intelligence technologies and provides AI context, legal background, and public comment

synthesis, as appropriate, for each of the patent RFC questions. Part II follows a similar

format for the second IP RFC solicitation dedicated to non-patent intellectual property

protections for artificial intelligence technologies, such as trademark, copyright, and trade

secret.

From the synthesis of the public comments, a number of themes emerged:

General Themes

• Many comments addressed the fact that AI has no universally recognized definition.

Due to the wide-ranging definitions of the term, often comments urged caution with

respect to specific IP policymaking in relation to AI.

• The majority of public commenters, while not offering definitions of AI, agreed that

the current state of the art is limited to “narrow” AI. Narrow AI systems are those that

perform individual tasks in well-defined domains (e.g., image recognition, translation,

etc.). The majority viewed the concept of artificial general intelligence (AGI)—

intelligence akin to that possessed by humankind and beyond—as merely a

theoretical possibility that could arise in a distant future.

• Based on the majority view that AGI has not yet arrived, the majority of comments

suggested that current AI could neither invent nor author without human

intervention. The comments suggested that human beings remain integral to the

Category of Responses to Oct. 30, 2019

RFC (other IP)

No. of

submissions

Bar associations

Trade associations/Advocacy groups

Companies

Academia

Practitioners

Individuals (not in other categories)

Total

Table 2

iii

operation of AI, and this is an important consideration in evaluating whether IP law

needs modification in view of the current state of AI technology.

• Across all IP topics, a majority of public commenters expressed a general sense that

the existing U.S. intellectual property laws are calibrated correctly to address the

evolution of AI. However, commenters appear split as to whether any new classes of

IP rights would be beneficial to ensure a more robust IP system.

Patent Themes

• A majority of commenters agreed that AI is viewed best as a subset of computer-

implemented inventions. Therefore, this majority felt that current USPTO guidance,

especially on patent subject matter eligibility and disclosure of computer-

implemented inventions, is equipped to handle advances in AI. However, some

commenters stressed that it may be difficult to enable (i.e., teach the public to make

and use) certain AI inventions, as required by 35 U.S.C. § 112(a), and offered the topic

for further exploration by the USPTO.

• Most public commenters agreed that the growing ubiquity of AI would affect how the

USPTO and courts would assess the legal hypothetical standard of a “person having

ordinary skill in the art,” this standard being critical to the determination of whether

a patent right should issue.

• While no majority coalesced around AI’s impact on prior art (i.e., the body of

knowledge known at the time a patent application is filed), a number of issues were

referred to the USPTO for further consideration, including that AI may generate a

proliferation of prior art amounting to a never before seen volume and the ensuing

difficulty in finding relevant prior art in view of the increased volume.

Other IP Themes

• Again, while a majority of commenters stated that current IP laws are calibrated

correctly in the copyright, trademarks, trade secrets, and data fields, many agreed

that existing commercial law principles might adequately fill any gaps left by IP law

in the wake of advances in AI (e.g., contract law).

• Specifically on trademarks, most commenters agreed that AI would improve

efficiency of examination of trademark applications. Although this sentiment was also

generally shared in regard to patent examination.

• Many comments expressed that the use of copyrighted material to “train” AI may

violate the reproduction right of a copyright owner under 17 U.S.C. §106(1), and that

this use may or may not be a non-infringing fair use.

• Most commenters found that existing fair use law does not require modification, as

fair use is a flexible doctrine and is capable of adapting to the use of copyrighted

works in the context of AI.

• The topics of trade secrets and data issues generated an expansive range of

comments, touching on issues of bias, transparency, privacy, and debates over

whether advances in AI warrant a sui generis IP system for data rights.

The USPTO will use this report to focus issues for continued exploration of other measures

it may take to bolster the understanding and reliability of IP rights for emerging

technologies, such as AI. These steps may include further engagement with the public,

additional guidance for stakeholders, and continued training for examiners on emerging

technologies.

Corrigendum: As originally published, footnote 29 relied on four public submissions. After

initial report publication, one such submitter requested that their public submission not be

relied on. This subsequent publication reflects that request.

Disclaimer: The USPTO appreciates the public’s feedback and engagement on issues related to

AI technology. The agency has considered all the comments and has included a summary of the

comments in this report. The full comments may be viewed at

https://www.uspto.gov/initiatives/artificial-intelligence

. The views, thoughts, and opinions

expressed in the comments do not necessarily state or reflect those of the USPTO, the

administration, or any other federal government entity. Reference herein to a comment made

by any specific entity does not constitute or imply its endorsement, recommendation, or

favoring by the USPTO, the administration, or any other federal government entity.

PART I—Responses to the RFC on Patenting Artificial Intelligence Inventions,

issued on August 27, 2019

A summary of the comments received in response to the RFC on Patenting AI Inventions

issued on August 27, 2019, is included below, organized by the question appearing in the

RFC. Commenters included foreign patent offices, bar associations, industry associations,

academia, and various stakeholders, both national and international. Representatives from

electronics, software, automobile, medical, and pharmaceutical industries responded to the

RFC.

1. What are elements of an AI invention? For example: The problem to be addressed

(e.g., application of AI); the structure of the database on which the AI will be

trained and will act; the training of the algorithm on the data; the algorithm

itself; the results of the AI invention through an automated process; the

policies/weights to be applied to the data that affects the outcome of the results;

and/or other elements.

This question sought to identify broadly the elements of an AI invention that may be subject

to patentability.

Among the responses, four common answers arose:

(1) The various elements disclosed in the question constitute a non-exclusive list of

elements of an AI invention.

(2) AI can be understood as computer functionality that mimics cognitive functions

associated with the human mind (e.g., the ability to learn).

(3) AI inventions can be categorized (in no particular order) as follows:

(a) inventions that embody an advance in the field of AI (e.g., a new neural

network structure of an improved machine learning (ML) model or algorithm)

(b) inventions that apply AI (to a field other than AI)

Response from AIPPI, at 2; Response from EPSON, at 2; Response from IBM (Nov. 8, 2019),

at 2; Response from JEITA, at 2.

Response from IBM (Nov. 8, 2019), at 2; Response from Juniper Networks, at 1; Response

from Schwegman Lundberg & Woessner, P. A., at 1.

Response from CCIA (1st Response), at 1; Response from Ericsson, at 2; Response from

Internet Association, High Tech Inventors Alliance, the Software and Information Industry

Association, and ACT, at 9; Response from IPO (Nov. 11, 2019), at 3; Response from JPAA, at

1-2; Response from Siemens, at 1-2; Response from AIPPI Japan, at 1-2; Response from JPMA,

at 1-2; Response from JPO, at 1; Response from Merck, at 2.

(4) Undue effort should not be expended on defining AI, which is dynamic and will be

subject to fundamental change in the coming years.

2. What are the different ways that a natural person can contribute to conception

of an AI invention and be eligible to be a named inventor? For example: Designing

the algorithm and/or weighting adaptations; structuring the data on which the

algorithm runs; running the AI algorithm on the data and obtaining the results.

As with other fields of technology, the development of AI may present many opportunities

for invention. For example, designing an AI algorithm, implementing particular hardware to

enhance an AI algorithm, or applying methods of preparing inputs to an AI algorithm may

present patent considerations. Many innovators may also be involved in the development of

an AI system. Provided with the potential range of innovation and the possibility that more

than one person may be involved in the development of an AI system, the law requires that

a determination be made as to who has legally contributed to the conception of an AI

invention and can be named as an inventor.

35 U.S.C. § 100 defines “inventor” as “the individual or, if a joint invention, the individuals

collectively who invented or discovered the subject matter of the invention.”

Moreover, 35

U.S.C. § 116 provides that an invention may be made by two or more persons jointly even

though “(1) they did not physically work together or at the same time, (2) each did not make

the same type or amount of contribution, or (3) each did not make a contribution to the

subject matter of every claim of the patent.”

The Federal Circuit has made clear that “conception” is the touchstone of inventorship.

Conception requires an inventor to have a specific solution to a problem rather than a

general goal for success.

Conception is finished “only when the idea is so clearly defined in

the inventor’s mind that only ordinary skill would be necessary to reduce the invention to

practice, without extensive research or experimentation.”

Similarly, to be a joint inventor,

one must: “(1) contribute in some significant manner to the conception or reduction to

practice of the invention, (2) make a contribution to the claimed invention that is not

3a-3c: Response from CCIA (1st Response), at 1; Response from Ericsson, at 2; Response

from FICPI, at 2-3; Response from Internet Association, High Tech Inventors Alliance, the

Software and Information Industry Association, and ACT, at 9-10; Response from IPO (Nov.

11, 2019), at 3; Response from JPAA, at 1-2; Response from Siemens, at 1-2.

Response from Ericsson, at 2; Response from EPSON, at 2; Response from Novartis, at 3;

Response from NSIP Law, at 4; Response from Schwegman Lundberg & Woessner, at 3.

35 U.S.C. § 100 (2018).

Burroughs Wellcome Co. v. Barr Labs., Inc., 40 F.3d 1223, 1227-28 (Fed. Cir.

1994) (citations omitted); see also In re Verhoef, 888 F.3d 1362, 1366 (Fed. Cir. 2018).

Id.

insignificant in quality, when that contribution is measured against the dimension of the full

invention, and (3) do more than merely explain to the real inventors well-known concepts

and/or the current state of the art.”

The vast majority of public commenters asserted that current inventorship law is equipped

to handle inventorship of AI technologies.

One commenter went as far as to state that

“there is no urgency to revise the law with respect to inventorship.”

Many of these

commenters suggested that assessment of conception should be fact-specific, as in the

analysis done today.

For example, one commenter stressed that there are different ways in

which a natural person may contribute to the conception of an invention and that each

contribution “should be evaluated on a case-by-case basis,” as is the law today.

A related

view was that a data scientist carrying out the task of building and testing a use of an AI

technology invention is doing nothing more than reducing the invention to practice.

In the

words of one commenter, “running [an] AI algorithm on the data and obtaining the results is

unlikely to qualify as a contribution [to conception].”

3. Do current patent laws and regulations regarding inventorship need to be

revised to take into account inventions where an entity or entities other than a

natural person contributed to the conception of an invention?

AI provides unique policy considerations stemming from its potential for autonomous

creation. Present AI technology appears to be within the realm of narrow, application-

specific objectives, but the notion of artificial general intelligence (AGI)—intelligence akin to

In re Verhoef, 888 F.3d at 1366 (Fed. Cir. 2018).

See, e.g., Response from IPO (Nov. 11, 2019), at 4 (“[T]here is nothing unique about how a

natural person contributes to the conception of an AI-related invention versus any other

highly technical field.”); Response from NAPP, at 1 (“an AI invention should be determined

in the same way as for other kinds of inventions …”).

Response from IBM (Nov. 8, 2019), at 5; see also Response from SIIA (Nov. 8, 2019), at 5

(indicating that the USPTO has all the tools it needs under the current statutory framework).

See, e.g., Response from AAIH, at 3 (suggesting that current law is a fact-specific analysis);

Response from BPLA, at 3-4 (noting that the use of AI in the inventive process does not

negate inventorship by a natural person); Response from FICPI, at 3 (noting the fact-specific

nature of the inquiry).

Response from AIPLA (Nov. 8, 2019), at 3; see also Response from IEEE-USA, at 4 (“The

ways that a natural person can contribute to conception of an AI invention are either the

same as or analogous to the ways that a natural person can contribute to conception of an

invention in computer-implemented technology”); Response from Novartis, at 4 (Whether

AI inventions rise to conception “depend[s] on the facts of a given case and situation …”).

See, e.g., Response from Maughan, at 2.

Response from ABA IPL (Nov. 8, 2019), at 11; see also Response from Edward Ryan, at 2

(“one should not be able to simply push a button and be named an inventor.”); Response

from RF SUNY, at 1 (“simply running the AI algorithm on data and obtaining results may not

constitute a meaningfully creative or inventive contribution …”).

that possessed by humankind and beyond—is worthy of consideration. Thus, the instant

question also contemplates a future state in which the capability of AI to invent approaches

or exceeds that of human intelligence.

As previously discussed under question two, conception is the formation in the mind of the

inventor of a definite and permanent idea of the complete and operative invention. As stated

above, an “inventor” is defined in 35 U.S.C. § 100(a) as “the individual or, if a joint invention,

the individuals collectively who invented or discovered the subject matter of the

invention.”

Title 35 of the United States Code is replete with language indicating that the

inventor of a patent application must be a natural person. For example, 35 U.S.C. § 101 states,

“Whoever invents or discovers any new and useful process, machine, manufacture, or

composition of matter … may obtain a patent therefore, subject to the conditions and

requirements of this title” (emphasis added). “Whoever” denotes whatever person, a person

being a human being—a natural person.

By the use of “whoever,” § 101 limits patent

protection to inventions and discoveries by natural persons.

35 U.S.C. § 115 provides additional clarification that the inventor must be a natural person.

That is, § 115 uses pronouns specific to natural persons—“himself” and “herself”—when

referring to the “individual” who believes himself or herself to be the original inventor or an

original joint inventor of a claimed invention in the application,

and states that the inventor

who executes an oath or declaration must be a “person.”

In fact, there are numerous other

patent statutes that refer to the inventor as a “person.”

The USPTO’s understanding of the

patent statutes and the Federal Circuit case law concerning the concept that inventorship

requires that an inventor must be a natural person is reflected in the numerous references

to the inventor as a “person” in Title 37 of the Code of Federal Regulations.

See also 35 U.S.C. § 115(a) (“each individual who is [an] inventor … shall execute an oath

or declaration”); 35 U.S.C. § 100(g) (“The terms ‘joint inventor’ and ‘coinventor’ mean any 1

of the individuals who invented or discovered the subject matter of a joint invention … ”).

Merriam-Webster.com, https://www.merriam-webster.com/dictionary/whoever (last

accessed Apr. 6, 2020).

35 U.S.C. § 115(b)(2) (“An oath or declaration under subsection (a) shall contain

statements that … such individual believes himself or herself to be the original inventor or

an original joint inventor of a claimed invention in the application.”).

35 U.S.C. § 115(h)(1) (“Any person making a statement required under this section may

withdraw, replace, or otherwise correct the statement at any time.”).

