Copyright Office AI Study on AI-Generated Content

59942

methodology and assumptions used in Labor, 200 Constitution Ave. NW, Room LIBRARY OF CONGRESS S–3323, Washington, DC 20210. the estimate; Copyright Office • Electronic submission: You may • Suggest methods to enhance the submit comments and attachments quality, utility, and clarity of the [Docket No. 2023–6] electronically at http://information to be collected; and www.regulations.gov. Follow the online • Minimize the burden of the Artificial Intelligence and Copyright instructions for submitting comments. information collection on those who are : U.S. Copyright Office, Library AGENCY: to respond, including through the use of FOR FURTHER INFORMATION CONTACT of Congress. Anjanette Suggs, Office of Workers’ appropriate automated, electronic, : Notice of inquiry and request for ACTIONmechanical, or other technological Compensation Programs, Division of comments. collection techniques or other forms of Federal Employees Longshore, and information technology, e.g., permitting Harbor Workers’ Compensation, OWCP/ : The United States Copyright SUMMARYelectronic submission of responses. DFELHWC, at suggs.anjanette@dol.gov Office is undertaking a study of the (email); (202) 354–9660. Background documents related to this copyright law and policy issues raised information collection request are by artificial intelligence (‘‘AI’’) systems. : SUPPLEMENTARY INFORMATIONavailable at https://regulations.gov and To inform the Office’s study and help I. Background at DOL–OWCP/DFELHWC located at assess whether legislative or regulatory steps in this area are warranted, the 200 Constitution Avenue NW, Room S– The Office of Workers’ Compensation Office seeks comment on these issues, 3323, Washington, DC 20210. Questions Programs (OWCP) is the agency including those involved in the use of about the information collection responsible for administration of the copyrighted works to train AI models, requirements may be directed to the Longshore and Harbor Workers’ the appropriate levels of transparency person listed in the Compensation Act (LHWCA), and the FOR FURTHER and disclosure with respect to the use Federal Employees’ Compensation Act section of this notice. INFORMATION of copyrighted works, and the legal (FECA). 33 U.S.C. 939 (LHWCA) and 5 III. Current Actions status of AI-generated outputs. U.S.C. 8104 and 8111 (FECA) authorizes OWCP to pay for approved vocational This information collection request : Written comments are due no DATESrehabilitation services to eligible concerns the Rehabilitation Plan and later than 11:59 p.m. Eastern Time on workers with work-related disabilities. Award, OWCP–16. OWCP/DFELHWC Wednesday, October 18, 2023. Written In order to decide whether to approve has updated the data with respect to the reply comments are due no later than a rehabilitation plan, OWCP must number of respondents, responses, 11:59 p.m. Eastern Time on Wednesday, receive a copy of the plan, supporting burden hours, and burden costs November 15, 2023. vocational testing materials and the supporting this information collection : For reasons of governmental ADDRESSESestimated cost to implement the plan, request from the previous information efficiency, the Copyright Office is using broken down to show the fees, supplies, collection request. the regulations.gov system for the tuition and worker maintenance Type of Review: Extension, without submission and posting of public payments that are contemplated. OWCP change, of a currently approved comments in this proceeding. All also must receive the signatures of the collection. comments should be submitted worker and the rehabilitation counselor electronically through regulations.gov. Agency: Office of Workers’ to show that the worker agrees to follow Specific instructions for submitting Compensation Programs, Division of the proposed plan, and that the comments are available on the Federal Employees’ Longshore, and proposed plan is appropriate. Form Copyright Office website at https://Harbor Workers’ Compensation, OWCP/ OWCP–16 is the standard format for the copyright.gov/policy/artificial- DFELHWC. collection of this information. The intelligence. If electronic submission is OMB Number: 1240–0045. regulations implementing these statutes not feasible, please contact the Office Affected Public: Not-for-profit allow for the collection of information using the contact information below for institutions, Businesses or other for- needed for OWCP to determine if a special instructions. profits. rehabilitation plan should be approved : Number of Respondents: 3,413. FOR FURTHER INFORMATION CONTACTand payment of any related expenses Rhea Efthimiadis, Assistant to the Frequency: On occasion. should be authorized. (LHWCA, 702.506 General Counsel, by email at meft@Number of Responses: 3,413. and 702.507, (FECA, 20 CFR 10.518, copyright.gov or telephone at 202–707– Annual Burden Hours: 1,707 hours. 10.519) 8350. Total Respondent or Recordkeeper : II. Desired Focus of Comments SUPPLEMENTARY INFORMATIONCost: $0. I. Introduction OWCP is soliciting comments OWCP Form 16, Rehabilitation Plan concerning the proposed information and Award. Over the last year, artificial collection (ICR) titled, ‘‘Rehabilitation Comments submitted in response to intelligence (‘‘AI’’) systems and the Plan and Award’’, OWCP–16. OWCP/ this notice will be summarized in the rapid growth of their capabilities have DFELHWC is particularly interested in request for Office of Management and attracted significant media and public comments that: Budget approval of the proposed attention. One type of AI, ‘‘generative • Evaluate whether the collection of information collection request; they will AI’’ technology, is capable of producing information is necessary for the proper become a matter of public record and outputs such as text, images, video, or performance of the functions of the will be available at https:// audio (including emulating a human Agency, including whether the www.reginfo.gov. voice) that would be considered information has practical utility; copyrightable if created by a human • Evaluate the accuracy of OWCP/ Anjanette Suggs, author. The adoption and use of 1DFELHWC’s estimate of the burden Certifying Officer. related to the information collection, [FR Doc. 2023–18668 Filed 8–29–23; 8:45 am] Generative AI technologies produce outputs 1including the validity of the based on ‘‘learning’’ statistical patterns in existing BILLING CODE 4510–CH–P

VerDate Sep<11>2014 17:31 Aug 29, 2023 Jkt 259001 PO 00000 Frm 00079 Fmt 4703 Sfmt 4703 E:\FR\FM\30AUN1.SGM 30AUN1

