Generative AI – Intellectual property cases and policy tracker

This class action complaint has been brought by two professors at SUNY Downstate Health Sciences University against Apple in the US District Court Northern District of California. The Plaintiffs argue that Apple infringed their copyright by reproducing their registered works when amassing databases of training materials and using that data to train its "Apple Intelligence" AI models (including Apple Foundation Models and Apple's OpenELM models). The Complaint argues that Apple reproduced and used data sets that included Books3, a data set of pirated, copyrighted books, and also used 'Applebot' a web-crawling software program that they alleged copied mass quantities of internet data. The Complaint also focuses on the use of unauthorised copies of eBooks sold to users through Apple Books. The Complaint is for direct copyright infringement.

The case has been consolidated with Hendrix v Apple, and Alexander v Apple.

This class action complaint has been issued in the US District Court Northern District of California by two authors (Rebecca Makkai and Jason Reynolds) against MosaicML and its parent company Databricks. Makkai owns registered copyrights in a number of books including The Hundred Year House, while Reynolds owns registered copyrights in books including As Brave as You.

The plaintiffs allege that their copyright works were included in the training dataset for MosaicML Pretrained Transformer (MPT) a series of large language models created by MosaicML and distributed by Databricks (including MPT-7B launched in May 2023, and MPT-30B launched in June 2023).  MosaicML has noted that a large quantity of data in the MPT training datasets comes from a component dataset called "RedPajama – Books". The complaint asserts that this is hosted on the Hugging Face website and its Books component is a copy of the Books3 dataset, which is itself a component of The Pile, which is derived from the Bibliothik shadow library comprising approximately 196,640 books.  The complaint against MosaicML is for direct copyright infringement.  The complaint against Databricks is for vicarious infringement (Databricks having acquired MosaicML in July 2023).

The case as been consolidated with O'Nan v Databricks (which should be tracked for updates) and re-titled In Re Mosaic Litigation.

In this class action filed by three authors against MosaicML (and its parent company Databricks) in the US District Court Northern District of California San Francisco Division, the Plaintiffs have brought a claim of direct copyright infringement relating to the training of MosaicML's Pretrained Transformer (MPT) models including MPT-7B and MPT-30B.  The complaint alleges that the MPTs were trained on a large quantity of data taken from a component dataset called 'RedPajama – Books' which was a dataset hosted on Hugging Face and in respect of which the 'Books' component is a copy of the "Books3 dataset", which is itself a component of The Pile dataset. The complaint also alleges vicarious infringement against Databricks.

The case has been consolidated with the Makkai claim against Databricks and re-titled In Re Mosaic Litigation.

The Plaintiffs filed a Second Amended Consolidated Complaint in January 2026, following leave of the Court. The Defendants' Motion to Dismiss in relation to the complaint concerning DBRX models has been denied. The Plaintiffs have filed a Motion for Partial Summary Judgment on whether the Defendants' downloading of their books from an unauthorised online source and the reproduction and distribution of their books to create and operate a commercial library on public cloud spaces amounted to copyright infringement.

In this class action complaint filed by three authors against Nvidia in the US District Court Northern District of California San Francisco Division, the Plaintiffs have brought a claim of direct copyright infringement against Nvidia relating to its NeMo Megatron LLM series released in September 2022.

The complaint alleges that the Plaintiff's registered copyrights were included in the training dataset used by Nvidia to develop its models. Each of the models is hosted on a website called Hugging Face, with a model card that provides information about the model, including its training dataset, in which it is stated that the model was trained on 'The Pile' dataset prepared by EleutherAI (the complaint therefore alleges that the LLM series was trained on one or more of the Plaintiffs' works).

The case has been consolidated with Dubus v Nvidia.

The Plaintiffs have filed a First Amended Consolidated Complaint, including to assert new claims for vicarious and contributory infringement. Nvidia has filed a Motion to Dismiss certain allegations. The court has dismissed the claim for vicarious liability (with leave to amend) but has allowed all other claims against Nvidia (including the claim of contributory infringement) to proceed.

This class action complaint has been issued in the US District Court Northern District of California by two authors owning registered copyrights in certain books that were alleged to be included in the training dataset Nvidia used to train its NeMo Megatron models, released in September 2022.  The complaint alleges that each of the NeMo Megatron models is hosted on a website called Hugging Face and each has a model card that provides information about the model, including its training dataset – for each of the NeMo Megatron models, the model card states that "the model was trained on 'The Pile' dataset prepared by Eleuther AI" (which includes the Book3 dataset, derived from the Bibliothik shadow library). The complaint is for direct copyright infringement.

The case has been consolidated with Nazemian v Nvidia.

This class action complaint has been brought by a number of authors in the US District Court Southern District of New York against Microsoft relating to its Megatron LLM.  The complaint is for copyright infringement in relation to the reproduction of the Plaintiffs' works by downloading from pirated sources and by training on reproduced copies of works.

Plaintiffs E. Molly Tanzer and Jennifer Gilmore, both authors, have brought a class action complaint against Salesforce, Inc., in the US District Court, Northern District of California, alleging direct copyright infringement.  The Plaintiffs claim that Salesforce unlawfully copied "hundreds of thousands" of copyrighted books without the authors' consent to train its large language models (LLMs), which power the company's Agentforce AI platform. The Complaint specifically identifies two datasets - RedPajama and ThePile - as sources of the allegedly pirated works.