See, e.g., 35 U.S.C. § 102(a) (“A person shall be entitled to a patent unless … ”); 35 U.S.C. §

116(c) (“Whenever through error a person is named in an application for patent as the

inventor … .”); 35 U.S.C. § 185 (“Notwithstanding any other provisions of law any person, and

his successors, assigns, or legal representatives, shall not receive a United States patent for

an invention if that person, or his … ”); 35 U.S.C. § 256(a) (“Whenever through error a person

is named in an issued patent as the inventor … .”).

See, e.g., 37 CFR 1.27(a)(1) (“A person, as used in paragraph (c) of this section, means any

inventor or other individual”); 37 CFR 1.41(d) (“ . . . the name and residence of each person

The use of an AI system as a tool by a natural person(s) does not generally preclude a natural

person(s) from qualifying as an inventor (or joint inventors) if the natural person(s)

contributed to the conception of the claimed invention. That is, the activities by a natural

person(s) that would ordinarily qualify as a contribution to the conception of an invention

are unaffected by the fact that an AI system is used as a tool in the development of the

invention. For example, depending on the specific facts of each case, activities such as

designing the architecture of the AI system, choosing the specific data to provide to the AI

system, developing the algorithm to permit the AI system to process that data, and other

activities not expressly listed here may be adequate to qualify as a contribution to the

conception of the invention.

The majority of commenters responding to this question reflected the view that there is no

need for revising patent laws and regulations on inventorship to account for inventions in

which an entity or entities other than a natural person contributed to the conception of an

invention.

One commenter remarked that “conception is inherently a human activity … an

entity or entities other than a natural person cannot contribute to the conception of an

invention.”

Many comments took issue with the question’s premise that under the state of

the art, a machine could conceive of an invention. As one commenter put it, “the current state

of AI technology is not sufficiently advanced at this time and in the foreseeable future so as

to completely exclude the role of a human inventor in the development of AI inventions.”

believed to be an actual inventor should be provided when the application papers pursuant

to § 1.53(b) or § 1.53(c) are filed.”); 37 CFR 1.53(d)(4) (“accompanied by a statement

requesting deletion of the name or names of the person or persons who are not inventors of

the invention being claimed in the new application”); 37 CFR 1.63(a)(3) (“An oath or

declaration under this section must: Include a statement that the person executing the oath

or declaration believes . . . ."); 37 CFR 1.324(b)(1) (“A statement from each person who is

being added as an inventor and each person who is currently named as an inventor . . . .”).

Note, also, the requirement under 7 CFR 1.76(b)(1) that the inventor be identified by their

“legal name.”

Response from Abadi, at 2; Response from ABA IPL (Nov. 8, 2019), at 12; Response from

AIPLA (Nov. 8, 2019), at 4; Response from AIPPI Japan, at 4; Response from the BADC, at 4;

Response from BPLA, at 3-4; Response from Ericsson, at 3; Response from Internet

Association, High Tech Inventors Alliance, the Software and Information Industry

Association, and ACT, at 10-11; Response from IBM, at 5; Response from JEITA, at 3-4;

Response from JPAA, at 3; Response from Juniper Networks, at 3; Response from Gaudry, at

2; Response from Rubin, at 5; Response from Merck, at 3; Response from NAPP, at 2;

Response from NSIP, at 4; Response from Kumar, at 2; Response from Davis, at 4-5; Response

from R Street Institute (Nov. 8, 2019), at 2-3; Response from Zubek, at 1; Response from

Naimpally, at 1; Response from Schwegman Lundberg & Woessner, at 5.

Response from BADC, at 4.

Response from AIPPI, at 5.

Others characterized modern-day AI as a tool to aid natural persons in the inventive

process.

Some commenters suggested that the USPTO should revisit the question when machines

begin achieving AGI (i.e., when science agrees that machines can “think” on their own).

minority of commenters suggested that AGI was a present reality that needed to be

addressed today.

Others warned that if such a change was made to recognize non-natural

person inventors, the USPTO should carefully consider the practical effects of such a change:

How would a continuation be treated? How would a machine sign an oath or declaration?

Would a flood of applications ensue? Would certain types of AI dominate technology

development in the future?

4. Should an entity or entities other than a natural person, or company to which a

natural person assigns an invention, be able to own a patent on the AI invention?

For example: Should a company who trains the artificial intelligence process that

creates the invention be able to be an owner?

Ownership of a patent entitles the patent owner the right to exclude others from making,

using, offering for sale, selling, or importing into the United States the invention claimed in

the patent.

For applications filed on or after September 16, 2012, the original applicant is

See, e.g., Response from FICPI, at 3 (“the AI system being used should be considered a ‘tool’

in the inventing process.”); Response from Brindisi, at 4 (“AIs of our era are still tools devised,

applied and exploited by humans.”) (emphasis in original); Response from Ford, at 1 (“AI-

created inventions are the product of a tool that facilitates discovery by the true inventor …

”).

Response from Davis, at 5 (“Looking far (far!) ahead, programs someday may begin to

learn on their own … Then we may have a deeper quandary.”); Response from Gaudry, at 2

(“general [artificial] intelligence is a very long ways off, such that we need not worry about

adjusting patent law now for this distant and remote possibility.”); Response from Michael

Murial and Andrew Noble, at 8 (“Unless and until the scientific community declares that AI

has allowed computers to achieve ‘consciousness’ such that a computer is capable of

‘conceiving’ something, any question about what to do when an AI system ‘invents’

something is purely hypothetical.”).

Response from RF SUNY, at 2-3; Response from Sanker, at 1; Response from Siemens, at 2.

Response from Askeladden, at 4 (questioning the Constitutional authority to recognize AI

as an inventor); Response from JEITA, at 3-4 (explaining an influx of applications may result

from recognizing AI as an inventor); Response from EPSON, at 2-3 (expressing concerns over

an influx of applications, “chaos in the business community,” and the broader legal

personality of machine issues); Response from IPO (Nov. 11, 2019), at 6 (raising practical

concerns, such as how one would depose a machine); Response from JIPA (Nov. 6, 2019), at

2-3 (expressing broad practical concerns, such as those about legal rights normally reserved

for natural persons being vested in machines); Response from Tata Consultancy, at 2-3

(raising practical concerns, such as execution of documents by a machine, effects on

continuing applications, and assignment of rights).

35 U.S.C. § 154(a)(1).

presumed to be the owner of an application for an original patent unless there is an

assignment.

For applications filed before September 16, 2012, the ownership of the patent

(or the application for the patent) initially vests in the named inventors of the invention on

the patent.

A patent or patent application is assignable by an instrument in writing, and the

assignment of the patent, or patent application, transfers to the assignee(s) an alienable

(transferable) ownership interest in the patent or application.

The vast majority of commenters stated that no changes should be necessary to the current

U.S. law—that only a natural person or a company, via assignment, should be considered the

owner of a patent or an invention.

However, a minority of responses stated that while

inventorship and ownership rights should not be extended to machines, consideration

should be given to expanding ownership to a natural person: (1) who trains an AI process,

or (2) who owns/controls an AI system.

5. Are there any patent eligibility considerations unique to AI inventions?

In assessing the patent eligibility of AI inventions, all judicially created exceptions to the

statutory categories are relevant (i.e., laws of nature, natural phenomena, and abstract

ideas). In January 2019, the USPTO issued the 2019 Revised Patent Subject Matter Eligibility

See 37 C.F.R. § 3.73(a).

See Beech Aircraft Corp. v. Edo Corp., 990 F.2d 1237, 1248 (Fed. Cir. 1993).

35 U.S.C. § 261.

Response from ABA IPL (Nov. 8, 2019), at 12-13; Response from AIPLA (Nov. 8, 2019), at

4 (“Ownership of patent rights should remain reserved for only natural or juridical persons

at this time. Changing the ownership regime to allow an AI entity to own a patent would raise

broad fundamental issues relating to incentives for inventing and ‘AI personhood,’ which go

far beyond the scope of this discussion.”); Response from AIPPI, at 5-6; Response from AIPPI

Japan, at 4-5; Response from Askeladden, at 4; Response from BADC, at 5; Response from

BPLA, at 4; Response from CCIA (1st Response), at 3; Response from EPSON, at 3; Response

from Internet Association, High Tech Inventors Alliance, the Software and Information

Industry Association, and ACT, at 11; Response from IBM (Nov. 8, 2019), at 5; Response from

Wong, at 1; Response from Lori Pressman, at 2; Response from Zubek, at 2; Response from

IEEE-USA, at 6; Response from Juniper Networks, at 3; Response from Merck, at 3; Response

from R Street Institute (Nov. 8, 2019), at 3.

Response from IPO (Nov. 11, 2019), at 6 (“Generally, a natural person who trains the AI

process that creates an AI Generated invention should be able to be an owner.”).

Response from IBM (Nov. 8, 2019), at 4 (“Thus, at whatever point we deem machines

capable of invention, their inventions and the corresponding patents should be owned by

those that own them (e.g., those that own the machines).”); Response from Siemens, at 2

(“Attributing inventor or ownership rights to machines doesn’t feel right. Therefore, we

suggest expanding the right of the inventors to legal persons controlling the Al systems.”).

Guidance (PEG), which extracts and synthesizes key concepts identified by the courts as

abstract ideas to offer greater clarity in this area of the law.

Many commenters asserted that there are no patent eligibility considerations unique to AI

inventions.

That is, AI inventions should not be treated any differently than other

computer-implemented inventions. This is consistent with how the USPTO examines AI

inventions today. AI inventions are treated like all other inventions that come before the

Office. In fact, the USPTO has been examining and issuing patents claiming AI inventions for

years. Claims to an AI invention that fall within one of the four statutory categories and are

patent-eligible under the Alice/Mayo

test will be patent subject matter eligible under 35

U.S.C. § 101.

Some commenters stated that many AI inventions are at risk under the subject matter

eligibility analysis because they can be characterized as certain methods of organizing

human activity, mental processes, or mathematical concepts.

However, as one commenter

noted, the complex algorithms that underpin AI inventions have the ability to yield

technological improvements.

In addition, claims directed to an abstract idea will still be

patent-eligible if the additional claim elements, considered individually or as an ordered

combination, amount to significantly more than the abstract idea so as to transform it into

patent-eligible subject matter.

6. Are there any disclosure-related considerations unique to AI inventions? For

example, under current practice, written description support for computer-

implemented inventions generally require sufficient disclosure of an algorithm

to perform a claimed function, such that a person of ordinary skill in the art can

reasonably conclude that the inventor had possession of the claimed invention.

Does there need to be a change in the level of detail an applicant must provide in

order to comply with the written description requirement, particularly for deep-

learning systems that may have a large number of hidden layers with weights

that evolve during the learning/training process without human intervention or

knowledge?

This guidance was subsequently updated in October 2019; the substantive aspects of the

January 2019 PEG were unchanged. The current guidance documents on subject matter

eligibility, including the 2019 PEG and the examples, are available at

www.uspto.gov/PatentEligibility.

See, e.g., Response from AIPLA (Nov. 8, 2019), at 4; Response from IPO (Nov. 11, 2019), at

7-8; Response from Ford, at 1.

Alice Corp. Pty. Ltd. v. CLS Bank Int’l, 573 U.S. 208, 221 (2014); Mayo Collaborative Servs. v.

Prometheus Labs., Inc., 566 U.S. 66 (2012).

See, e.g., Response from ABA IPL (Nov. 8, 2019), at 13-14; Response from IBM (Nov. 8,

2019), at 5.

Response from ABA IPL (Nov. 8, 2019), at 15.

35 U.S.C. § 112(a)

has three separate and distinct disclosure requirements: written

description, enablement, and best mode.

These requirements apply to all applications

examined before the USPTO, including those directed to AI inventions.

The Manual of Patent Examining Procedure

(MPEP) and examiner training

provide

examination guidance regarding 35 U.S.C. § 112(a) that is consistent with the USPTO’s

understanding of the statute and legal precedent. In addition, in January 2019, the USPTO

issued guidance (January 2019 § 112 Guidance) to assist examiners in the examination of

claims in patent applications that contain functional language, particularly patent

applications in which functional language is used to claim computer-implemented

inventions.

The January 2019 § 112 Guidance may be especially helpful for evaluating AI

inventions, considering that patent applications related to AI inventions often include

computer-implemented inventions claimed, at least in part, with functional language.

Under current USPTO examination guidance, a determination of whether the disclosure

requirements are satisfied will depend on the facts of each application, including the subject

matter being claimed. To satisfy the written description requirement of 35 U.S.C. § 112(a),

applications for AI inventions that include claims to computer-implemented inventions that

recite functional language should provide sufficient detail in the specification regarding the

hardware, as well as software, to show that the inventor had possession of the full scope of

the claimed invention. In particular, the specification should disclose the computer and the

algorithm (e.g., detailed steps or procedures, formulas, diagrams, and/or flowcharts) that

perform the claimed function in sufficient detail such that one of ordinary skill can

reasonably conclude that the inventor possessed the claimed subject matter.

The majority of commenters shared the sentiment that there are no unique disclosure

considerations for AI inventions. One commenter stated that the principles set forth in the

Section 4 of the Leahy-Smith America Invents Act (AIA) designated pre-AIA 35 U.S.C. § 112,

¶¶ 1 through 6, as 35 U.S.C. § 112(a) through (f), effective as to applications filed on or after

September 16, 2012; see Public Law 112-29, 4(c), 125 Stat. 284, 296 (2011). AIA 35 U.S.C. §

112(a) and pre-AIA 35 U.S.C. § 112, ¶ 1 are collectively referred to in this paper as 35 U.S.C.

§ 112(a); AIA 35 U.S.C. § 112(b) and pre-AIA 35 U.S.C. § 112, ¶ 2 are collectively referred to

in this paper as 35 U.S.C. § 112(b).

Although this paper is limited to analyzing AI issues related to 35 U.S.C. § 112(a), issues

related to indefiniteness under 35 U.S.C. § 112(b) may arise for 35 U.S.C. § 112(f) functional

claim limitations where the specification does not provide sufficient corresponding

structure.

See MPEP §§ 2161-65, particularly § 2161.01, § 2181(IV), and § 2185.

See https://www.uspto.gov/patent/laws-and-regulations/examination-

policy/examination-guidance-and-training-materials. Note that examiners were recently

trained on examining computer-implemented functional claim limitations for compliance

with 35 U.S.C. § 112 (training completed March 14, 2019).

Examining Computer-Implemented Functional Claim Limitations for Compliance With 35

U.S.C. § 112, 84 Fed. Reg. 57 (Jan. 7, 2019).

USPTO’s examiner training materials regarding computer-implemented inventions “are

similarly applicable to AI-related inventions as to conventional algorithmic solutions.”