59943

generative AI systems by millions of companies based on the training process Works (‘‘CONTU’’), which agreed with Americans —and the resulting volume for, and outputs derived from, the Office but declined to discuss the of AI-generated material—have sparked generative AI systems. As concerns and issue in depth because ‘‘[t]he 5widespread public debate about what uncertainties mount, Congress and the development of this capacity for these systems may mean for the future Copyright Office have been contacted by ‘artificial intelligence’ has not yet come of creative industries and raise many stakeholders with diverse views. to pass, and, indeed, it has been significant questions for the copyright The Office has publicly announced a suggested to this Commission that such system. broad initiative earlier this year to a development is too speculative to 3Some of these questions relate to the explore these issues. This Notice is part In the consider at this time.’’ 12scope and level of human authorship, if of that initiative and builds on the intervening years, as AI moved out of any, in copyright claims for material Office’s research, expertise, and prior the realm of speculation, the Office produced in whole or in part by work, as well as information that continued to participate in discussions generative AI. Over the past several stakeholders have provided to the on AI issues, from a 1991 conference years, the Office has begun to receive Office. hosted by the World Intellectual applications to register works II. The Copyright Office’s Past Work on Property Organization (‘‘WIPO’’) to containing AI-generated material, some 13 Machine Learning and AI more recent events the Office co-hosted of which name AI systems as an author with WIPO and with the U.S. Patent or co-author. At the same time, The Copyright Office has long been 14 4 and Trademark Office. copyright owners have brought engaged in questions involving machine 15infringement claims against AI Last year, in two separate copyright learning and copyright. In 1965, the registration matters, the Office publicly Office’s annual report noted that developments in computer technology addressed the question of copyright in data, which may include copyrighted works. Kim Martineau, What is generative AI?, IBM Research had begun to raise ‘‘difficult questions AI-generated material. In the first Blog (Apr. 20, 2023), https://research.ibm.com/ of authorship’’—namely the question of instance, the Office refused to register a blog/what-is-generative-AI (‘‘At a high level, the authorship of works ‘‘‘written’ by claim for two-dimensional artwork generative models encode a simplified computers.’’ As the then-Register of representation of their training data and draw from described as ‘‘autonomously created by 6it to create a new work that’s similar, but not Copyrights observed: a computer algorithm running on a identical, to the original data.’’). The Office has The Office’s Review machine.’’defined ‘‘generative AI’’ and other key terms in a The crucial question appears to be whether 16glossary at the end of this Notice. the ‘‘work’’ is basically one of human See, e,g., Microsoft FY23 Second Quarter authorship, with the computer merely being 2 CONTU was created ‘‘to assist the President Earnings Conference Call Transcript, Microsoft (Jan. 10 an assisting instrument, or whether the and Congress in developing a national policy for 24, 2023), https://www.microsoft.com/en-us/ traditional elements of authorship in the both protecting the rights of copyright owners and Investor/events/FY-2023/earnings-fy-2023-q2.aspx work (literary, artistic, or musical expression (Microsoft CEO Satya Nadella stating that ‘‘[m]ore ensuring public access to copyrighted works when or elements of selection, arrangement, etc.) than one million people have used Copilot to they are used in computer and machine duplication date’’); Krystal Hu, ChatGPT sets record for fastest- were actually conceived and executed not by systems.’’ CONTU, Final Report of the National growing user base—analyst note, Reuters (Feb. 2, Commission on New Technological Uses of man but by a machine. 72023), https://www.reuters.com/technology/chatgpt- Copyrighted Works at 3 (July 31, 1978) (‘‘CONTU Because the answer depends on the sets-record-fastest-growing-user-base-analyst-note- Final Report’’) One of its statutory mandates was to 2023-02-01/. circumstances of a work’s creation, the study ‘‘the creation of new works by the application See, e.g., James Vincent, The scary truth about or intervention of [ ] automatic systems or machine head of the Office’s Examining Division 3AI copyright is nobody knows what will happen reproduction.’’ National Commission on New (and future Register) Barbara Ringer next, The Verge (Nov. 15, 2022), https:// Technological Uses of Copyrighted Works, Public warned that the Office could not ‘‘take www.theverge.com/23444685/generative-ai- Law 93–573, sec. 201(b)(2), 88 Stat. 1873 (1974). the categorical position that registration copyright-infringement-legal-fair-use-training-data CONTU Final Report at 44–46 (recommending 11(discussing the ‘‘key [legal] questions from which the same ‘‘approach [that] is followed by the will be denied merely because a the topic’s many uncertainties unfold’’); see Kevin Copyright Office today in conducting examinations computer may have been used in some Roose & Cade Metz, How to Become an Expert on for determining registrability for copyright of works manner in creating the work.’’ As she A.I., N.Y. Times (Apr. 4, 2023), https:// created with the assistance of computers’’). 8 noted, ‘‘a typewriter is a machine that www.nytimes.com/article/ai-artificial-intelligence- Id. at 44. 12chatbot.html; Kim Martineau, What is generative is used in the creation of a manuscript[,] See U.S. Copyright Office, 94th Annual Report 13AI?, IBM Research Blog (Apr. 20, 2023), https:// but this does not result in the of the Register of Copyrights for the Fiscal Year research.ibm.com/blog/what-is-generative-AI; Ending September 30, 1991, at 2 (1991), https:// manuscript being uncopyrightable.’’ Harvard Online, The Benefits and Limitations of 9 copyright.gov/reports/annual/archive/ar-1991.pdf. This view was echoed a decade later by Generative AI: Harvard Experts Answer Your See Copyright in the Age of Artificial Questions, Harvard Online Blog (Apr. 19, 2023), 14 the National Commission on New Intelligence, U.S. Copyright Office (Feb. 5, 2020), https://www.harvardonline.harvard.edu/blog/ Technological Uses of Copyrighted https://www.copyright.gov/events/artificial- benefits-limitations-generative-ai; Arhan Islam, A intelligence/. History of Generative AI: From GAN to GPT–4, See Copyright law and machine learning for AI: Marktechpost (Mar. 21, 2023), https:// 15 See, e.g., Am. Compl. ¶¶ 8, 61, Getty Images Where are we and where are we going?, U.S. Patent www.marktechpost.com/2023/03/21/a-history-of- 5 (US), Inc. v. Stability AI, Inc., No. 1:23–cv–135, ECF and Trademark Office (Oct. 26, 2021), https://generative-ai-from-gan-to-gpt-4/. Generative AI is No. 13 (D. Del. Mar. 29, 2023) (alleging www.uspto.gov/about-us/events/copyright-law-and- also a point of contention in the labor disputes infringement based on use of copyrighted images to machine-learning-ai-where-are-we-and-where-are- between the Alliance of Motion Picture and train a generative AI model and on the possibility we-going. The Office also supported the U.S. Patent Television Producers and both the Writers Guild of of that model generating images ‘‘highly similar to and Trademark Office when it solicited public America and SAG–AFTRA (the guild representing and derivative of’’ copyrighted images). comments on the impact of AI on intellectual actors and other media professionals). See Andrew U.S. Copyright Office, Sixty-Eighth Annual property policy, including copyright. See U.S. Webster, Actors say Hollywood studios want their 6 Report of the Register of Copyrights for the Fiscal Patent and Trademark Office, Public Views on AI replicas—for free, forever, The Verge (July 13, Year Ending June 30, 1965, at 5 (1966), https://2023), https://www.theverge.com/2023/7/13/ Artificial Intelligence and Intellectual Property www.copyright.gov/reports/annual/archive/ar- Policy (Oct. 2020), https://www.uspto.gov/sites/ 23794224/sag-aftra-actors-strike-ai-image-rights. 1965.pdf. default/files/documents/USPTOAI-Report2020- See U.S. Copyright Office Review Board, 4 Id. 10-07.pdf. Decision Affirming Refusal of Registration of A 7 U.S. Copyright Office, Annual Report of the Recent Entrance to Paradise at 2 (Feb. 14, 2022), U.S. Copyright Office Review Board, Decision 8 16 Examining Division, Copyright Office, for the Fiscal https://www.copyright.gov/rulings-filings/review- Affirming Refusal of Registration of A Recent Year 1965, at 4 (1965), https://copyright.gov/ Entrance to Paradise at 2 (Feb. 14, 2022), https:// board/docs/a-recent-entrance-to-paradise.pdf reports/annual/archive/ar-examining1965.pdf. (noting visual work was submitted listing the www.copyright.gov/rulings-filings/review-board/ author as the ‘‘Creativity Machine’’). docs/a-recent-entrance-to-paradise.pdf. Id. 9