However, some commenters indicated that there are significant and unique challenges to

satisfying the disclosure requirements for an AI invention. One commenter noted that “AI

inventions can be difficult to fully disclose because even though the input and output may be

known by the inventor, the logic in between is in some respects unknown.”

These

characteristics of AI learning systems thus may drive further discussion regarding

enablement (see discussion regarding enablement below).

Several commenters noted that proper enforcement of the description requirement is

imperative for patent quality. For example, one commenter explained that it is “critical for

the USPTO to aggressively police the § 112 disclosure standards.”

7. How can patent applications for AI inventions best comply with the enablement

requirement, particularly given the degree of unpredictability of certain AI

systems?

According to current USPTO examination guidelines, the enablement requirement of 35

U.S.C. § 112(a) can be satisfied when the specification teaches one of ordinary skill in the art

how to make and use the full scope of the claimed invention without undue

experimentation.

When determining whether the specification satisfies the enablement

requirement and whether any necessary experimentation is undue, examiners are expected

to consider various factors called the “Wands factors.”

The Wands factors include: breadth

of claims, nature of the invention, state of the prior art, level of one of ordinary skill, level of

predictability in the art, amount of direction provided by the inventor, existence of working

examples, and quantity of experimentation necessary to make or use the invention based on

the content of the disclosure.

Generally, the amount of guidance or direction needed in the specification to enable the

invention is inversely related to the amount of knowledge in the state of the art, as well as

the predictability in the art.

The more that is known in the prior art regarding the nature

of the invention and the more predictable the art is, the less information is required to be

explicitly stated in the specification. Conversely, if less is known in the prior art about the

nature of the invention and the art is unpredictable, the specification should include more

Response from IPO (Nov. 11, 2019), at 14.

Response from IBM (Nov. 8, 2019), at 6.

Response from ABA IPL (Nov. 8, 2019), at 17.

See 84 Fed. Reg. 62; see also MPEP § 2164.01.

See 84 Fed. Reg. 62; see also MPEP § 2164.01(a).

Id.

See MPEP § 2164.03.

information as to how to make and use the invention in order to be enabling.

Thus, whether

a specification provides enabling support for the claimed invention is intensely fact-specific.

The commenters suggest that there are differing views on the predictability of AI systems.

One commenter stated that “most current AI systems behave in a predictable manner and

that predictability is often the basis for the commercial value of practical applications of

these technologies.”

Similarly, another commenter explained that “AI inventions are

inherently no more unpredictable than the underlying ML algorithm on which they rely.”

On the other hand, one commenter noted that some AI inventions may operate in a black box

because there is an “inherent randomness in AI algorithms.”

Some commenters suggested

that the principles applied in life sciences technology may be helpful when analyzing the

disclosure requirement for AI inventions. For example, one commenter explained that “the

greater degree of unpredictability associated with AI-based inventions makes it appropriate

to apply the written description requirement and the enablement factors from In re

Wands.”

8. Does AI impact the level of a person of ordinary skill in the art? If so, how? For

example: Should assessment of the level of ordinary skill in the art reflect the

capability possessed by AI?

AI is capable of being applied to various disciplines, from the life sciences and robotic

systems to agriculture and manufacturing processes. The ubiquitous nature of AI requires

an assessment of how it is affecting seemingly disparate fields of innovation. That is, AI may

have the potential to alter the skill level of the hypothetical “ordinary skilled artisan,”

thereby affecting the bar for nonobviousness.

An invention that would have been obvious to a person of ordinary skill before the effective

filing date of the claimed invention is not patentable.

As reiterated by the Supreme Court

in KSR International Co. v. Teleflex Inc., obviousness is a question of law based on underlying

factual inquiries.

These factual inquiries include the scope and content of the prior art, the

differences between the claimed invention and the prior art, and the level of ordinary skill in

the art.

Id.

Response from AIPLA (Nov. 8, 2019), at 8.

Response from Schwegman Lundberg & Woessner, at 9.

Response from IBM (Nov. 8, 2019), at 6.

Response from Genentech (Nov. 8, 2019), at 9.

While the “person of ordinary skill in the art” also has an impact on disclosure

requirements, with the instant question, the USPTO sought to hear from the public as to how

AI is impacting the level of ordinary skill in the art in assessing nonobviousness.

35 U.S.C. § 103 (2018).

550 U.S. 398, 406 (2007).

Graham v. John Deere Co., 383 U.S. 1, 17-18 (1966).

The person of ordinary skill in the art is a legal fiction, a person presumed to know the

relevant prior art.

Factors considered in determining the level of ordinary skill in the art

may include the type of problems encountered in the art, prior art solutions to those

problems, the rapidity with which innovations are made, the sophistication of the

technology, and the educational level of active workers in the field.

Each case may vary, not

every one of the aforementioned factors may be present, and one or more factors may

predominate the analysis.

Many commenters asserted that AI has the potential to affect the level of ordinary skill in an

art.

Furthermore, numerous commenters suggested that the present legal framework for

assessing the person of ordinary skill in the art is “adequate to determine the impact of AI-

based tools in a given field.”

Some commenters elaborated that the level of skill in any art

has traditionally grown over time based on the introduction of new technologies and that

“once conventional AI systems become widely available … such accessibility would be

expected to enhance the abilities of a person of ordinary skill in [an] art.”

In the words of

one commenter:

Just as the existence of test tubes impacts the level of a person of ordinary skill

in the chemical arts, and just as the existence of general purpose computers

impacts the level of a person of ordinary skill in the software arts (and many

Custom Accessories, Inc. v. Jeffrey-Allan Indus., Inc., 807 F.2d 955, 962 (Fed. Cir. 1986).

In re GPAC Inc., 57 F.3d 1573, 1579 (Fed. Cir. 1995).

Id.

Response from Abadi, at 3; Response from AIPPI, at 8; Response from CCIA (1st Response),

at 6; Response from Edward Ryan, at 4; Response from EPO, at 5; Response from Ericsson,

at 4; Response from Genentech (Nov. 8, 2019), at 10; Response from Internet Association,

High Tech Inventors Alliance, the Software and Information Industry Association, and ACT,

at 16; Response from IBM (Nov. 8, 2019), at 7; Response from IEEE-USA, at 8; Response from

Glucoft, at 2; Response from JIPA (Nov. 6, 2019), at 6; Response from JPAA, at 5; Response

from JPO, at 4; Response from KINPA, at 3; Response from NAPP, at 3; Response from R Street

Institute (Nov. 8, 2019), at 5; Response from Abbott, at 11; Response from Siemens, at 3.

Response from Novartis, at 11; see also Response from Juniper Networks, at 5 (“AI

inventions do not require any changes to the current legal requirements of the level of a

person [of] ordinary skill in the art”); Response from Merck, at 4 (“The ‘person of ordinary

skill in the art’ standard … should not change …”).

Response from BADC, at 6-7; see also Response from AIPPI Japan, at 8 (“advances in AI

technologies should be reflected in the determination of inventive step in the form of

improvement of level of technology used by a person skilled in the art”); Response from

Novartis, at 10 (“We believe AI must ultimately impact the definition and skill level of a

person of ordinary skill in the art, just as microscopes, calculators, and more conventional

software applications have in the past.”).

others), so [too] would AI affect the level of skill in the arts where it can be

made useful.

However, some commenters cautioned that such wide prevalence of AI systems has not yet

permeated all fields and counseled against declaring that all fields of innovation are now

subject to the application of “conventional AI.”

Others interpreted the question to assume

a future state in which AGI exists and machines have intelligence comparable to humans or

beyond. Those interpreting the question in this manner suggested that such machines are

not persons and, therefore, would not affect the legal standard of a “person” of ordinary skill

in the art.

9. Are there any prior art considerations unique to AI inventions?

The existence of prior art or the lack thereof has fundamental implications for determining

the fate of a filed U.S. patent application. The impact of AI on what can be considered prior

art, the quantity of prior art, and its accessibility are topics well worth considering.

35 U.S.C. § 102(a) states: “a person shall be entitled to a patent unless—(1) the claimed

invention was patented, described in a printed publication, or in public use, on sale, or

otherwise available to the public before the effective filing date of the claimed invention; or

(2) the claimed invention was described in a patent issued under section 151, or in an

application for patent published or deemed published under section 122(b) … ”

“The

categories of prior art documents and activities are set forth in AIA 35 U.S.C. 102(a)(1) and

the categories of prior art patent documents are set forth in AIA 35 U.S.C. 102(a)(2). These

documents and activities are used to determine whether a claimed invention is novel or

nonobvious.”

The majority of commenters stated that there were no prior art considerations unique to AI

inventions—that current standards were sufficient.

A minority of commenters indicated

Response from Edward Ryan, at 4.

See, e.g., Response from FICPI, at 5 (cautioning that AI’s impact on a field is “highly fact-

specific”); Response from Genentech (Nov. 8, 2019), at 10 (“the USPTO must be very cautious

in assessing which uses of [AI] are considered merely the exercise of ordinary skill in the

art”); Response from Internet Association, High Tech Inventors Alliance, the Software and

Information Industry Association, and ACT, at 17 (recognizing that “in some cases, applying

an existing ML model may not be simple, and hurdles overcome in order to achieve that

application may render application claims non-obvious … ”).

35 U.S.C. § 102(a).

MPEP § 2152.

Response from ABA IPL (Nov. 8, 2019), at 18; Response from AIPPI, at 8; Response from

Askeladden, at 6; Response from University of MD Center for Advanced Life Cycle

Engineering, at 3; Response from EPSON, at 5 (“We believe that the same prior art standards

that are applied to computer-implemented inventions should be applied to AI inventions.”);

Response from Ericsson, at 5; Response from International Federation of Intellectual

Property Attorneys, at 5-6; Response from Genentech, at 11; Response from Internet

that there were prior art considerations unique to AI inventions,

many of which focused

on the proliferation of prior art, such as the generation of prior art by AI,

and the difficulty

in finding prior art, such as source code related to AI.

A minority of commenters indicated

that while no prior art considerations unique to AI inventions currently existed, depending

on how sophisticated AI becomes in the future, unique AI prior art could become relevant.

Among all the responses, a common theme was the importance of examiner training and

providing examiners with additional resources for identifying and finding AI-related prior

art.

Association, High Tech Inventors Alliance, the Software and Information Industry

Association, and ACT, at 17; Response from IEEE-USA, at 8 (“The rules and procedures

governing prior art considerations for computer-implemented technology inventions and

broadly applicable enabling technology inventions should govern the prior art

considerations for AI inventions.”); Response from IPO (Nov. 11, 2019), at 17; Response from

Juniper Networks, at 5; Response from Rubin, at 10; Response from Lori Pressman, at 4;

Response from NAPP, at 3; Response from NSIP, at 10; Response from Maughan, at 3;

Response from Kumar, at 3; Response from Davis, at 7; Response from Schwegman Lundberg

& Woessner, at 10; Response from Naimpally, at 2; Response from Siemens, at 3; Response

from RF SUNY, at 5; Response from Edward Ryan, at 4.

Response from BADC, at 8; Response from Cardozo Intellectual Property Law Society, at 5

(“Already, some companies … have been using AI to generate patents in an attempt to

prevent adjacent patent claims. It is foreseeable that a company could use this technique to

generate massive amounts of prior art for the express purpose of rendering potential future

inventions unpatentable.”); Response from CCIA (1st Response), at 7; Response from IBM

(Nov. 8, 2019), at 8; Response from JEITA, at 7; Response from JIPA (Nov. 6, 2019), at 6;

Response from JPPA, at 5; Response from JPO, at 4; Response from Novartis, at 11; Response

from Prevencio, at 2; Response from Tata Consultancy, at 5.

See, e.g., Response from IBM (Nov. 8, 2019), at 8 (“AI will dramatically expand the scope of

prior art available. First and foremost, AI has the capability of generating a tremendous

amount of prior art.”).

See, e.g., Response from CCIA (1st Response), at 7 (“While standard AI techniques are more

likely to be described in the literature than is the case in software, there is still a significant

proportion of AI technology that is undocumented except in source code. This source code

may or may not be available and is generally considered difficult to search for.”).

Response from AIPLA (Nov. 8, 2019), at 8-9 (“In the event that an AI entity is considered

an inventor, the definition of ‘analogous’ may have to be significantly expanded, depending

on the capabilities of the inventive AI.”); Response from AIPPI Japan, at 9.

Response from BADC, at 8; Response from Internet Association, High Tech Inventors

Alliance, the Software and Information Industry Association, and ACT, at 17 (“Consistent

with the suggestions above that the USPTO provide a more proactive approach to examiner

technical training, the Associations recommend that the USPTO become more proactive

when it comes to providing prior art to examiners.”); Response from IPO (Nov. 11, 2019), at

17; Response from Juniper Networks, at 5 (“While examiners have been examining AI

10. Are there any new forms of intellectual property protections that are needed for

AI inventions, such as data protection?

Data is a foundational component of AI. Access to data for initial development and ongoing

training is necessary for AI development. This means that data and datasets, including their

collection and compiling, have value, particularly “big data" (i.e. extremely large data sets

that may be analyzed computationally to reveal patterns, trends, and associations). Data

protection under current U.S. law is limited in scope, and the U.S. does not currently have

intellectual property rights protections solely focused on data for AI algorithms.

Commenters were nearly equally divided between the view that new intellectual property

rights were necessary to address AI inventions and the belief that the current U.S. IP

framework was adequate to address AI inventions. Generally, however, commenters who

did not see the need for new forms of IP rights suggested that developments in AI technology

should be monitored to ensure needs were keeping pace with AI technology developments.

The majority of opinions requesting new IP rights focused on the need to protect the data

associated with AI, particularly ML. For example, one opinion stated that “companies that

collect large amounts of data have a competitive advantage relative to new entrants to the

market. There could be a mechanism to provide access to the repositories of data collected

by large technology companies such that proprietary rights to the data are protected but new

market entrants and others can use such data to train and develop their AI.”

Similarly, another commenter stated that “there may be gaps in IP protection for AI, and

specifically gaps in IP protection for the trained model and its associated coefficients.”

contrast, another opinion shared that training data is currently “protectable as a trade secret

or, in the event that the training data provides some new and useful outcome, then as a

patent.”

One commenter’s opinion was that the U.S. should not adopt the “European

database protection” scheme

, largely because it is “compromised by the fact that the

protection is tied to the investment expended to collect and/or verify the data.”

inventions for decades, Examiner training to identify relevant prior art would be

beneficial.”); Response from NSIP, at 10.

Response from AIPLA (Nov. 8, 2019), at 9.

Response from IBM (Nov. 8, 2019), at 8.

Response from Edward Ryan, at 4-5; see also Response from Schwegman Lundberg &

Woessner, at 11 (arguing that data used to train an ML algorithm is protectable under trade

secret or copyright law); Response from ABA IPL (Nov. 8, 2019), at 19 (arguing that training

technologies are protected under existing legal frameworks).

The USPTO has interpreted this as a reference to the Directive of the European

Parliament and of the Council of 11 March 1996 on the legal protection of databases.

Council Directive 96/9, 1996 O.J. (L 077) 20-28 (EC) available at https://eur-

lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A31996L0009.

Response from Gaudry, at 5.

In addition to data protection rights, one commenter suggested that if the patent system fails

to protect certain types of data, such as bioinformatics, “some alternative, sui generis form of

protection might be required to ensure that bioinformatics and other practical applications

of AI in biotechnology are protected forms of intellectual property.”

Another commenter

similarly stated that while it was currently “undecided as to whether new IP rights are

needed for AI inventions,” “should current systems prove unable to provide adequate

incentives to sufficiently effectuate the promise of AI, or should important gaps arise in those

systems, we believe it would be appropriate to consider new forms of IP,” including “an IP

right for trained models, and an IP right for nonpublic data where its generation required

substantial effort and investment (similar to the regulatory data protection (RDP) rights

available in [the] industry for proprietary clinical and other data submitted to FDA and other

regulatory authorities).”

Commenters did not provide concrete proposals on how any newly created IP rights should

function, and many, from both sides of the divide, called on the USPTO to further consult the

public on the issue.

11. Are there any other issues pertinent to patenting AI inventions that we should

examine?

The USPTO recognizes that the implications of AI-related issues on intellectual property

rights may be far-reaching. The agency has attempted to be comprehensive in posing

questions to the public on all related aspects of IP protection. Despite its best efforts,

however, the USPTO recognizes that there may be other issues that the public might wish to

bring to light. With this question, the USPTO intended to capture any issues not previously

addressed.

Speaking only to issues not dealt with elsewhere in the August 27, 2019 RFC (e.g., those

pursuant to 35 U.S.C. §§ 101, 103, 112),

a major theme the commenters stressed was the

need for examiner technical training and a call for memorializing guidance specific to AI for

patent examiners.

One commenter advised the USPTO to “invite and request industry trade

groups to adopt formal recommendations for patent applications and patent examination”

Response from Genentech (Nov. 8, 2019), at 11.

Response from Novartis, at 11.

Some responses used this question to reiterate the importance of subject matter eligibility,

obviousness, and disclosure requirements relative to AI inventions.

See, e.g., Response from CCIA (1st Response), at 8; Response from Internet Association,

High Tech Inventors Alliance, the Software and Information Industry Association, and ACT,

at 17-18; Response from Baysinger, at 3 (suggesting that the USPTO hire attorneys “versed

in data science with computer science backgrounds to form a think tank as the office

continues to receive AI patents” because “a case by case analysis and approach watching the

examinations would assist in setting standards and clarity for both examiners and

inventors.”).

on “questions relating to [the person of ordinary skill] in the art of AI, including training,

reference material, and benchmarks for improved performance.”

One commenter stated that “extending legal protection to AI [created] inventions may

require substantial changes in traditional legal approaches and frameworks, including

notions of property, ownership, and other non-IP legal principles akin to the development of

corporate law as we know it today.”

Another commenter stressed the importance of an

“open ecosystem of research” to U.S. economic and scientific leadership in the field of AI and

stated that the USPTO should “consider the economic and scientific risks of granting patents

to AI algorithm inventions and basic research that may have the potential to become

foundational.”

One commenter asked whether, given the dynamic nature of AI systems (i.e.,

“some systems constantly incorporate (i.e., learn from) additional examples/experience

while in operation”), the claims to such inventions need to be constantly updated as well.

Another commenter stated that “unexpected results should be weighed as heavily as they

are in evaluating biotechnology and pharmacological inventions.”

Other commenters called for further consideration of issues outside of patentability, such as

patent infringement and patent enforcement.

12. Are there any relevant policies or practices from other major patent agencies

that may help inform USPTO’s policies and practices regarding patenting of AI

inventions?

The USPTO participates in numerous global activities. On a multilateral level, the USPTO

represents the U.S. government on AI-related activities at the World Intellectual Property

(WIPO) and, the Organization for Economic Cooperation and Development. Additionally the

USPTO engages in cooperation directly with other intellectual property offices, both one-on-

one, for example through bilateral exchanges on the patentability of AI inventions, and

multilaterally in groups like the IP5 Taskforce on New Emerging Technologies and AI

Response from Rubin, at 14-15. Note: “POSITA” is a reference to “a person of ordinary skill

in the art.”

Response from AIPLA (Nov. 8, 2019), at 9; see also Response from EPO, at 5 (“The impact

of the development of AI technology on society, including changes in the right holders’

position, employment market and ethical challenges that AI poses are also areas which may

need to be addressed as the technological development progresses.”); Response from JPO, at

4 (“When a claim is made for a substance (compound, composition, pharmaceutical, etc.)

whose physical properties are predicted by AI, it might be helpful if the details on what is

required to be patentable (i.e., whether only calculation results are sufficient or chemical

experiments are additionally required) are described in Specification.”).

Response from Menart, at 2.

Response from Davis, at 7.

Response from NAPP, at 4.

See, e.g., Response from Genentech (Nov. 8, 2019), at 11-12; Response from IBM (Nov. 8,

2020), at 9.

(NET/AI). Through these channels, the USPTO is able to share its policies and be informed of

relevant policies and practices from other major patent agencies.

Commenters highlighted the respective work at other patent offices, particularly at the

European Patent Office (EPO) and Japan Patent Office (JPO). Commenters cited the reports,

guidelines, and patent examination examples on AI issued by these offices, all of which were

suggested to be informative to the USPTO.

One commenter noted that “the JPO and more

recently, KIPO (Korean Patent Office), have established separate and specific ‘AI examination

groups’ that can focus on only AI matters, so such could be adopted in the USPTO as well.”

Commenters also called attention to the Intellectual Property Office of Singapore (IPOS),

which reportedly created an expedited examination path for AI technologies.

Others

generally sought to have the USPTO continue its multilateral engagements on AI through

WIPO and the IP5.

One commenter specifically stated that “it is our desire as an

organization of international scope to see that the respective laws and administrative

practices of IP5 are evolving in a common direction.”

On the other hand, one commenter

“caution[ed] against further attempts to harmonize patent laws and procedures, especially

as it relates to patenting AI” because “U.S. patent law has long been the gold standard for

patent protection and a major driver in the success of the U.S. innovation economy.”

100

See Response from AIPPI Japan, at 10 (suggesting that the USPTO provide criteria in

guidelines with many examples, like the EPO and JPO); see also Response from BADC, at 9-

10 (noting examination guidelines of other patent offices); Response from BPLA, at 5;

Response from FICIPI, at 6; Response from JPAA, at 6; Response from Abbott, at 11; Response

from Alliance for AI in Healthcare, at 5; Response from JEITA, at 8; Response from JIPA (Nov.

6, 2019), at 8.

Response from KINPA, at 4.

See, e.g., Response from NSIP Law, at 11.

See, e.g., Response from Novartis, at 13 (“consult [with other patent offices] and share

learnings and best practices as appropriate and consistent with the U.S. system’s

Constitutional goals”); Response from Ericsson, at 5; Response from Merck, at 5.

Response from AIPPI, at 10; see also Response from IBM (Nov. 8, 2019), at 9 (“IBM asks

the Office to continue to monitor the development of IP policy surrounding AI in other

jurisdictions and ensure that the IP policies in the U.S. do not comparatively disadvantage AI

inventions in the U.S.”).

100

Response from IEEE-USA, at 10.

PART II—Responses to the RFC on Intellectual Property Protection for Artificial

Intelligence Innovation, issued on October 30, 2019

A summary of the comments received in response to the RFC on the impact of AI on IP policy

areas other than patent law, including copyrights, trademarks, database protections, and

trade secret law, issued on October 30, 2019, is included below, organized by the question

appearing in the RFC. Commenters included bar associations, industry associations,

academia, and various stakeholders, both national and international. Representatives from

electronics, software, media, and pharmaceutical industries responded to the RFC.

1. Should a work produced by an AI algorithm or process, without the involvement

of a natural person contributing expression to the resulting work, qualify as a

work of authorship protectable under U.S. copyright law? Why or why not?

Under current U.S. law, a work created without human involvement would not qualify for

machines would qualify for copyright protection if other conditions are met. The Supreme

Court has long recognized copyright protection for creative works,

101

even when an author

is assisted by a machine.

102

The U.S. Copyright Office, in its Compendium of Practices (Third Edition), has addressed the

question of human contribution to creative works. It notes that:

[T]he Copyright Act protects “original works of authorship.” 17 U.S.C. § 102(a). To

qualify as a work of “authorship,” a work must be created by a human being. ….

Works that do not satisfy this requirement are not copyrightable. …. Similarly, the

Office will not register works produced by a machine or mere mechanical process

101

See Feist Publ’ns, Inc., v. Rural Tel. Serv. Co., 499 U.S. 340 (1991), a case establishing that

information alone, without a minimum of original creativity, cannot be protected by

102

See Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884), a case that upheld the

power of Congress to extend copyright protection to photography when the photographer

demonstrates creativity by posing a subject; selecting and arranging the costume, draperies,

and other various accessories; and arranging and disposing the light and shade. The court

described copyright as “as the exclusive right of a man to the production of his own genius

or intellect.” Id. at 58 (emphasis added). See also Bleistein v. Donaldson Lithographing Co.,

188 U.S. 239 (1903), a case in which advertisements for a circus were held to be protected

by copyright. The court noted that the “copy”—i.e., the work that is protected by copyright—

“is the personal reaction of an individual upon nature. Personality always contains

something unique. It expresses its singularity even in handwriting, and a very modest grade

of art has in it something irreducible which is one man’s alone. That something he may

that operates randomly or automatically without any creative input or intervention

from a human author.

103

Accordingly, the U.S. Copyright Office will not grant a copyright registration unless the

author is a human being.

104

A draft update to the Compendium further specifies that works

“produced by a machine or mere mechanical process that operates randomly or

automatically without any creative input or intervention from a human author” will not be

granted copyright registration.

105

The United States is a member of the Berne Convention, the leading multilateral agreement

establishing the framework for international copyright protection, which has been

incorporated in large part in the TRIPs Agreement and subsequent U.S. free trade

agreements.

106

The Berne Convention has been interpreted to require protection only for

works that are original and created with human involvement.

107

The vast majority of commenters acknowledged that existing law does not permit a non-

human to be an author (outside of the work-for-hire doctrine, which creates a legal fiction

103

U.S. Copyright Office, Compendium of U.S. Copyright Office Practices, (3d ed. 2017) § 313.2

(“Compendium 3”) (emphasis added).

104

Compendium 3 at § 306.

105

A draft Compendium proposing updates that reiterated this principle specifically with

respect to machine-generated works was published for public comment on March 15, 2019,

and states:

Similarly, the Office will not register works produced by a machine or mere mechanical

process that operates randomly or automatically without any creative input or

intervention from a human author. The crucial question is “whether the ‘work’ is

basically one of human authorship, with the computer [or other device] merely being

an assisting instrument, or whether the traditional elements of authorship in the work

(literary, artistic, or musical expression or elements of selection, arrangement, etc.)

were actually conceived and executed not by man but by a machine.”

See U.S. Copyright Office, Public Draft Compendium of U.S. Copyright Office Practices, §

313.2 (March 15, 2019), https://www.copyright.gov/comp3/chap300/chap300-draft-3-

15-19.pdf (quoting U.S. Copyright Office, Report to the Librarian of Congress by the Register

of Copyrights 5 (1966)).

106

Berne Convention for the Protection of Literary and Artistic Works (Paris Act, 1971).

107

See Sam Ricketson, People, or Machines: The Berne Convention and the Changing Concept

of Authorship, Horace S. Manges Lecture, 16 Colum. -VLA J. L. & Arts 1 (1991-1992). Ricketson

puts forth that although the Berne Convention did not define authorship, there was

nonetheless a basic agreement among the contracting states that the meaning of the term

referenced human beings, and because of this, it was thought unnecessary to define the term.

Id.

for non-human employers to be authors under certain circumstances);

108

they also

responded that this should remain the law. One comment stated: “A work produced by an AI

algorithm or process, without intervention of a natural person contributing expression to

the resulting works, does not, and should not qualify as a work of authorship protectable

under U.S. copyright law.”

109

Multiple commenters noted that the rationale for this position

is to support legal incentives for humans to create new works.

110

Other commenters noted

that AI is a tool, similar to other tools that have been used in the past to create works:

“Artificial intelligence is a tool, just as much as Photoshop, Garage Band, or any other

consumer software in wide use today … the current debate over whether a non-human object

or process can be ‘creative’ is not new; the government has long resisted calls to extend

authorship to corporations or entities that are not natural humans .”

111

A minority of commenters suggested that a sufficiently creative work made by AI without

human intervention should be copyrightable and that copyright law should allow authorship

to inhere either in the owner/controller of the AI system or in the person/user who fixes the

work in its final form.

112

108

The Copyright Act provides that “[i]n the case of a work made for hire, the employer or

other person for whom the work was prepared is considered the author for purposes of this

title, and, unless the parties have expressly agreed otherwise in a written instrument signed

by them, owns all of the rights comprised in the copyright.” 17 U.S.C. §201(b). The doctrine

applies whether the employer is a human being or a corporation but the actual creator of

any work protected by copyright has always needed to be human.

109

Response from ABAB IPL (Jan. 9, 2020), at 4 (emphasis added). Other commenters

highlight some of the potentially unforeseen and deleterious consequences of changing the

law: “The result will not just be one more disruption in the work force; we will be a much

poorer society because those AI-created works, no matter their superficial similarity to

works of human provenance, will lack the experience and emotion of the human artist…

[Human] experience makes the arts so fundamentally important to every human society. AI

generated works rehash, mash-up, and rework what they are fed; they do not tell or emote.”

Response from the Authors Guild, Inc., at 5.