VerDate Sep<11>2014 17:31 Aug 29, 2023 Jkt 259001 PO 00000 Frm 00080 Fmt 4703 Sfmt 4703 E:\FR\FM\30AUN1.SGM 30AUN1

59944

1. The Office’s AI Initiative Boardexplained that the work could ‘author,’ used in both the Constitution and the Copyright Act, excludes non- not be registered because it was made In response to growing humans.’’ In the context of generative ‘‘without any creative input or Congressionaland public interest, 272324AI, this means that ‘‘[i]f a work’s intervention from a human author,’’ and the Office launched a comprehensive AI traditional elements of authorship were that ‘‘statutory text, judicial precedent, Initiative in early 2023. The Initiative produced by a machine, the work lacks and longstanding Copyright Office identified a number of steps that the human authorship and the Office will practice’’ all require human authorship Office would take to further explore the not register it.’’ The Guidance as a condition of copyrightability.The copyright policy questions surrounding 2818instructed applicants seeking to register Office’s registration denial, as well as AI, including hosting public listening works containing more than de minimis the supporting legal analysis, was sessions and publishing a notice of AI-generated material to disclose that recently affirmed in federal district inquiry.At the same time, the Office 25the work contains such material and court. created a website, www.copyright.gov/ 19provide a brief explanation of the A second registration application, ai, to provide information about the human author’s contributions. submitted in 2022, involved a work Initiative, including planned events and 29containing both human authorship and opportunities for public engagement. b. Public Listening Sessions generative AI material. The work was a In April and May 2023, the Office a. March 2023 Registration Guidance graphic novel with text written by the held four public listening sessions to At the outset of the Initiative, the human applicant and illustrations gather input on the copyright issues Office issued a statement of policy created through the use of Midjourney, raised by generative AI. Each session providing registration guidance on a generative AI system. After soliciting focused on a different category of works containing AI-generated material information from the applicant about creative work: literary works, including (‘‘AI Registration Guidance’’).The AI the process of the work’s creation, the 26print journalism and software; works of Registration Guidance reiterated the Office determined that copyright visual art; audiovisual works, including principle that copyright protection in protected both the human-authored text video games; and musical works and the United States requires human and human selection and arrangement sound recordings. Over the four authorship. Under well-established case of the text and images, but not the AI- listening sessions, nearly 90 participants law, the Guidance explained, ‘‘the term generated images themselves.The 20representing individual artists, Office explained that where a human academic experts, legal practitioners, author lacks sufficient creative control granting a narrow exemption after concluding that technology companies, and industry over the AI-generated components of a the specific use as described was likely to be fair associations shared their views with the because it was limited to a ‘‘researcher or group of work, the human is not the ‘‘author’’ of Office. Transcripts, videos recordings, researchers seeking to investigate a particular set of those components for copyright questions that require examination of a large and agendas for each session are purposes.The Office continues to number of works;’’ access to the works in full 21available on the Office’s website. 30receive applications to register works would be limited to researchers solely for purposes of verifying research results; and the researchers incorporating AI-generated material, c. Educational Webinars would not use the works ‘‘for their expressive involving different levels of human In June and July 2023, the Office held purposes.’’ U.S. Copyright Office, Section 1201 contributions. Rulemaking: Eighth Triennial Proceeding to two public webinars on generative AI, 22Determine Exemptions to the Prohibition on each of which drew an audience of Circumvention, Recommendation of the Register of nearly 2,000. The first webinar focused The Review Board is a three-member body that Copyrights 107–13 (Oct. 2021). 17hears administrative appeals of copyright on registration of works containing AI- See Letter from Sen. Chris Coons, Chair, and 23registration decisions. Review Board decisions Sen. Thom Tillis, Ranking Member, Subcomm. on generated material. It included an constitute final agency actions and are subject to Intell. Prop. of the S. Comm. on the Judiciary, to overview of the Office’s general rules on judicial review. See 37 CFR 202.5(f), (g). Kathi Vidal, Under Secretary of Commerce for how to register works containing U.S. Copyright Office Review Board, Decision Intell. Prop. and Director, U.S. Patent and 18material created or owned by someone Affirming Refusal of Registration of A Recent Trademark Office, and Shira Perlmutter, Register of Entrance to Paradise at 3 (Feb. 14, 2022), https:// Copyrights, U.S. Copyright Office (Oct. 27, 2022) other than the applicant, followed by www.copyright.gov/rulings-filings/review-board/ and Letter from Kathi Vidal, Under Secretary of examples illustrating how those rules docs/a-recent-entrance-to-paradise.pdf. Commerce for Intell. Prop. and Director, U.S. Patent apply to works that incorporate AI- Mem. Op., Thaler v. Perlmutter, No. 22–cv– and Trademark Office, and Shira Perlmutter, 19generated material. The second 1564, ECF No. 24 (D.D.C. Aug. 18, 2023). Register of Copyrights, to Sen. Chris Coons, Chair, 31and Sen. Thom Tillis, Ranking Member, Subcomm. webinar convened experts on different U.S. Copyright Office, Cancellation Decision re: 20on Intell. Prop. of the S. Comm. on the Judiciary Zarya of the Dawn (VAu001480196) at 1 (Feb. 21, regions of the world to discuss (Dec. 12, 2022), https://www.copyright.gov/laws/ 2023), https://www.copyright.gov/docs/zarya-of-the- international developments in hearings/Letter-to-USPTO-USCO-on-National- dawn.pdf (letter from the Office to applicant generative AI and copyright law. These Commission-on-AI-1.pdf (Senate letter requesting canceling the original certificate and issuing a new the Office to provide guidance on what the law experts discussed how other countries one covering only the expressive material created around generative AI should be in the future and by the applicant). are addressing copyright issues, the Office’s response explaining that it intended, Id. at 9. including authorship, training, and 21among other things, to issue a notice of inquiry on In addition to registration, the Office has exceptions and limitations. They 22questions involving copyright and AI). considered AI in the regulatory context of the provided an overview of legislative See, e.g., Virtual AI Townhall hosted by Karla section 1201 rulemaking. Section 1201 of the 24Ortiz featuring the U.S. Copyright Office, Concept Copyright Act sets up a triennial proceeding to Art Ass’n (Nov. 2, 2022), https://address possible exceptions to a statutory ban on Id. at 16191. www.conceptartassociation.com/calendar/virtual- 27circumventing technological protection measures Id. at 16192. ai-townhall-featuring-us-copyright-office (event that that control access to copyrighted works. See 17 28featured two senior attorneys from the Office). Id. at 16193. U.S.C. 1201(a)(1)(C) (charging Register of 29Spring 2023 AI Listening Sessions, U.S. Copyright Office Launches New Artificial Copyrights with making recommendation as to 2530Intelligence Initiative, U.S. Copyright Office (Mar. Copyright Office, https://www.copyright.gov/ai/ whether particular users of copyrighted works are 16, 2023), https://www.copyright.gov/newsnet/ listening-sessions.html. adversely affected in ability to engage in 2023/1004.html. noninfringing uses). In the most recent proceeding, The transcript and recording of the registration 31the Register was asked to consider text and data webinar are available at https://www.copyright.gov/ Copyright Registration Guidance: Works 26mining activities as part of this analysis, and she events/ai-application-process/. In the coming Containing Materials Generated by Artificial Intelligence, 88 FR 16190 (Mar. 16, 2023). A copy months, the Office intends to provide further concluded that existing copyright case law did not support the conclusion that all such activity is fair of the guidance is available at https://copyright.gov/ guidance to copyright applicants seeking to register ai/aipolicyguidance.pdf. use. The Register did, however, recommend works containing AI-generated material.