110

Response from ABA IPL (Jan. 9, 2020), at 4-5; Response from IPO (Jan. 10, 2020), at 2.

111

Response from Public Knowledge, at 1.

112

See, e.g., Response from the ITIF, at 4. Another commenter analogized AI to a player piano:

“Someone has created the roll that will tell the player piano which sounds or notes to

produce. This is analogous to choosing which data to use in training the algorithm. . . . The

imperfection in the analogy is that the human’s actions in designing and ‘training’ a player

piano are perfectly determinative. . . . This will not necessarily be true for an algorithm. . . .

All of this is to say that an AI algorithm is a machine built through human effort. Any works

that result from that machine should be considered works of authorship attributable to the

humans who constructed that machine, just as they would be for other types of machines.”

Response from Feamster, at 2; see also Response from Professor Nina Brown, Syracuse

University, at 7. “My recommendation is not to vest ownership in the algorithm/computer

itself, as it is a piece of chattel rendering it incapable of owning anything. Rather, I

2. Assuming involvement by a natural person is or should be required, what kind of

involvement would or should be sufficient so that the work qualifies for copyright

protection? For example, should it be sufficient if a person (i) designed the AI

algorithm or process that created the work; (ii) contributed to the design of the

algorithm or process; (iii) chose data used by the algorithm for training or

otherwise; (iv) caused the AI algorithm or process to be used to yield the work; or

(v) engaged in some specific combination of the foregoing activities? Are there

other contributions a person could make in a potentially copyrightable AI-

generated work in order to be considered an “author”?

U.S. law requires a minimum threshold of human creativity to qualify for copyright

protection. A work’s copyrightability depends on whether creative expression, contributed

by someone who can reasonably be described as an author of the work, is evident in the

resultant work.

In Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884), the U.S. Supreme Court

recognized the way in which the photographer posed his subject, selected and arranged the

costume and lighting, and otherwise demonstrated creative authorship in his choices. The

court also recognized the possibility that many photographs would not qualify for copyright

protection in instances when a photographer did not exhibit the requisite level of creativity.

Thus, according to current case law, some level of authorial creativity would be necessary

for copyright protection.

More broadly speaking, commenters’ response to this question either referred back to their

response to the first question without comment (stating that human involvement is

necessary for copyright protection) or referred back and made some further observations or

clarifications, often pointing out that each scenario will require fact-specific, case-by-case

consideration. Several commenters raised or reiterated their view that natural persons, for

the foreseeable future, will be heavily involved in the use of AI, such as when designing

models and algorithms, identifying useful training data and standards, determining how

technology will be used, guiding or overriding choices made by algorithms, and selecting

which outputs are useful or desirable in some way. The commenters thus predicted that the

outputs of AI will be heavily reliant on human creativity.

recommend that ownership is vested with the person/entity responsible for fixing the work

…” Id.

3. To the extent an AI algorithm or process learns its function(s) by ingesting large

volumes of copyrighted material, does the existing statutory language (e.g., the

fair use doctrine) and related case law adequately address the legality of making

such use? Should authors be recognized for this type of use of their works? If so,

how?

Existing statutory and case law should adequately address the legality of machine “ingestion”

in AI scenarios. Mass digitization and text and data mining (TDM), as relevant examples of

other activities with copyright implications, may be considered copyright infringement or

fair use, depending on the facts and circumstances at issue. Copyright law in its current form

appears to be adaptable to new technologies and circumstances, including those raised by

AI.

Copying substantial portions of expressive (copyrighted) works, even for non-expressive

purposes implicates the reproduction right and, absent an applicable exception, is an act of

113

Depending on the copyrighted work and the activity taking place,

it may or may not be eligible for an exception to the reproduction right.

114

Regardless of whether an “ingestion” use is determined to be an infringement or not in a

given situation, there is a separate issue of whether authors of ingested works should be

remunerated for these types of uses. Many publishers now include TDM terms in their

contracts and expressly set a licensing fee for for-profit entities or permit licensing at no

additional cost for researchers and public research organizations, while ensuring that the

licensed content is machine-readable and searchable.

115

Advocates for authors have

suggested that when copyrighted works are used as inputs into AI systems to train the AI to

create works of authorship or engage in other activities that result in remuneration, the

113

Violent Hues Prods., LLC, 922 F.3d 255, 265 (4th Cir. 2019) (“As a basic matter, copyright

infringement is a strict liability offense, in which a violation does not require a culpable state

of mind”); EMI Christian Music Grp., Inc. v. MP3tunes, LLC, 844 F.3d 79, 89 (2d Cir. 2016)

(“Copyright infringement is a strict liability offense in the sense that a plaintiff is not required

to prove unlawful intent or culpability, and a user does not have to share copyrighted works

in order to infringe a copyright”)(internal citations omitted).

114

One scenario is TDM, which is commonly understood to mean an automated process of

selecting and analyzing large amounts of text or data resources for purposes such as

identifying patterns or finding relationships. Another scenario is mass digitization of a

certain body of work or a certain medium of work (e.g., Google Books digitizing the world’s

non-digitized books to make them searchable).

115

STM, Text and Data Mining: Building a healthy and sustainable knowledge ecosystem for

Europe, Dec. 2017, https://www.stm-

assoc.org/2017_12_20_2017_12_STM_Text_and_Data_Mining_Summary.pdf.

authors should be entitled to a share of the revenues generated by the AI. The recognition

sought is not attribution but rather remuneration.

116

The ingestion of copyrighted works for purposes of machine learning will almost by

definition involve the reproduction of entire works or substantial portions thereof.

Accordingly, whether this constitutes copyright infringement will generally be determined

by considering the applicability of the fair use doctrine, an exception set forth in section 107

of the Copyright Act, 17 U.S.C. § 107. Fair use is applied on a case-by-case basis, requiring

courts to weigh several statutory factors, and is highly fact-dependent.

117

116

See USPTO AI: Intellectual Property Policy Considerations event (Jan. 31, 2019), at minute

43, available at

https://rev-vbrick.uspto.gov/#/videos/d6e591c3-64cf-4d74-ab35-

9f387a2da4b2, highlighting the views of Mary Rasenberger, Executive Director of the

Authors Guild. Rasenberger asserts that companies “should not get a pass at paying royalties

to use copyrighted works en masse to train their computers.” She highlights that computers

read and consume expressive works and, indeed, are changing the ways in which people are

consuming creative works. She concludes that “we want AI to be able to use existing works

and learn from them, but it doesn’t have to be free. AI reading involves copying; it extracts

huge value from copying … . The solution, I think, is fairly simple—we’ve done it before in

transaction costs are too high … . We have to figure out a way to pay for computer reading.”

117

“Notwithstanding the provisions of sections 106 and 106A, the fair use of a copyrighted

work, including such use by reproduction in copies or phonorecords or by any other means

specified by that section, for purposes such as criticism, comment, news reporting, teaching

(including multiple copies for classroom use), scholarship, or research, is not an

infringement of copyright.

In determining whether the use made of a work in any particular case is a fair use the

factors to be considered shall include:

(1) the purpose and character of the use, including whether such use is of a

commercial nature or is for nonprofit educational purposes;

(2) the nature of the copyrighted work;

(3) the amount and substantiality of the portion used in relation to the copyrighted

work as a whole; and

(4) the effect of the use upon the potential market for or value of the copyrighted

work.

The fact that a work is unpublished shall not itself bar a finding of fair use if such finding

is made upon consideration of all the above factors.” 17 U.S.C. § 107.

Fair use is a legal doctrine that promotes freedom of expression by permitting the

unlicensed use of copyright-protected works in certain circumstances. Section 107 provides

the statutory framework for determining whether something is a fair use and identifies

certain types of uses—such as criticism, comment, news reporting, teaching, scholarship,

and research—as examples of activities that may qualify as fair use. Section 107 calls for

When AI algorithms or processes “learn” their functions by ingesting copyrighted works,

reproductions of those works are made in the process as the works are digitized and/or

“read” by the AI algorithms or processes. Some mass digitization scenarios may be a fair

use,

118

whereas others may be infringements.

119

Although mass digitization for purposes of

machine learning (ML) “ingestion” processes —and large-scale ingestion of already-digitized

works—has not yet been tested by the courts, some rights holders argue that AI trainers

should be required to compensate the authors and rights holders whose copyrighted works

their machines are ingesting as a simple matter of doing business.

120

consideration of four factors in evaluating a question of fair use. Courts evaluate fair use

claims on a case-by-case basis, and the outcome of any given case depends on a fact-specific

inquiry. This means there is no formula to ensure that a predetermined percentage or

amount of a work—or a specific number of words, lines, pages, or copies—may be used

without permission.

118

See Authors Guild, Inc. v. HathiTrust, 755 F.3d 87 (2d Cir. 2014) (search and accessibility

uses of digitized books were found to be fair use); see also Authors Guild v. Google, Inc., 804

F.3d 202 (2d Cir. 2015) (Google’s unauthorized digitizing of tens of millions of copyright-

protected books, Google’s creation of a search functionality, and Google’s display of snippets

from those books were found to be non-infringing fair uses); A.V. ex rel. Vanderhye v.

iParadigms, L.L.C., 562 F.3d 630 (4th Cir. 2009)(unauthorized digital archiving of student

papers for purposes of preventing plagiarism constituted fair use because the activity was

unrelated to the works’ expressive content).

119

See Fox News Network, LLC v. TVEyes, Inc., 883 F.3d 169 (2d Cir. 2018), involving a

comprehensive database of television programs downloaded without authorization by a

monitoring service that was found to infringe the distribution right in those programs by

making them available for downloading by subscribers who conducted searches for videos

of interest. The court rejected the defendant’s fair use defense, but the defendant’s initial

copying of the programs was not an issue before the court.

120

See, e.g., Response from the AAP, at 2; Response from the Authors Guild, at 10; Response

from STM, at 4. STM further explained that the deployment of AI tools and mining tools is

becoming more commonplace and that publishers make works and datasets available in

ways that facilitate machine reading and learning. Id. Another commenter made a more

granular suggestion about establishing collecting societies to ensure that micropayments be

made to individual copyright owners. See Response from Getty Images, at 3. Also note that,

to address the challenge of mass digitization generally, including whether and how to

compensate the authors whose works are being digitized, some countries provide for

collective rights management under which the exercise of copyright for individual works by

collective rights management organizations (CMOs) facilitates the dissemination of works,

at lower transaction costs than with individual licensing, by administering the remuneration

collected from the users and distributing it to the appropriate rights holders. Under this

approach, known as extended collective licensing, after determining that rights holders and

users want to participate in a collective licensing arrangement, the government authorizes a

CMO to negotiate licenses for a particular class of works (e.g., textbooks, newspapers, or

magazines) or a particular class of uses (e.g., reproduction of published works for

Most commenters referred to the existing statute and its interpretation by the courts and

noted that the use of copyrighted material to “train” AI processes (constituting ML) may

violate the reproduction right of a copyright owner under 17 U.S.C. §106(1), and that this use

may or may not be a non-infringing fair use. Most commenters found that existing law does

not require modification, as fair use is a flexible doctrine and is capable of adapting to the

use of copyrighted works in an AI context. Many commenters included hypotheticals or

specifically walked through the fair use factors, citing well-known fair use cases, including

Authors Guild v. Google, Inc., (the “Google Books” case) (finding that Google’s digitizing of

TVEyes, Inc., (finding that TVEyes’ use of Fox’s copyrighted content in order to allow a

customer to find material of interest to them was not fair use).

Some commenters were explicit in their view that the use of copyrighted works for ML

should be permissible and compensated;

121

one commenter suggested an opt-in or opt-out

mechanism for rights holders in the event some form of blanket licensing was

implemented.

122

Another commenter noted that any perception that the Google Books case

educational or scientific purposes). When the CMO negotiates a license with a particular

user, that license is automatically extended—by operation of law—to works in the specified

class owned by all the rights owners, regardless of whether they belong to the collective

organization or not. All copyright owners are entitled to receive a share of the royalties that

the collective receives for that work from its licensees. In some countries, copyright owners

may be allowed to opt out of some uses of their works or demand individual remuneration

if they believe they are entitled to a larger share of the royalties for the use of their works.

As mentioned above, this model is potentially relevant for those who believe that “mass

ingestion” should provide remuneration to authors whose works are digitized. There has not

yet been a sustained conversation among experts as to whether this model would be

practical or desirable in the United States.

121

See, e.g., Response from the AAP, at 2; Response from the Authors Guild, at 10; Response

from STM, at 4. STM further explained that the deployment of AI tools and mining tools is

becoming more commonplace and that STM publishers “increasingly publish copyright

works and associated datasets with AI ingestion technologies in mind. In other words,

facilitate machine reading, learning, etc. It follows that licensing…should in most instances

be the method of choice for enabling access to copyright works.” Id. Another commenter

made a more granular suggestion about establishing collecting societies to ensure that

micropayments be made to individual copyright owners: “New statutory guardrails are

needed to ensure existing law is sensibly and fairly adapted to consider unique attributes

associated with the use of large volumes of copyright work. In this regard, it is essential that

the ingestion of any volume of copyrighted material in connection with AI learning is not

considered a ‘transformative’ fair use by default.” See Response from Getty Images, at 3.

122

Response of Lori Pressman, at 2. “Rather than have to have an argument that the mass

use of copyrighted works is ‘fair’ because of the character of the output, it would be better to

allow creators to opt in, or out of such use. Thus, I advocate using technology, such as an

electronic watermark, to accord copyright holders control over the use of their copyrighted

provides “carte blanche for copying entire works into databases misapprehends the limits of

that decision.”

123

Another commenter considered unlicensed use of copyrighted material by

search engines as copyright infringement and stated: “Tech platforms that appropriate vast

quantities of news content for this purpose should pay for the privilege of doing so, no less

than they should pay for the electricity that powers their computers or motorists for the fuel

that powers their cars.”

124

Commenters did not specifically take up the question of

recognition of the source materials, although several noted, as mentioned above, that

compensation would be necessary or appropriate.

Another, smaller subset of commenters expressed their view that using content to train,

tune, and/or test an AI system should automatically be presumed a fair use. But among most

of these comments, there were some qualifiers and an acknowledgment that “the legality of

these kinds of uses will be a fact-specific decision that augurs against the development of

bright-line rules.”

125

A subtheme of this set of comments suggested that allowing AI systems

to ingest content and be trained free from copyrightability will promote innovation.

4. Are current laws for assigning liability for copyright infringement adequate to

address a situation in which an AI process creates a work that infringes a

copyrighted work?