VerDate Sep<11>2014 17:31 Aug 29, 2023 Jkt 259001 PO 00000 Frm 00081 Fmt 4703 Sfmt 4703 E:\FR\FM\30AUN1.SGM 30AUN1

59945

developments and highlighted possible To the extent that commenters believe from artists and performers concerned areas of convergence and divergence. such permission and/or compensation is about generative AI systems’ ability to necessary, the Office seeks their views mimic their voices, likenesses, or styles. d. Engagement With Stakeholders on what kind of remuneration system(s) Although these personal attributes are In addition to the public events might be feasible and effective. The not generally protected by copyright described above, the Office has spoken Office also seeks information regarding law, their copying may implicate with a broad spectrum of stakeholders, the retention of records necessary to varying state rights of publicity and participating in dozens of meetings with identify underlying training materials unfair competition law, as well as have academics, trade groups, individual and the availability of this information relevance to various international treaty creators, technology companies, and to copyright owners and others. obligations. 38creative industries. These meetings On the second issue, the Office seeks 33 V. Overview of Notice have provided valuable information on comment on the proper scope of the technical aspects of generative AI copyright protection for material created The purpose of this Notice is to models and systems, how creators are using generative AI. Although we collect factual information and views using generative AI, and the continuing believe the law is clear that copyright relevant to the copyright law and policy questions copyright applicants have protection in the United States is issues raised by recent advances in about registering works that include AI- limited to works of human authorship, generative AI. The Office undertakes 35generated material. questions remain about where and how this study pursuant to its statutory to draw the line between human mandate in title 17 to ‘‘[c]onduct IV. The Current Inquiry creation and AI-generated content. For studies’’ and ‘‘[a]dvise Congress on Drawing on our prior AI Initiative example, are there circumstances where national and international issues work, including discussions with a human’s use of a generative AI system relating to copyright, other matters stakeholders, the Office has identified a could involve sufficient control over the arising under this title, and related wide range of copyright policy issues technology, such as through the matters.’’ It intends to use this 39arising from the development and use of selection of training materials and information to advise Congress by AI. These relate to: (1) the use of multiple iterations of instructions providing analyses of the current state copyrighted works to train AI models; (‘‘prompts’’), to result in output that is of the law, identifying unresolved (2) the copyrightability of material human-authored? Resolution of this issues, and evaluating potential areas for generated using AI systems; (3) potential question will affect future registration congressional action. The Office will liability for infringing works generated decisions. While the Office is separately also use this record to inform its using AI systems; and (4) the treatment working to update its registration regulatory work and to offer information of generative AI outputs that imitate the guidance on works that include AI- and resources to the public, courts, and identity or style of human artists. The generated material, this Notice other government entities considering 36Office seeks public comments on these explores the broader policy questions these issues. and related issues. related to copyrightability. The questions are grouped into As to the first issue, the Office is On the third question, the Office is several categories. This Notice begins aware that there is disagreement about interested in how copyright liability with several general high-level whether or when the use of copyrighted principles could apply to material questions and then inquires about AI works to develop datasets for training AI created by generative AI systems. For training, including questions of 37models (in both generative and non- example, if an output is found to be transparency and accountability; generative systems) is infringing. This substantially similar to a copyrighted generative AI outputs, including 34Notice seeks information about the work that was part of the training questions of copyrightability, collection and curation of AI datasets, dataset, and the use does not qualify as infringement, and labeling or how those datasets are used to train AI fair, how should liability be apportioned identification of such outputs; and other models, the sources of materials between the user whose instructions issues related to copyright. Because of ingested into training, and whether prompted the output and developers of the importance of using shared language permission by and/or compensation for the system and dataset? in discussing these issues, the questions copyright owners is or should be Lastly, in both our listening sessions are followed by a glossary of key terms required when their works are included. and other outreach, the Office heard for the purposes of this Notice. The Office welcomes input from commenters on the definitions. The transcript and recording of the See Mem. Op., Thaler v. Perlmutter, No. 22– 32 35international webinar are available at https:// cv–1564, ECF No. 24 (D.D.C. Aug. 18, 2023) VI. Instructions and Questions (affirming the Office’s registration denial of AI- www.copyright.gov/events/international-ai- copyright-webinar/. generated work). The Office does not expect that every Additionally, the Office has offered guidance to For example, the Office has received questions 33 36 party choosing to respond to this Notice The Mechanical Licensing Collective (‘‘The MLC’’), about how to apply its guidance that applicants will address every question raised explaining that AI-generated music is not eligible disclose more than de minimis amounts of AI- for the statutory mechanical blanket license in generated material in their works. See AI below. The questions are designed to section 115 of the Copyright Act and that The MLC Registration Guidance, 88 FR at 16193 (explaining gather views from a broad range of should not disburse royalties for such musical that ‘‘AI-generated content that is more than de parties. The Office does request that, works. See Letter from Suzanne V. Wilson, General minimis should be explicitly excluded from the when responding to a question, Counsel and Associate Register of Copyrights, U.S. application’’). Copyright Office, to Kris Ahrend, Chief Exec. commenters clearly identify each Some of these questions are currently before 37Officer, The MLC, at 2–3 (Apr. 20, 2023), https:// the courts in lawsuits that have already been filed www.copyright.gov/ai/USCO-Guidance-Letter-to- over generative AI systems. See, e.g., J.L. v. The-MLC-Letter-on-AI-Created-Works.pdf. See U.S. Copyright Office, Authors, Attribution, Alphabet Inc., 3:23–cv–03340 (N.D. Cal.); Kadrey v. 38 and Integrity: Examining Moral Rights in the United In some cases, a non-generative AI model may Meta Platforms, Inc., 3:23–cv–3417 (N.D. Cal.); 34 States 112–116 (Apr. 2019), https://be trained on copyrighted material. In other cases, Silverman v. OpenAI, Inc., 4:23–cv–3416 (N.D. www.copyright.gov/policy/moralrights/full- the same AI model may be capable of being Cal.); Tremblay v. OpenAI, Inc., 3:23–cv–3223 (N.D. report.pdf (discussing how such interests are deployed in both a generative AI system and a non- Cal.); Getty Images (US), Inc. v. Stability AI, Inc., generally protected under state right of publicity generative one. The Office’s consideration of 1:23–cv–0135 (D. Del.); Andersen v. Stability AI laws). training is framed broadly in order to encompass Ltd., 3:23–cv–0201 (N.D. Cal.); Doe v. GitHub, Inc., these and other situations. 4:22–cv–6823 (N.D. Cal.). 17 U.S.C. 701(b)(1), (b)(4). 39

VerDate Sep<11>2014 17:31 Aug 29, 2023 Jkt 259001 PO 00000 Frm 00082 Fmt 4703 Sfmt 4703 E:\FR\FM\30AUN1.SGM 30AUN1

59946

model, are there other ways to question for which they submit a important a factor is international ‘‘unlearn’’ inferences from training? response, address questions separately, consistency in this area across borders? 5. Is new legislation warranted to 7.4. Absent access to the underlying and provide the factual, legal, or policy address copyright or related issues with dataset, is it possible to identify whether basis for their responses. Commenters generative AI? If so, what should it an AI model was trained on a particular should make clear whether they are entail? Specific proposals and piece of training material? submitting in a personal capacity or on legislative text are not necessary, but the 8. Under what circumstances would behalf of an organization or entity they Office welcomes any proposals or text the unauthorized use of copyrighted are authorized to represent. Commenters for review. works to train AI models constitute fair are particularly encouraged to explain use? Please discuss any case law you any technical understandings that Training believe relevant to this question. inform their legal and policy If your comment applies only to a 8.1. In light of the Supreme Court’s viewpoints, as well as whether their specific subset of AI technologies, recent decisions in Google v. Oracle answers are applicable only to certain please make that clear. industries, technologies, or types of and Andy Warhol America 416. What kinds of copyright-protected copyrighted works. Although some Foundation v. Goldsmith, how should 42training materials are used to train AI questions seek technical information the ‘‘purpose and character’’ of the use models, and how are those materials about generative AI systems, of copyrighted works to train an AI collected and curated? commenters do not need to be affiliated model be evaluated? What is the 6.1. How or where do developers of with a technical entity to answer these relevant use to be analyzed? Do different AI models acquire the materials or questions. stages of training, such as pre-training datasets that their models are trained and fine-tuning, raise different 43General Questions on? To what extent is training material considerations under the first fair use first collected by third-party entities The Office has several general factor? (such as academic researchers or private questions about generative AI in 8.2. How should the analysis apply to companies)? addition to the specific topics listed entities that collect and distribute 6.2. To what extent are copyrighted below. Commenters are encouraged to copyrighted material for training but works licensed from copyright owners raise any positions or views that are not may not themselves engage in the for use as training materials? To your elicited by the more detailed questions training? knowledge, what licensing models are further below. 8.3. The use of copyrighted materials currently being offered and used? 1. As described above, generative AI in a training dataset or to train 6.3. To what extent is non- systems have the ability to produce generative AI models may be done for copyrighted material (such as public material that would be copyrightable if noncommercial or research purposes. 44domain works) used for AI training? it were created by a human author. How should the fair use analysis apply Alternatively, to what extent is training What are your views on the potential if AI models or datasets are later material created or commissioned by benefits and risks of this technology? adapted for use of a commercial developers of AI models? How is the use of this technology nature? Does it make a difference if 456.4. Are some or all training materials currently affecting or likely to affect funding for these noncommercial or retained by developers of AI models creators, copyright owners, technology research uses is provided by for-profit after training is complete, and for what developers, researchers, and the public? developers of AI systems? purpose(s)? Please describe any relevant 2. Does the increasing use or 8.4. What quantity of training storage and retention practices. distribution of AI-generated material materials do developers of generative AI 7. To the extent that it informs your raise any unique issues for your sector models use for training? Does the views, please briefly describe your or industry as compared to other volume of material used to train an AI personal knowledge of the process by copyright stakeholders? model affect the fair use analysis? If so, which AI models are trained. The Office 3. Please identify any papers or how? is particularly interested in: studies that you believe are relevant to 8.5. Under the fourth factor of the fair 7.1. How are training materials used this Notice. These may address, for use analysis, how should the effect on and/or reproduced when training an AI example, the economic effects of the potential market for or value of a model? Please include your generative AI on the creative industries copyrighted work used to train an AI understanding of the nature and or how different licensing regimes do or duration of any reproduction of works could operate to remunerate copyright that occur during the training process, 141 S. Ct. 1183 (2021). owners and/or creators for the use of 41143 S. Ct. 1258 (2023). as well as your views on the extent to 42their works in training AI models. The See Pre-training, Fine-tuning, and Foundation which these activities implicate the 43Office requests that commenters provide Models, GenLaw: Glossary (June 1, 2023), https://exclusive rights of copyright owners. a hyperlink to the identified papers. genlaw.github.io/glossary.html (explaining that pre- 7.2. How are inferences gained from training is a relatively slow and expensive process 4. Are there any statutory or the training process stored or that ‘‘results in a general-purpose or foundation regulatory approaches that have been model’’ whereas fine-tuning ‘‘adapts a pretrained represented within an AI model? adopted or are under consideration in model checkpoint to perform a desired task using 7.3. Is it possible for an AI model to other countries that relate to copyright additional data’’). ‘‘unlearn’’ inferences it gained from and AI that should be considered or For example, the generative AI model, Stable 44training on a particular piece of training Diffusion, was reportedly developed in part by avoided in the United States?How 40material? If so, is it economically researchers at the Ludwig Maximilian University of Munich but is used by the for-profit company feasible? In addition to retraining a Stability AI. See Kenrick Cai, Startup Behind AI For example, several jurisdictions have 40Image Generator Stable Diffusion Is In Talks To adopted copyright exceptions for text and data Raise At A Valuation Up To $1 Billion, Forbes mining that could permit use of copyrighted summary of the use of training data protected under (Sept. 7, 2022), https://www.forbes.com/sites/ material to train AI systems. Separately, the copyright law.’’ See Artificial Intelligence Act, amend. 399, art. 28b(4)(c), EUR. PARL. DOC. P9kenrickcai/2022/09/07/stability-ai-funding-round-1- European Parliament passed its version of the billion-valuation-stable-diffusion-text-to-image/ Artificial Intelligence Act on June 14, 2023, which TA (2023)0236 (2023), https://www.europarl.europa.eu/doceo/document/TA-9-2023-0236?sh=31e11f5a24d6. includes a requirement that providers of generative AI systems publish ‘‘a sufficiently detailed EN.html. 17 U.S.C. 107(1). 45