While an AI machine cannot currently own intellectual property rights, it may be able to

infringe others’ rights. Federal copyright law sets forth a straightforward standard for

owner” is liable for copyright infringement.

126

If the AI’s owner takes sufficient action to

cause the AI’s infringement—through programming, data inputs, or otherwise—the owner

could directly or contributorily infringe. Alternatively, if AI becomes more autonomous, it is

conceivable that an AI owner might be vicariously liable for the AI’s copyright infringement

works. The watermark would allow copyright holders to place restrictions on how their

work is used.” Id.

123

Response from Kernochan Center, Columbia Law School, at 5.

124

Response from the News Media Alliance, at 5.

125

Response from SIIA (Jan. 10, 2020), at 7; see also Response from BSA, at 4-5 (It is

impossible to draw a generalized conclusion that all applications of AI involving the

reproduction of copyrighted works will be a fair use. But the case law suggests strongly that

the use of copyrighted works for the training of AI systems will be a fair use when the

reproductions are used to generate new insights whose value is unrelated to the expression

in the underlying works.”). See also Response from the Berkman Klein Center for Internet

and Society, at 1. (“This section of my comments divides relevant AI applications into two

categories: ‘non-expressive’ uses and ‘market-encroaching’ uses. It then explains why the

former are clearly fair use and the latter are not.”).

126

17 U.S.C. § 501(a).

when the owner possesses the right and ability to supervise the infringing conduct and a

financial interest in the infringement.

127

Most commenters noted that existing laws are sufficient and can be applied to situations

involving AI,

128

while also acknowledging that a natural or legal person will need to be held

responsible.

129

Some commenters took a more skeptical approach: “Current laws may rely

on certain general law doctrines. Existing laws governing liability for copyright infringement

and the fair use defense are probably not clear with respect to AI generated works.”

130

Another commenter noted that “enforcement under theories of contributory and vicarious

infringement will require the courts to consider novel issues regarding, among other things,

agent[sic], control, and foreseeability of the AI device’s acts. And, as AI becomes increasingly

autonomous, changes to the law may prove necessary.”

131

Another commenter noted that

current rules on damages are not appropriate, since “knowing” infringement will be harder

to define.

132

5. Should an entity or entities other than a natural person, or company to which a

natural person assigns a copyrighted work, be able to own the copyright on the

AI work? For example: Should a company who trains the artificial intelligence

process that creates the work be able to be an owner?

outlines a finite set of scenarios giving rise to copyright ownership. Under the copyright law,

one can own a copyright by: (1) being the author or a joint author, (2) being deemed to be

the author under the work-made-for-hire doctrine, or (3) obtaining an assignment of the

133

127

See A&M Records, Inc. v. Napster, Inc., 239 F.3d 1004 (9th Cir. 2001).

128

See, e.g., Response from IPO (Jan. 10, 2020), at 4 (citing Sony Corp. of Am. v. Universal City

Studios, Inc., 464 U.S. 417 (1984) as guidance concerning how the courts should assess the

potential liability of creators and users of technology that might be used to infringe

copyrights).

129

Response from the Copyright Clearance Center, at 4.

130

Response from AIPLA (Jan. 10, 2020), at 7.

131

Response from NYIPLA, at 7; see also Response from RIAA, at 7. (“There will be a

continuum of human contribution associated with various AI outputs, and the difficult issue

will be determining where on that continuum of human involvement is sufficient to justify

that can only be resolved on a case-by-case basis.”).

132

Response from KEI, at 2. We take this terminology to be responsive to the “knowing”

standard described in the Nimmer treatise: “Basically, there are three levels of awards: a

basic award, an increased measure, and a decreased measure. In brief, willfulness warrants

the increase, innocence the decrease, and all other cases are computed according to the

standard measure. For ease of terminology, intermediate between willful and innocent

conduct lies the domain of ‘knowing infringement. ’” 4 Nimmer on Copyright § 14. 04 (2019).

133

See 17 U.S.C. §§ 201-205.

Commenters generally felt ownership of copyrights in the AI context can be dealt with by

commercial negotiations.

134

However, answers to this question demonstrated a mixed

understanding of what was being asked.

135

One response, for example, said, “No. Ownership

should vest in the author (or employer of the author, in the case of works made for hire) and

may then be assigned to another natural or juridical person.”

136

A similar response did not

answer in the negative but noted:

There is nothing under U.S. law to prevent an entity from being an owner of

copyrightable works that are part of AI systems or are created by persons

using AI as a tool. Copyrights in AI related works are assignable to entities just

as any other copyrights are under Section 201(d) of the Copyright Act.

137

In general, responses to this question referred back to the commenters’ answers from

questions 1 and 2, which acknowledged that a non-human cannot be an author and noted

that the use of AI is reliant on human creativity.

6. Are there other copyright issues that need to be addressed to promote the goals

of copyright law in connection with the use of AI?

As referenced above, the term “AI” can comprise a range of meanings. For example,

generative algorithms (i.e., algorithms that possess the ability to create data) are responsible

for producing unique works of varying complexity. These works can result from

collaborative efforts between a human creator and an AI program, or they can result from an

independent AI process or algorithm. Therefore, no bright-line rule about “AI and

authorship” or “AI and copyrightability” can be made; rather, it depends on a human being’s

role in tandem with AI in generating a creative output that is potentially copyrightable.

A frequent response was that it is too soon to answer this question. “Generally, AI

technologies are still in their infancy and there is no known instance of a machine-generated

creative output without some human intervention and/or direction, and so it is difficult to

answer, in a useful way, some of the ... inquiries, which contemplate circumstances that have

not yet come to pass.”

138

And: “[T]he scenario envisioned by [question 1] assumes a degree

of autonomy by a ‘general AI system’ that is more of an aspiration than a present reality in

134

Response from ITIF, at 11; Response from Boomy Corp., at 13 (“These relationships are

governed by contracts, and it’s working.”). Boomy is a music and artificial intelligence

business. It uses a variety of algorithms, some assisted by machine learning, to generate

musical compositions with user input.

135

See, e.g., Response from SIIA (Jan. 10, 2020), at 8 (“SIIA finds the question a little

confusing.”).

136

Response from AIPLA, at 7.

137

Response from ABA IPL (Jan. 9, 2020), at 10.

138

Response from the Entertainment Software Association, at 3. One outlier response is from

Getty Images, which supports taking action now as opposed to a wait-and-see approach.

Response from Getty Images, at 1.

our field.”

139

There were both general statements and detailed examples of how “AI” as we

know it today cannot produce copyrightable works without human input in the following

areas: the level of design, identification of useful training data and standards, determination

of how technologies will be used in commerce and research, guidance or override of choices

made by algorithms, and selection of outputs that are useful or desirable.

140

Relatedly, many commenters noted that there is no standard working definition of AI that

would allow them to make pointed, consistent responses to the questions. One commenter

noted that AI “is mostly a colloquial stand-in for ‘algorithms’ and ‘automation,’ or perhaps

more specifically ‘algorithms created with ML,’ however, there is no static definition of ‘ML’

either.”

141

Another theme identified by some commenters is the issue of bias. These commenters found

that the ambiguous state of copyright enforcement with respect to ML drives developers to

use “low-quality data that exists in the public domain,” thereby inadvertently biasing the

many AI applications that are built on that data.

142

Finally, a handful of commenters raised

the issue of “deep fakes” (in which a person in an existing image or video is replaced with

someone else’s likeness) and the legitimate uses of AI-based technology for creative

purposes, suggesting that rights of publicity and other related laws may need to be tweaked

to take into account cases in which a “deep fake” is derived from copyright-protected

material.

143

7. Would the use of AI in trademark searching impact the registrability of

trademarks? If so, how?

To search for confusingly similar designs, global trademark offices index all incoming mark

drawings through human-assigned design codes. But different humans may perceive the

same design element differently, and some may not identify hidden elements in certain

design marks because of optical illusions, such as hidden designs created by letters or other

visual phenomena, even though others might. Properly trained AI software could be used to

supplement the human process of assigning design codes by identifying all perceptible

design elements in a mark. If the design codes assigned to images are more comprehensive,

any image search results will be likewise more comprehensive.

139

Response from Genentech (Jan. 10, 2020), at 6.

140

See, e.g., Response from the Copyright Clearance Center, at 4; see also Response from IPO,

at 3.

141

Response from Boomy, at 3.

142

Response from the R Street Institute (Jan. 10, 2020), at 1; see also Response from Adobe,

at 6; Response from Wikimedia Foundation, at 7 (“When selecting training data, developers

of AI/ML systems will often select public domain or freely licensed works, both for ease of

access and to avoid potential infringement liability exposure.”).

143

See, e.g., Response from ITIF, at 14-15; see also Response from the Authors Guild, at 12.

However, assembling the data sets that would serve as the raw materials to train the AI

software for image search or design coding raises potential impacts on registrability.

Globally, trademark law focuses on consumer perception, but different populations and

cultures may perceive the same stimuli differently. If the data is shared amongst national

offices, AI image searching could produce results that, while accurately reflecting an

aggregate global perception, may not accurately reflect consumer perception in the relevant

territory as to what could be a confusing element. Relatedly, one commenter noted that the

sharing of likelihood of confusion determination data between national offices is potentially

problematic due to the fact that many countries’ likelihood of confusion tests are

“perceptibly different” from others’.

144

Use of AI software by the USPTO

Most commenters agreed that AI software will improve USPTO examination searching

efficiency. A few commenters predicted that customers may face more likelihood of

confusion refusals as a result. One commenter noted that “AI would not itself change the legal

standard for registrability. However, the AI software may change the way the legal standard

is applied to analyze the availability of a particular mark.”

145

A majority of commenters insisted that AI software should be used only to supplement

human examiners’ searches, not to replace their searches or make registrability decisions.

The examination and infringement test of likelihood of confusion is based on human

perception. Humans are necessary for evaluating search results because “humans inherently

incorporate practical considerations into their arguments and decisions.”

146

The use of AI

without human involvement to “determine the outcome of an evaluation of a likelihood of

confusion may be too rigid and not allow for the subjective relative weights of different

factors that play a role in the outcome,” observed another commenter.

147

Additionally, humans remain necessary to evaluate trademark distinctiveness and

determine whether confusion is likely because, as a few commenters noted, existing AI

software is not good enough to accurately assess confusion. One commenter noted that “AI

at its current level of accuracy does not always make suitable decisions, and there is also the

risk of an AI malfunction resulting in an incorrect conclusion.”

148

A few commenters raised concerns that AI algorithms can be unintentionally biased, leading

to less accurate search results. One commenter noted that “the inherent biases of the

algorithm(s) should be well known and accounted for, and addressed prior to the trademark

granting decision-making process.”

149

One commenter went on to note that human feedback

144

See, e.g., Response from Dr. Dev S. Gangjee (Nov. 8, 2019), at 1.

145

Response from Intel, at 9-10.

146

Response from AIPLA (Jan. 10, 2020), at 8.

147

Response from ABA IPL (Jan. 9, 2020), at 12.

148

Response from JIPA (Jan. 8, 2020), at 1-2.

149

Response from David Branca, at 2.

in developing the AI algorithm(s) is necessary to improve accuracy and avoid bias and that

“human feedback remains important in the implementation phase to ensure that AI search

tools continue to be effective, consistent and accurate.”

150

Moving beyond the USPTO’s use of AI software in searching, one commenter suggested that

the USPTO could use AI software as a training tool to improve consistency in examination by

identifying how similar cases were handled by different examining attorneys.

151

Another

noted that AI software could be useful for assisting USPTO examining attorneys in identifying

altered or fraudulent specimen images submitted to the USPTO in trademark applications.

152

One commenter observed that there is already positive engagement among global trademark

offices in discussions about the use of AI software for examination and administration.

153

Another commenter believed that AI tools (including search tools) should be shared among

global trademark offices to promote consistency and defray costs of the tools for smaller

national offices.

154

That commenter urged that national trademark offices should be

transparent with their customers when they use specific AI tools in trademark examination

and other official proceedings.

Use of AI software by trademark owners

A few commenters believed that AI software would improve the accuracy of trademark

clearance and searching when used by trademark owners and would improve business

decisions by better predicting the risk of registration refusals and third-party objections. One

commenter noted that “AI could be used to more objectively evaluate the risks of adoption

of a mark and in particular the risks of refusal by the USPTO” and that the impact of this risk

assessment could be fewer applications filed.

155

However, one commenter observed that the use of AI trademark tools by large, multinational

companies could diminish the ability of small and medium-sized businesses to protect their

IP, presumably because small and medium-sized businesses do not have the same access to

these predictive tools for making business decisions.

156

Another commenter expressed

concern about small and medium-sized businesses relying on AI software to create a

trademark or to “recommend a certain market to them based on certain parameters that may

not be in there [sic] best interest or not fully informing them of a risk in that market.”

157

150

Response from INTA, at 2.

151

Response from ABA IPL (Jan. 9, 2020), at 12.

152

Id.

153

Response from INTA, at 1-2.

154

Response from IPO (Jan. 10, 2020), at 6.

155

Response from ABA IPL (Jan. 9, 2020), at 11.

156

Response from Alève Mine, at 2.

157

Response from A-CAPP at Michigan State University, at 1.

8. How, if at all, does AI impact trademark law? Is the existing statutory language

in the Lanham Act adequate to address the use of AI in the marketplace?

Section 32 of the Lanham Act, 15 U.S.C. § 1114, provides a civil action for trademark

infringement against “any person” who uses an infringing registered mark without the

consent of the registrant. AI software used in connection with trademark infringement

cannot itself be held liable for an infringing act, as it is not a “person.” But the people who

create or employ AI software in commercial transactions are “persons,” and as such, they can

be indirectly liable for infringement perpetuated through the use of AI software. Over the

last 30 years, courts have demonstrated flexibility in interpreting the U.S. trademark statutes

that assign liability to various actors and intermediaries as the context for infringement has

expanded beyond brick-and-mortar operations to online infringement and counterfeiting,

including, e.g., the use of such tactics as unauthorized use of others’ trademarks, both in

software coding and more visible ways, in efforts to manipulate web search results to

influence human consumer behavior in commercial transactions. The use of AI in

commercial transactions is yet another of those evolving business models.