VerDate Sep<11>2014 17:31 Aug 29, 2023 Jkt 259001 PO 00000 Frm 00083 Fmt 4703 Sfmt 4703 E:\FR\FM\30AUN1.SGM 30AUN1

59947

model be measured? Should the 10.1. Is direct voluntary licensing obtaining appropriate licenses for inquiry be whether the outputs of the AI feasible in some or all creative sectors? training? Who, if anyone, should be 10.2. Is a voluntary collective system incorporating the model responsible for securing them (for compete with a particular copyrighted licensing scheme a feasible or desirable example when the curator of a training work, the body of works of the same approach? Are there existing dataset, the developer who trains an AI 49author, or the market for that general collective management organizations model, and the company employing that class of works? that are well-suited to provide those model in an AI system are different 9. Should copyright owners have to licenses, and are there legal or other entities and may have different affirmatively consent (opt in) to the use impediments that would prevent those commercial or noncommercial roles)? of their works for training materials, or 12. Is it possible or feasible to identify organizations from performing this role? should they be provided with the means the degree to which a particular work Should Congress consider statutory or to object (opt out)? contributes to a particular output from other changes, such as an antitrust 9.1. Should consent of the copyright a generative AI system? Please explain. exception, to facilitate negotiation of owner be required for all uses of 13. What would be the economic collective licenses? copyrighted works to train AI models or impacts of a licensing requirement on 10.3. Should Congress consider only commercial uses? the development and adoption of establishing a compulsory licensing 479.2. If an ‘‘opt out’’ approach were generative AI systems? regime? If so, what should such a 50adopted, how would that process work 14. Please describe any other factors regime look like? What activities should for a copyright owner who objected to you believe are relevant with respect to the license cover, what works would be the use of their works for training? Are potential copyright liability for training subject to the license, and would there technical tools that might facilitate AI models. copyright owners have the ability to opt this process, such as a technical flag or out? How should royalty rates and terms Transparency & Recordkeeping metadata indicating that an automated be set, allocated, reported and service should not collect and store a 15. In order to allow copyright owners distributed? work for AI training uses? to determine whether their works have 10.4. Is an extended collective 489.3. What legal, technical, or practical been used, should developers of AI licensing scheme a feasible or 51obstacles are there to establishing or models be required to collect, retain, desirable approach? using such a process? Given the volume and disclose records regarding the 10.5. Should licensing regimes vary of works used in training, is it feasible materials used to train their models? based on the type of work at issue? to get consent in advance from Should creators of training datasets have 11. What legal, technical or practical copyright owners? a similar obligation? issues might there be with respect to 9.4. If an objection is not honored, 15.1. What level of specificity should what remedies should be available? Are be required? Collective licensing is one alternative to a existing remedies for infringement 49 15.2. To whom should disclosures be direct licensing regime, in which copyright owners appropriate or should there be a made? negotiate and enter into private agreements on an separate cause of action? 15.3. What obligations, if any, should individual basis. Under a collective licensing 9.5. In cases where the human creator arrangement, rights are aggregated and administered be placed on developers of AI systems does not own the copyright—for by a management organization. The management that incorporate models from third organization negotiates the terms of use and example, because they have assigned it parties? distributes payment to participating copyright or because the work was made for hire— 15.4. What would be the cost or other owners. See WIPO, WIPO Good Practice Toolkit for should they have a right to object to an CMOs at 6 (2021), https://www.wipo.int/ impact of such a recordkeeping system AI model being trained on their work? publications/en/details.jsp?id=4561. for developers of AI models or systems, A compulsory or ‘‘statutory’’ license allows for If so, how would such a system work? 50 creators, consumers, or other relevant certain uses of a copyrighted work ‘‘without the 10. If copyright owners’ consent is parties? consent of the copyright owner provided that the required to train generative AI models, person adhered to the provisions of the license, 16. What obligations, if any, should how can or should licenses be obtained? most notably paying a statutorily established there be to notify copyright owners that royalty to the copyright owner.’’ Music Licensing their works have been used to train an Reform: Hearing Before the Subcomm. on Intell. Id. at 107(4). AI model? 46 Prop. of the S. Comm. on the Judiciary, 109th Cong. For example, the European Union’s Directive 17. Outside of copyright law, are there (2005) (statement of Marybeth Peters, Register of 47on Copyright in the Digital Single Market provides Copyrights), http://copyright.gov/docs/ existing U.S. laws that could require for two copyright exceptions or limitations for text regstat071205.html. developers of AI models or systems to and data mining (which may be used in the training ‘‘An Extended Collective Licensing scheme is of generative AI systems): one for purposes of 51 retain or disclose records about the one where a relevant licensing body, subject to scientific research and one for any other purpose. materials they used for training? certain safeguards, is authori[z]ed to license The latter is available only to the extent that specified copyright works on behalf of all rights rightsholders have not expressly reserved their Generative AI Outputs holders in its sector (including non-members), and rights to the use of their works in text and data not just members who have given specific mining. See Directive 2019/790 of the European If your comment applies only to a permission for it to act.’’ Extended Collective Parliament and of the Council of 17 April 2019 on particular subset of generative AI Licensing (ECL) scheme definition, LexisNexis copyright and related rights in the Digital Single technologies, please make that clear. Glossary (2023), https://www.lexisnexis.co.uk/legal/ Market and amending Directives 96/9/EC and 2001/ glossary/extended-collective-licensing-ecl-scheme; 29/EC, 2019 O.J. (L 130), https://eur-lex.europa.eu/ Copyrightability see also Letter from Karyn A. Temple, Acting eli/dir/2019/790/oj. Register of Copyrights, U.S. Copyright Office, to For example, some AI companies have 18. Under copyright law, are there 48 Rep. Robert Goodlatte, Chair, and Rep. John reportedly started to allow copyright owners to tag circumstances when a human using a Conyers, Ranking Member, H. Comm. on the their works as not available for AI training. See generative AI system should be Judiciary (Sept. 29, 2017), https://Emilia David, Now you can block OpenAI’s web www.copyright.gov/policy/massdigitization/house- crawler, The Verge (Aug. 7, 2023), https:// considered the ‘‘author’’ of material letter.pdf; Letter from Karyn A. Temple, Acting www.theverge.com/2023/8/7/23823046/openai- produced by the system? If so, what Register of Copyrights, U.S. Copyright Office, to ¨data-scrape-block-ai; Melissa Heikkila , Artists can factors are relevant to that Sen. Charles Grassley, Chair, and Sen. Dianne now opt out of the next version of Stable Diffusion, determination? For example, is selecting Feinstein, Ranking Member, S. Comm. on the MIT Tech. Review (Dec. 16, 2022), https:// Judiciary (Sept. 29, 2017), https://www.technologyreview.com/2022/12/16/1065247/ what material an AI model is trained on www.copyright.gov/policy/massdigitization/senate- artists-can-now-opt-out-of-the-next-version-of- and/or providing an iterative series of stable-diffusion/. letter.pdf.