Most commenters addressing this question noted that either the use of AI software would

have no impact on trademark law or, alternatively, that the existing statutory and common

law framework for trademarks in the United States is sufficiently flexible to address any such

impact. One commenter noted that the focus on voice-activated AI assistants that provide

product suggestions to consumers or order goods for consumers could put more emphasis

on phonetic similarities between marks and correspondingly less emphasis on visual or

connotative similarities.

158

The commenter noted that no change to the statute would be

needed to accommodate this shift toward phonetics; “however, this [shift] could affect the

fact finders’ balancing scale when assessing likelihood of confusion.”

159

Human involvement

In responding to question 8, commenters also noted the need for continued human

involvement in the use of AI software in connection with trademark issues. One commenter

suggested that the Lanham Act should be amended to require that a trademark be

“distinctive to natural people” to be eligible for protection.

160

Another commenter noted that

if an owner used AI software to select a mark, the Lanham Act should expressly require that

the applicant who is using the mark in commerce be a natural or juridical person.

161

In the criminal and civil enforcement arena, one commenter noted that AI software is

currently being used to identify online counterfeit goods and issue automated takedown

notices.

162

This commenter urged that, in these circumstances, the human brand owner be

158

Response from Intel, at 11.

159

Id.

160

Response from Lori Pressman, at 3.

161

Response from IPO (Jan. 10, 2020), at 7.

162

Response from A-CAPP, at 2.

required to review and verify the AI recommendations before sending any AI-recommended

takedown notices.

Transparency as to the use of AI

As to the roles and responsibilities of those creating and using AI software in commercial

transactions, one commenter insisted that online platforms using AI software to assist

consumers should have a responsibility to avoid deceiving consumers.

163

This commenter

cautioned that “the use of AI can add complication through its shift of the retail experience,

making it more predictive, meaning that consumers are shown a preselected grouping of

products based on a variety of set factors,” but that the consumer may be unaware of that

preselection, resulting in potential deception.

164

To alleviate this deception and create

transparency about the use of AI, this commenter suggested that platforms should cobrand

with the AI creator to encourage a relationship between the AI creator and the end

consumer, who is experiencing the results from an AI algorithm.

Liability under trademark law

Several commenters noted that the use of AI software in consumer transactions raises

questions about who will be legally liable for infringement facilitated by the AI software. One

commenter noted that if AI acts on its own and infringes another’s trademark, “it may be

difficult to identify who is the infringing or violating entity or what the infringement or

violation is.”

165

One commenter believes that it should be incumbent on the creator of the AI

to ensure that it is not violating another’s trademark or other intellectual property.

166

One commenter noted directly importing liability standards developed for internet service

providers may not provide a perfect fit because the creation and use of AI algorithms to offer

consumer products raises a number of novel issues.

167

Some commenters thought it

currently unclear how AI could affect the likelihood of confusion test for infringement and

the role of the “average consumer” in that test. If the test of the average consumer loses

relevance for evaluating likelihood of confusion as consumers rely more on AI to make their

purchases, one commenter noted that “brand owners and trademark practitioners may need

to re-evaluate the strength of infringement theories that rely principally on initial interest

and point of sale confusion and instead explore theories of infringement that place greater

emphasis on the harm caused by post purchase confusion.”

168

Other commenters opined that while the AI technology is not a legal person for purposes of

infringement liability, the creator or platform could be liable for purchases of counterfeit

goods facilitated by the AI. One commenter observed that “‘intent’ of the defendant is an

163

Id. at 2-3.

164

Response from A-CAPP, at 2.

165

Response from JIPA (Jan. 8, 2020), at 2.

166

Response from A-CAPP, at 2.

167

Response from INTA, at 2-3.

168

Id. at 3.

important element in question when looking at the Lanham Act” and queried how AI could

affect the intent element of a counterfeiting claim.

169

The commenter asks, “Can AI intend for

something to happen? Is the onus on the user of the AI or the creator of the AI to not ‘intend’

for a counterfeit mark to be used?”

170

To address this uncertainty, this commenter suggested

that the law should impose a rebuttable presumption of willful intent if a person knowingly

provides or feeds false or infringing data into an AI consumer purchasing recommendation

algorithm or withholds information that would preclude the AI from recommending

infringing goods.

171

Another commenter suggested that platforms using AI to recommend consumer purchasing

choices may have a corresponding responsibility to inform consumers of potentially

suspicious (e.g., counterfeit) goods reflected in such recommendations.

172

This commenter

also raised the question as to whether there should be some responsibility for the AI creator

or the platform to achieve some level of “appropriate” accuracy of the predictions that the AI

algorithm makes, either via legal regulation or industry standard(s).

173

AI-generated works

One commenter identified an issue at the intersection of copyright and trademark law that

exists in the non-AI context but that could also arise in the AI environment. Namely, if a

person instructs or uses AI to create works “in the style of” a well-known copyright creator,

and that copyright creator’s name has been used as a source identifier for the copyrighted

works but is now being used to identify the new work, would the Lanham Act, general unfair

competition laws, or right of publicity laws adequately address potential misappropriation

of the source identifier in that situation?

174

Resulting services provided by AI

One commenter observed that AI cannot qualify as an “applicant” for trademark registration

because the Lanham Act contemplates that only legal entities may qualify as applicants.

175

Nonetheless, one commenter noted the AI “may provide services in connection with a

mark.”

176

Other commenters, however, expressed uncertainty about contemplating AI as a

good or a service and expressed a desire for clarification on how, for purposes of Nice

Classification (the international classification of goods and services applied in the

registration of marks by WIPO), the USPTO would classify the underlying service performed

by AI—would it be classified as the underlying service itself or as computer software?

169

Response from A-CAPP, at 3.

170

Id.

171

Id. at 4.

172

Id. at 3.

173

Id.

174

Response from RIAA, at 7.

175

Response from IPO (Jan. 10, 2020), at 7.

176

Id.

Another commenter noted that this question implicates the Federal Circuit’s holding in In re

JobDiva, Inc., which held that even though a service may be performed by a company’s

software, the company itself may well be rendering the service.

177

A few commenters

observed that mark owners that use AI to provide services could encounter unreliable

examination results from examiners, who may encounter difficulty when determining

whether the mark was actually in use in commerce by the owner for the underlying services.

9. How, if at all, does AI impact the need to protect databases and data sets? Are

existing laws adequate to protect such data?

As noted above, data is the foundation for AI.

178

It is in this context that the question of

whether data or datasets should have protection, or already do have protection, is important

to consider. However, it should also be considered that access to datasets, or the lack thereof,

could impact the development of AI.

Databases and datasets are afforded some protection under copyright law, although

copyrightable

179

. That said, the databases or datasets containing the data may be protected

as compilations to the extent their selection and arrangement demonstrate the requisite

level of originality.

180

Of course, databases and datasets used to train an algorithm can be protected as trade

secrets with criminal remedies under the Economic Espionage Act and civil remedies under

the Defend Trade Secrets Act. To be a protectable trade secret, the dataset must derive

independent economic value from not being generally known or readily ascertainable

through proper means. Unlike copyright protection, trade secret protection can extend to

the underlying facts in a dataset.

177

Id.

178

See discussion regarding Question 10 in Part I.

179

Raw” data is used here in its traditional sense to refer to “individual facts, statistics, or

items of information,” see Random House Webster’s College Dictionary 346 (1991), in

which in which copyright does not subsist. Feist Publ'ns, 499 U.S. at 345. (defining ‘raw

data” as “wholly factual information not accompanied by any original written expression “)

The term “data” can be ambiguous and can sometimes refer more broadly to “electronic

information” or recorded information,” potentially encompassing material protected by

regardless of form or the media on which the data is recorded” for purposes of the Open

Government Data Act). “Data” is not defined in the copyright statute. See also responses

from AAP, at 6; and Adobe, at 2.

180

See Feist Publ’ns, 499 U.S. at 349; 17 U.S.C. § 101 (definition of “compilation”). This is

recognized internationally under WTO TRIPS Article 10(2) (“compilations of data or other

material, whether in machine readable or other form, which by reason of the selection or

arrangement of their contents constitute intellectual creations shall be protected as such.”)

Contract law also can be used to protect collected data and restrict access to the dataset

unless the user agrees not to copy or exploit the data commercially. Technical measures,

181

such as password protection on the website housing the database, can be used, and if the

database has any protection by copyright, the Digital Millennium Copyright Act

182

available for unauthorized uses. There are also state causes of action in tort. Lastly, some

commenters proposed that open data licenses may be an appropriate vehicle to facilitate the

sharing of certain data,

183

while other data may be more appropriately controlled by its

owners by traditional licensing methods.

184

These stakeholders can improve the quality of

the datasets before they are used to train the algorithm.

Commenters who answered this question mostly found that existing laws are adequate to

continue to protect AI-related databases and datasets and that there is no need for

reconsidering a sui generis database protection law, such as exists in Europe. Furthermore,

one commenter cautioned “that AI technology is developing rapidly and that any laws

proposed now could be obsolete by the time they are enacted.”

185

One commenter noted that “because databases and data sets may enjoy copyright

protections as compilations, it is likely that existing laws adequately protect them.”

186

Another commenter noted that “ingestion of the database to train the AI must be

permissioned and/or compensated, or otherwise compliant with copyright law.”

187

Many commenters felt that contractual arrangements and trade secrets offer appropriate

protection for databases and datasets under current law. One commenter noted that, as to

data, its members were “satisfied with reliance on licenses and other comparable

mechanisms for authorized access to such contents and data, with emphasis on contractual

freedom to design the terms of access that work best for the parties.”

188

However, another

commenter stated that a contract is only binding as to the two parties “and cannot be

enforceable against a third party, who has improperly acquired the data outside of the

contract.”

189

As for trade secret protection for datasets, one commenter observed that “trade

secret protection could be impractical or impossible for many business models, for example

where the AI-based data is distributed in a product, or where the results produced by the AI

will be made public.”

190

Another commenter noted similarly that with the use and

181

See 17 U.S.C. § 1201 (providing legal protection against circumvention of technological

protection measures).

182

35 U.S.C. § 512.

183

Response from CDT at 5; IBM at 8; ITIF at 9-10.

184

Response from AAP at 6; Response from STM at 1, 4, 6.

185

Response from IPO (Jan. 10, 2020), at 7.

186

Response from ACT, at 6.

187

Response from Association of American Publishers, at 5.

188

Response from the Entertainment Software Association, at 3.

189

Response from JIPA (Jan. 8, 2020), at 3.

190

Response from IPO (Jan. 10, 2020), at 7.

development of AI, “a lot of data will [be] shared between multiple business operators, where

such data will fall outside the category of trade secrets.”

191

Because copyright protection will not cover the data values in a compilation and a

rearrangement of the data would not necessarily be considered a derivative work,

192

one

commenter believed this might drive the need for consideration of sui generis IP rights.

However, this commenter advised caution in creating new IP rights noted that “providing sui

generis database rights similar to those that exist in Europe might not necessarily promote

innovation.”.

193

Other commenters remembered past debates in the United States on the

issue of database protection, one noting that “Congress considered, and rejected, adoption

of a sui generis form of protection for nonoriginal databases modeled on the EU Database

Directive.”

194

Another commenter noted that since that debate in 1996, it “is not aware of

any significant reason for movement toward a general database protection provision on a

sui generis basis.”

195

A smaller number of commenters did suggest a reconsideration of whether additional

protection of datasets and databases could be useful to spur investment in high-quality data

of vetted/assured provenance. One commenter noted an interest in exploring the possible

creation “of sui generis property rights to curated datasets in view of the: 1) reported

increasing efforts toward data curation, 2) fact that curation may not, in some circumstances,

meet the criteria required for copyright protection, 3) benefits of sharing, improving, and

incentivizing the creation of such datasets, including in the context of public-private

partnerships, 4) disadvantages of trade secrets, particularly in an academic environment,

and 5) different infringement protections needed for proprietary data sets that are used in a

‘once and done’ manner to train AI algorithms.”

196

Another commenter suggested that the

United States consider the Japanese model regarding “Protected Data” in the Unfair

Competition Prevention Act, May 2018, which promotes protection for data used for

exchange.

197

Some commenters went outside the IP framework entirely, noting that privacy, product

safety, and anti-discrimination laws could be implicated, and “it may be that access must be

191

Response from JIPA (Jan. 8, 2020), at 3.

192

A compilation is copyrightable only if the preexisting material that is compiled exhibits

originality in its selection, coordination, or arrangement. A derivative work consists of

revisions, annotations, elaborations or other modifications which, as a whole, represent an

original work of authorship. 17 U.S.C. § 101 (definitions of “compilation” and “derivative

work”). A rearrangement of data will not necessarily exhibit sufficient originality to qualify.

193

Response from IPO (Jan. 10, 2020), at 7. The reference to “the EU’s type of protection”

was an allusion to the European Union’s Directive 96/9/EC on the legal protection of

databases, which is also addressed below in the discussion of Question 13.

194

Response from CCIA (Jan. 10, 2020), at 11.

195

Response from CTA, at 5.

196

Response from AUTM (Jan. 10, 2020), at 5-6.

197

Response from JIPA (Jan. 8, 2020), at 3.

given to the public (or a proxy) for review and possible testing (for example of training data

that is suspected of wrongful bias).”

198

10. How, if at all, does AI impact trade secret law? Is the Defend Trade Secrets Act

(DTSA) 18 U.S.C. 1836 et seq., adequate to address the use of AI in the

marketplace?

Businesses and innovators have long used trade secret law as a means of protecting their

valuable IP. Trade secret law is often chosen for protection for IP for a variety of reasons. For

example, trade secrets do not require that the innovator pay what can be considerable up-

front expenses (e.g., filing and legal fees for obtaining patent grants) that might be too

burdensome for small business and individuals. Trade secrets can also be used for

information that would qualify for a patent, but the innovator does not want to disclose the

information. Trade secret law can also protect information that does not qualify for patent

protection at all, such as customer lists. Additionally, trade secrets have no expiration date

so long as the qualifying factors continue to be in force, which has resulted in many iconic

and long-term trade secrets seen in daily life.

Civil trade secret protection and enforcement have three sources in the United States. First,

trade secret law, which developed from our common law tradition of torts. Second, trade

secret protection is available in 49 states (except New York), the District of Columbia, Puerto

Rico, and the U.S. Virgin Islands, through enactment of variations on the Uniform Trade

Secrets Act (UTSA). The UTSA is essentially a statutory enactment of the common law

principles that preceded it and is supported by a considerable body of case law. The third

avenue of protection is fairly recent: the Defend Trade Secrets Act of 2016 (DTSA). The DTSA

codifies additional provisions, at the core of which is the establishment of the first federal

private civil cause of action for trade secret misappropriation. The DTSA does not supersede

the state statutes but rather provides a victim of misappropriation the choice of its venue—

state or federal.