VerDate Sep<11>2014 17:31 Aug 29, 2023 Jkt 259001 PO 00000 Frm 00084 Fmt 4703 Sfmt 4703 E:\FR\FM\30AUN1.SGM 30AUN1

59948

text commands or prompts sufficient to 26. If a generative AI system is trained this issue require legislative attention in claim authorship of the resulting on copyrighted works containing the context of generative AI? output? copyright management information, 34. Please identify any issues not 19. Are any revisions to the Copyright how does 17 U.S.C. 1202(b) apply to the mentioned above that the Copyright Act necessary to clarify the human treatment of that information in outputs Office should consider in conducting authorship requirement or to provide of the system? this study. additional standards to determine when 27. Please describe any other issues VII. Glossary of Key Terms content including AI-generated material that you believe policymakers should The Office has included definitions of is subject to copyright protection? consider with respect to potential 20. Is legal protection for AI-generated key terms as they are used in this Notice copyright liability based on AI- material desirable as a policy matter? Is to clarify the technical processes generated output. legal protection for AI-generated involved in generative AI systems. The Labeling or Identification material necessary to encourage following definitions are used for development of generative AI 28. Should the law require AI- purposes of this Notice only; they do technologies and systems? Does existing generated material to be labeled or not necessarily reflect the government’s copyright protection for computer code otherwise publicly identified as being legal position with respect to any that operates a generative AI system generated by AI? If so, in what context particular term. provide sufficient incentives? should the requirement apply and how Artificial Intelligence (AI): A general 20.1. If you believe protection is should it work? classification of automated systems desirable, should it be a form of 28.1. Who should be responsible for designed to perform tasks typically copyright or a separate sui generis right? identifying a work as AI-generated? associated with human intelligence or If the latter, in what respects should 28.2. Are there technical or practical cognitive functions. Generally, AI 55protection for AI-generated material barriers to labeling or identification technologies may use different differ from copyright? requirements? techniques to accomplish such tasks. 21. Does the Copyright Clause in the 28.3. If a notification or labeling This Notice uses the term ‘‘AI’’ in a U.S. Constitution permit copyright requirement is adopted, what should be more limited sense to refer to protection for AI-generated material? the consequences of the failure to label technologies that employ machine Would such protection ‘‘promote the a particular work or the removal of a learning, a technique further defined progress of science and useful arts’’? label? below. 52If so, how? 29. What tools exist or are in AI Model: A combination of computer development to identify AI-generated code and numerical values (or Infringement material, including by standard-setting ‘‘weights,’’ which is defined below) that 22. Can AI-generated outputs bodies? How accurate are these tools? is designed to accomplish a specified implicate the exclusive rights of What are their limitations? task. For example, an AI model may be preexisting copyrighted works, such as designed to predict the next word or Additional Questions About Issues the right of reproduction or the word fragment in a body of text. Related to Copyright derivative work right? If so, in what Examples of AI models include GPT–4, circumstances? 30. What legal rights, if any, currently Stable Diffusion, and LLaMA. 23. Is the substantial similarity test apply to AI-generated material that AI System: A software product or adequate to address claims of features the name or likeness, including service that substantially incorporates infringement based on outputs from a vocal likeness, of a particular person? one or more AI models and is designed generative AI system, or is some other 31. Should Congress establish a new for use by an end-user. An AI system 56standard appropriate or necessary? federal right, similar to state law rights may be created by a developer of an AI 24. How can copyright owners prove of publicity, that would apply to AI- model, or it may incorporate one or the element of copying (such as by generated material? If so, should it more AI models developed by third demonstrating access to a copyrighted preempt state laws or set a ceiling or parties. work) if the developer of the AI model floor for state law protections? What Generative AI: An application of AI does not maintain or make available should be the contours of such a right? used to generate outputs in the form of records of what training material it 32. Are there or should there be expressive material such as text, images, used? Are existing civil discovery rules protections against an AI system audio, or video. Generative AI systems sufficient to address this situation? generating outputs that imitate the may take commands or instructions 25. If AI-generated material is found artistic style of a human creator (such as to infringe a copyrighted work, who an AI system producing visual works independent fixation of other sounds, even though should be directly or secondarily ‘‘in the style of’’ a specific artist)? Who such sounds imitate or simulate those in the liable—the developer of a generative AI should be eligible for such protection? copyrighted sound recording.’’ model, the developer of the system See John S. McCain National Defense What form should it take? 55incorporating that model, end users of Authorization Act for Fiscal Year 2019, Public Law 33. With respect to sound recordings, 115–232, sec. 238(g)(2), 132 Stat. 1636, 1697–98 the system, or other parties? how does section 114(b) of the (2018) (defining ‘‘artificial intelligence’’ to include 25.1. Do ‘‘open-source’’ AI models Copyright Act relate to state law, such systems ‘‘developed in computer software, physical raise unique considerations with respect hardware, or other context that solves tasks as state right of publicity laws? Does 54to infringement based on their requiring human-like perception, cognition, planning, learning, communication, or physical outputs? 53 AI (July 18, 2023), https://ai.meta.com/llama/ action’’). license/ (requiring users of Llama 2 AI model to See James M. Inhofe National Defense 56U.S. Const. art. I, sec. 8, cl. 8. include an attribution notice and excluding use in Authorization Act for Fiscal Year 2023, Public Law 52 services with greater than 700 million monthly Some AI models are released by their 117–263, sec. 7223(4)(A), 136 Stat. 2395, 3669 53 active users). developers for download and use by members of the (2022) (defining ‘‘artificial intelligence system’’ as general public. Such so-called ‘‘open-source’’ Under 17 U.S.C. 114(b), the reproduction and ‘‘any data system, software, application, tool, or 54models may restrict how those models can be used derivative work rights for sound recordings ‘‘do not utility that operates in whole or in part using through the terms of a licensing agreement. See, extend to the making or duplication of another dynamic or static machine learning algorithms or e.g., Llama 2 Community License Agreement, Meta sound recording that consists entirely of an other forms of artificial intelligence’’).