One core principle of trade secret law, as codified in the DTSA, states that:

The term “owner,” with respect to a trade secret

, means the person or entity

in whom or in which rightful legal or equitable title to, or license in, the trade

secret is reposed.

199

Trade secret law does not address how the trade secret is created or by whom, instead

providing for the rights of the owner.

Commenters acknowledged the importance of trade secret law for the protection of IP, with

one stating that “trade secret law may be the only viable protection available to ensure that

bioinformatics and other practical applications of AI in biotechnology are protected forms of

198

Response from ABA IPL (Jan. 9, 2020), at 13-14.

199

18 U.S.C. § 1839(4).

intellectual property.”

200

Another succinctly stated: “Use of AI in the marketplace presents a

variety of considerations for application of trade secret law as other information

technologies but does not by itself warrant change to trade secret law.”

201

Some commenters

were concerned that application of current trade secret protections would have a negative

effect on human due process rights where AI is making decisions that have a legal or

significant effect on an individual.

202

Other commenters addressed issues that might arise in

the future. For example, one commenter observed that “trade secret protection over AI

systems may be imperfect if and/or when transparency is too robustly required to secure

regulatory approval in the life sciences.”

203

Commenting on the risk of forced data sharing,

another reflected: “While any IP system for data should probably tilt toward encouraging

data sharing … this does not mean a data-related IP system should default to no IP rights or

even forced sharing.”

204

Commenters suggested that, in the future, a sui generis form of

protection might be required.

205

Commenters also raised the issue that in the future, AI might make it easier to ascertain a

trade secret without breaking security measures. Specifically, one observed that a practical

effect of using AI technology may be that information (including know-how and insights)

may be more readily discoverable without defeating secrecy measures.

206

There was no

200

Response from Genentech (Jan. 10, 2020), at 8; see also Response from Intel, at 12 (“Trade

secret law is very important for AI technologies, particularly for protecting certain aspects

of AI that are difficult to protect using patents, such as training datasets and computational

architectures of AI systems. Trade secret law may also be important to protect the

implementations of AI technologies ….”); Response from ACT, at 6 (“Should changes to the

DTSA be considered, it is important that trade secret protections are not weakened.”).

201

Response from ABA IPL (Jan. 9, 2020), at 14; see also Response from IPO (Jan. 10, 2020),

at 8 (“An important means for protecting AI innovation will be trade secrets. If properly

enforced, the current trade secret laws in the U.S. (DTSA and various state statutes) suffice

to protect AI-related trade secrets.”); Response from CCIA (Jan. 10, 2020), at 11 (“The

existing framework of trade secrecy laws, consisting of the federal Defend Trade Secrets Act,

the Economic Espionage Act, and state trade secret law, is adequate to address the use of AI

in the marketplace.”).

202

See, e.g., Response from Wikimedia Foundation, at 8 (“concern about the application of IP

laws to prevent oversight into algorithmic decision-making, particularly where that decision

making will have a substantial effect on people’s lives …”).

203

Response from Genentech (Jan. 10, 2020), at 8-9.

204

Response from ITIF, at 13.

205

See, e.g., Response from Genentech (Jan. 10, 2020), at 9 (“If trade secret protection is too

difficult to maintain over AI systems for our industry, the Defend Trade Secrets Act would,

of course, be inadequate to address the use of AI in the marketplace. If trade secret protection

is insufficient, some alternative, sui generis form of protection might be required …”).

206

Response from IBM (Jan. 10, 2019), at 6.

consensus on whether a change to the law would be required to address this issue in the

future.

207

11. Do any laws, policies, or practices need to change in order to ensure an

appropriate balance between maintaining trade secrets on the one hand and

obtaining patents, copyrights, or other forms of intellectual property protection

related to AI on the other?

Very few commenters responded directly to this question. The consensus of the responses

was that the current balance is correct and that no changes are necessary.

208

One urged the

USPTO to proceed cautiously and deliberately.

209

Another questioned whether a sui generis

form of data protection might be needed in the future if the balance tipped “too much” in the

direction of trade secrets.

210

One commenter argued against trade secret protection for AI

because such protection does not align with the university culture of collaboration and

scientific publication.

211

12. Are there any other AI-related issues pertinent to intellectual property rights

(other than those related to patent rights) that the USPTO should examine?

As was the case with Question 11 in Part I, this question was intended to capture any issues

not previously addressed. A prevalent theme in the responses to this question centered on

207

See, e.g., Response from ABA IPL (Jan. 9, 2020), at 14 (“Application of AI to deconstruct a

public-facing (or other legitimately accessed) model or output data would not be actionable

under the DTSA or UTSA.”); Response from Intel, at 12 (“There is …. . . a non-negligible risk

that, in the future, AI itself would weaken the protection of trade secret law because AI may

be used to reverse engineer, or make public, what would have been traditionally protected

by trade secret law.”).

208

See, e.g., Response from CCIA (Jan. 10, 2020), at 11 (“No changes to laws, policies, or

practices are needed in order to maintain the balance between these different forms of

protection. AI does not present unique challenges to the IP system.”); Response from Intel,

at 12 (“Intel is not aware of any laws, policies, or practices that warrant change at this time

in order to ensure an appropriate balance between trade secret protection and other forms

of IP protection.”); Response from ABA IPL (Jan. 9, 2020), at 15 (“The Section has not

identified changes necessary for ‘an appropriate balance’ between maintaining trade secrets

and other intellectual property protection related to AI.”).

209

Response from Genentech (Jan. 10, 2020), at 9 (“We urge the USPTO and the Copyright

Office to proceed cautiously and deliberately so that innovation in AI, including exploration

of AI in the life sciences field is incentivized, and not inadvertently left unprotected.”).

210

Response from AIPLA (Jan. 10, 2020), at 16 (“Relying solely on trade secret protection

could tip the balance too far away from disclosure and thereby stifle innovation. The

increasing importance of AI technology highlights the need to further analyze whether and

how to protect data, including a sui generis form of data protection is warranted.”).

211

Response from AUTM (Jan. 10, 2020), at 2 (“Protecting AI innovations using trade secret

strategies does not align with the ethos of university culture.”).

data-related issues.

212

One commenter encouraged the continuation of the USPTO’s long-

standing support for sharing government data with the public.

213

Relatedly, another

highlighted a desire that the “USPTO should examine how ‘data’ is defined in various

government instruments, including trade agreements.”

214

That commenter believed that

doing so could ensure consistent data policies across the U.S. government.

215

Another

expanded the idea of data governance to the global level, pointing to the security of personal

data and the accountability of controllers of such data around the world as an area for future

work.

216

Others raised a concern about the potential for AI to “rapidly generate huge volumes of

IP.”

217

The USPTO should “examine whether the Lanham Act … [is] adequate to guard

against” AI producing “sound alike” works.

218

A separate commenter “strongly encourage[d]

USPTO … to provide mandatory training to staff on AI and its capabilities.”

219

Others raised

ethical issues that the USPTO should explore, such as the impact of deep fakes

220

and of bias

on AI systems

221

and the ramifications should AI be granted personhood.

222

13. Are there any relevant policies or practices from intellectual property IP

agencies or legal systems in other countries that may help inform USPTO’s

policies and practices regarding intellectual property rights (other than those

related to patent rights)?

Commenters pointed to work being done with respect to IP protection for AI technologies in

the European Commission of the EU, at WIPO, in the Organisation for Economic Co-operation

and Development (OECD), at international IP stakeholder associations, and by foreign

governments. Commenters encouraged coordination with foreign government agencies that

are assessing the impact of AI on IP, and with multilateral organizations. One noted that

“there is a need for countries to enact broadly similar (or at least not conflicting) rules and

212

See, e.g., Response from Starrett, at 17 (raising issues of data collection in military

operations that later become available for public use).

213

Response from IBM (Jan. 10, 2019), at 8.

214

Response from AAP, at 6.

215

Id.

216

Response from AIPLA (Jan. 10, 2020), at 16; see also Response from IPO (Jan. 10, 2020),

at 9 (“it would be worth paying careful attention to data privacy as it relates to AI.”);

Response from Pressman, at 4 (advocating for the federal government ensuring data

privacy).

217

Response from Obeebo, at 4; Response from KEI, at 3 (“The potential volume of AI-

generated IP claims is something that should be evaluated very carefully …”).

218

Response from RIAA, at 7.

219

Response from ACT, at 6.

220

Response from ITIF, at 13-15.

221

Response from Aimonetti, at 3.

222

Response from Shore, at 2.

criteria around many IP issues raised by AI and its use of data.”

223

However, another advised

that engagement “should prioritize alignment with U.S. law and precedent to the maximum

extent possible.”

224

One commenter pointed to the “AI Copyright Primer,” created by the commenter’s

organization, the Association Internationale pour la Protection de la Propriété Intellectuelle

(AIPPI) and the International Federation of Intellectual Property Attorneys (FICPI).

225

Another commenter called attention to the Toronto Declaration, “which outlines a human

rights framework to be applied to the use of AI/ML systems.”

226

Another commenter noted

the five OECD AI Principles.

227

With respect to the EU, commenters noted the 2019 Directive on Copyright in the Digital

Single Market. Under article 3 of the directive, non-commercial scientific research using

licensed content for text and data mining is a permitted exception. Other lawfully accessible

online content is also available for short-term mining or extraction under article 4 of the

directive, if the rights holder has not reserved its rights. One commenter observed that “these

provisions provide for a ‘copyright exception’-based approach to the use of content ingested

for AI purposes, and contemplate a viable market for licensing content for commercial AI use

…”.

228

Commenters were mixed as to whether this was a wise approach. One commenter

cautioned generally about text and data mining carve outs because “frameworks that provide

too broad a carve out for un-permissioned and uncompensated uses, even where such use is

of a commercial nature, may result in an eroding of rights accorded to rights holders that

curate and own copyrighted works or compilations of copyrighted works.”

229

Commenters also pointed to the EU Directive 96/9/EC of the European Parliament and the

council of March 11, 1996 on the legal protection of databases.

230

Other commenters noted

problems with that model and its rejection by the U.S. Congress in the mid-1990s.

231

Another

commenter observed that the Database Directive “has had no tangible impact on the

production of databases or the competitiveness of the industry.”

232

Another commenter pointed out the European approach to trade secret protection, which

“provides a broad enough definition of trade secrets such that artificial intelligence

223

Response from ITIF, at 15.

224

Response from ACT, at 7.

225

Response from AIPLA (Jan. 10, 2020), at 17.

226

Response from Wikimedia, at 9 (citing the Toronto Declaration at

https://www.amnesty.org/download/Documents/POL3084472018ENGLISH.PDF

227

Response from KEI, at 4 (citing https://www.oecd.org/going-digital/ai/principles/).

228

Response from Copyright Clearance Center, at 7.

229

Response from AAP, at 7.

230

Response from IBM (Jan. 10, 2019), at 8.

231

See, e.g., Response from IPO (Jan. 10, 2020), at 7.

232

Response from AUTM (Jan. 10, 2020), at 4.

algorithms and processes can be protected.”

233

This approach, according to this commenter,

should “serve as informing future USPTO’s policies and practices regarding Intellectual

Property Rights specific to AI innovation where at the present time employee job hopping is

common.”

234

One commenter also “recommend[ed] the EU’s Report on Liability for Artificial Intelligence

which provides a deep analysis of the liability issues of AI.”

235

Commenters also referenced existing legislation in the United Kingdom, Singapore,

Australia, China, Thailand, Mexico, South Korea, and Japan.

Specifically in the United Kingdom, the Copyrights, Designs and Patents Act of 1988, section

9(3), addresses computer-generated works. Similar provisions exist in Ireland and New

Zealand, according to one commenter.

236

Another noted similar provisions in Hong Kong

(SAR) and India.

237

Similar to specific references to the recent EU directive, commenters’

views were mixed as to whether any of these laws are a useful template for the United States.

provides an example of a well-considered legislative definition of a TDM exception under the

238

A few commenters suggested that the United States consider the

Japanese model regarding “Protected Data” in the Unfair Competition Prevention Act, May

2018, which promotes protection for data used for exchange.

239

South Korea’s relaxation of its definition of trade secret law in the Unfair Competition

Prevention and Trade Secret Protection Act, which made it no longer necessary to take

reasonable efforts to maintain the secrecy of the information, was acknowledged. However,

one commenter noted that “this approach is unlikely to be implemented in the US.”

240

The

same commenter pointed to South Korea as providing “protection of databases defined as

233

Response from AIPLA (Jan. 10, 2020), at 24.

234

Id.

235

Response from Kernochan Center, at 7 (citing Rep. of the Expert Group on Liability and

New Technologies—New Technologies Formation, Liability for Artificial Intelligence and

Other Emerging Digital Technologies (2019)).

236

Response from IBM (Jan. 10, 2019), at 8, noting Copyright, Designs and Patents Act,

1988, c. 48, § 9(3) (U.K.);

28/2000) (Ire.).

237

Response from KEI, at 5.

238

See, e.g., Response from IBM (Jan. 10, 2019), at 8. The database directive covers databases

in which there has been a substantial investment, qualitatively or quantitatively, in

obtaining, verifying, or presenting the contents and makes it unlawful to extract or reutilize

the whole or a substantial part of those contents. Directive 96/9/EC, Art. 7(1).

239

Response from AIPLA (Jan. 10, 2020), at 24-25; Response from JIPA (Jan. 8, 2020), at 3.

240

Response from AIPLA (Jan. 10, 2020), at 25.

compiled matters whose subject matters are systematically arranged or composed, so that

they may be individually approached or retrieved.”

241

One commenter noted that certain aspects of U.S. copyright law are unique to the United

States, and references to other countries’ laws may not be appropriate at this time. For

example, with respect to sound recordings, there are differences in the basic requirements

for copyrightability that could lead to differences in how AI-generated sound recordings are

protected.

242

Overall, the majority of the commenters expressed that the existing copyright,

trademark, and trade secret law framework is sufficiently robust and flexible to adequately

address issues raised by AI.

241

Id. at 20.

242

Therefore, RIAA, for example, believes it is “premature to consider incorporation of

foreign practices at this time, especially with respect to exceptions or limitations to

Appendix I—

Patenting Artificial Intelligence Inventions RFC Response Summary