VerDate Sep<11>2014 17:31 Aug 29, 2023 Jkt 259001 PO 00000 Frm 00085 Fmt 4703 Sfmt 4703 E:\FR\FM\30AUN1.SGM 30AUN1

59949

: The U.S. Nuclear Regulatory adams.html. To begin the search, select from a human user, which are SUMMARYsometimes called ‘‘prompts.’’ Examples Commission (NRC) invites public ‘‘Begin Web-based ADAMS Search.’’ For of generative AI systems include comment on the renewal of Office of problems with ADAMS, please contact Midjourney, OpenAI’s ChatGPT, and Management and Budget (OMB) the NRC’s Public Document Room (PDR) Google’s Bard. approval for an existing collection of reference staff at 1–800–397–4209, 301–

Machine Learning: A technique for information. The information collection 415–4737, or by email to building AI systems that is is entitled, ‘‘NRC Form 5, Occupational PDR.Resource@nrc.gov. The supporting statement and NRC Form 5 are available characterized by the ability to Dose Record for a Monitoring Period.’’ automatically learn and improve on the in ADAMS under Accession Nos. : Submit comments by October 30, DATESbasis of data or experience, without ML23082A250 and ML23082A254. 2023. Comments received after this date • NRC’s PDR: The PDR, where you relying on explicitly programmed will be considered if it is practical to do rules. Machine learning involves may examine and order copies of so, but the Commission is able to ensure 57ingesting and analyzing materials such publicly available documents, is open consideration only for comments as quantitative data or text and obtain by appointment. To make an received on or before this date. inferences about qualities of those appointment to visit the PDR, please : You may submit comments ADDRESSESmaterials and using those inferences to send an email to PDR.Resource@nrc.gov by any of the following methods; accomplish a specific task. These or call 1–800–397–4209 or 301–415– however, the NRC encourages electronic inferences are represented within an AI 4737, between 8 a.m. and 4 p.m. eastern comment submission through the model’s weights. time (ET), Monday through Friday, Federal rulemaking website: Training Datasets: A collection of except Federal holidays. • Federal Rulemaking Website: Go to training material (as defined below) that • NRC’s Clearance Officer: A copy of https://www.regulations.gov and search is compiled and curated for use in the collection of information and related for Docket ID NRC–2022–0212. Address machine learning. Examples of training instructions may be obtained without questions about Docket IDs in datasets include BookCorpus, ImageNet, charge by contacting the NRC’s Regulations.gov to Stacy Schumann; and LAION. Clearance Officer, David C. Cullison, telephone: 301–415–0624; email: Training Material: Individual units of Office of the Chief Information Officer, Stacy.Schumann@nrc.gov. For technical material that are used for purposes of U.S. Nuclear Regulatory Commission, questions, contact the individual(s) training an AI model. They may include Washington, DC 20555–0001; telephone: listed in the FOR FURTHER INFORMATIONa combination of text, images, audio, or 301–415–2084; email: section of this document. CONTACTother categories of expressive material, Infocollects.Resource@nrc.gov. • Mail comments to: David C. as well as annotations describing the Cullison, Office of the Chief Information B. Submitting Comments material. An example of training Officer, Mail Stop: T–6 A10M, U.S. material would be an individual image The NRC encourages electronic Nuclear Regulatory Commission, and an associated text ‘‘label’’ that comment submission through the Washington, DC 20555–0001. describes the image. Federal rulemaking website (https:// For additional direction on obtaining Weights: A collection of numerical www.regulations.gov). Please include information and submitting comments, values that define the behavior of an AI Docket ID NRC–2022–0212, in your see ‘‘Obtaining Information and model. Weights are stored within an AI comment submission. Submitting Comments’’ in the model and reflect inferences from the The NRC cautions you not to include section of SUPPLEMENTARY INFORMATIONtraining process. identifying or contact information in this document. comment submissions that you do not : Dated: August 24, 2023. FOR FURTHER INFORMATION CONTACT want to be publicly disclosed in your David C. Cullison, Office of the Chief Suzanne V. Wilson, comment submission. All comment Information Officer, U.S. Nuclear General Counsel and Associate Register of submissions are posted at https:// Regulatory Commission, Washington, Copyrights. www.regulations.gov and entered into DC 20555–0001; telephone: 301–415– Maria Strong, ADAMS. Comment submissions are not 2084; email: Infocollects.Resource@Associate Register of Copyrights and Director routinely edited to remove identifying nrc.gov. of Policy and International Affairs. or contact information. : SUPPLEMENTARY INFORMATION[FR Doc. 2023–18624 Filed 8–29–23; 8:45 am] If you are requesting or aggregating comments from other persons for I. Obtaining Information and BILLING CODE 1410–30–P submission to the NRC, then you should Submitting Comments inform those persons not to include A. Obtaining Information identifying or contact information that NUCLEAR REGULATORY they do not want to be publicly Please refer to Docket ID NRC–2022– COMMISSION disclosed in their comment submission. 0212 when contacting the NRC about Your request should state that comment the availability of information for this [NRC–2022–0212] submissions are not routinely edited to action. You may obtain publicly Information Collection: NRC Form 5, remove such information before making available information related to this Occupational Dose Record for a action by any of the following methods: the comment submissions available to Monitoring Period • Federal Rulemaking Website: Go to the public or entering the comment into

https://www.regulations.gov and search ADAMS. : Nuclear Regulatory AGENCYfor Docket ID NRC–2022–0212. Commission. II. Background • NRC’s Agencywide Documents : Renewal of existing information ACTION In accordance with the Paperwork Access and Management System collection; request for comment. Reduction Act of 1995 (44 U.S.C. (ADAMS): You may obtain publicly Chapter 35), the NRC is requesting available documents online in the public comment on its intention to ADAMS Public Documents collection at See National Artificial Intelligence Initiative 57 request the OMB’s approval for the https://www.nrc.gov/reading-rm/ Act of 2020, 15 U.S.C. 9401(11).

VerDate Sep<11>2014 17:31 Aug 29, 2023 Jkt 259001 PO 00000 Frm 00086 Fmt 4703 Sfmt 4703 E:\FR\FM\30AUN1.SGM 30AUN1