The case has been consolidated with Alexander v. Salesforce, Inc and a Consolidated Amended Complaint filed.

This complaint has been brought in the US District Court of Montana Butte Division by Darius James, an author, who complains about the use of his works in the training set (including the RedPajama dataset) used by Snowflake to train its Arctic LLMs. The complaint is for direct copyright infringement.

This class action complaint was brought in the US District Court Northern District of California against OpenAI. The original Plaintiff is a YouTube user and video creator and the complaint relates to the "surreptitious, non-consensual transcription of millions of YouTube users' videos" to train the Defendants' AI software products. The complaint refers to a New York Times report that claimed Whisper (OpenAI's automatic speech recognition system, released in 2022) is capable of transcribing audio from YouTube videos, and that an OpenAI team had transcribed more than one million hours of videos from YouTube. The claim is for unjust enrichment and unfair competition.

The complaint was amended to bring in a new plaintiff (Ruslana Petryazhna) and to add complaints of breaches of the Massachusetts Unfair and Deceptive Business Practices Act, and for direct copyright infringement.

It is notable that the complaint, as initially filed, did not include one of copyright infringement. It is assumed (as asserted by OpenAI) that this is because there will have been no registrations of some of the works in issue in this case.

In February 2025, the Plaintiffs filed a Statement of Non-Opposition to OpenAI's Motion to Dismiss in relation to the state law claims for unjust enrichment and unfair competition. The statement of non-opposition did not impact the claim for direct copyright infringement. The complaint is now proceeding only in the name of Ruslana Petryazhna as class plaintiff and a Second Amended Complaint has been filed.

The case has been transferred to the Southern District of New York.

This class action complaint was brought in the US District Court Northern District of California against Google/YouTube concerning Google's Gemini products. The original Plaintiff is a YouTube user and video creator. The complaint relates to the "surreptitious, non-consensual transcription of millions of YouTube users' videos" to train the Defendants' AI software products. The complaint refers to a New York Times article that reported that Google had transcribed YouTube videos to harvest text for its language models, having changed its terms of service in 2023. The claim as originally drafted was for unjust enrichment and unfair competition.

The complaint was amended to bring in claims under Massachusetts Unfair and Deceptive Business Practices Act, and for direct copyright infringement (by a new Plaintiff, Ruslana Petryazhna, and on behalf of a copyright class).

In February 2025, the Plaintiffs filed a Statement of Non-Opposition to OpenAI's Motion to Dismiss in relation to the state law claims for unjust enrichment and unfair competition. 

On 30 April 2025, the Plaintiff filed a Notice of Voluntary Dismissal of the complaint, without prejudice.

This class action complaint has been brought in the US District Court Northern District of California against Nvidia (there are separate claims against Google/YouTube and OpenAI – the three cases have been related) concerning the training of Nvidia's Cosmos AI software.  The Plaintiff is a YouTube user and video creator and the complaint relates to the "surreptitious, non-consensual transcription of millions of YouTube users' videos" to train the Defendants' AI software products in violation of YouTube's terms of service and at the expense of video creators.

Nvidia has filed a Motion to Dismiss on the grounds that the Plaintiffs lack standing (for not asserting that the Plaintiff has suffered or will suffer a concrete, particularised injury) and that the claims are pre-empted by the Copyright Act. 

The complaint has been amended to bring in claims under Massachusetts Unfair and Deceptive Business Practices Act, and for direct copyright infringement (by a new Plaintiff, and on behalf of a copyright class). Nvidia has filed a Motion to Dismiss the Amended Claim.

On 24 March 2025, the Plaintiffs voluntarily dismissed their claims against Nvidia, without prejudice.

This complaint has been brought by author Elizabeth Lyon (and a proposed class) against Adobe in the US District Court Northern District of California San Jose Division, relating to Adobe's SlimLM models and the SlimPajama dataset (said to be derived from the RedPajama dataset).  The complaint is for direct copyright infringement. This is the first GenAI complaint brought against Adobe.

This case concerns whether copyright can be registered in a creative work made by artificial intelligence – specifically a piece called 'A Recent Entrance to Paradise' which was created autonomously by an AI tool (the AI tool, Creativity Machine, was created by Dr Thaler who listed the system as the work's creator and himself as the 'Copyright Claimant' as 'a work-for-hire to the owner of the Creativity Machine').

The work was denied registration by the US Copyright Office on the basis there was no human author to support a claim to copyright registration. The proceedings in the US District Court for the District of Columbia sought to overturn the USCO refusal to register. The case was therefore a judicial review hearing of the Copyright Office's decision as a final agency decision.

Following cross motions for summary judgment, on 18 August 2023, Judge Beryl A. Howell issued an Order (and accompanying Memorandum Opinion) dismissing the Plaintiff's motion for summary judgment and granting the Defendants' cross-motion for summary judgment.

The Judge concluded that the Registrar had not acted arbitrarily or capriciously in reaching its conclusion that the copyright registration should be denied.    Thaler's argument is that AI generated works deserve copyright protection as a matter of policy. The Judge said that "copyright has never stretched so far, however, as to protect works generated by new forms of technology absent any guiding human hand … human authorship is a bedrock requirement of copyright".

Dr Thaler filed a Notice of Appeal to the US Court of Appeals for the District of Columbia Circuit. Oral argument was heard by the US Court of Appeals for the DC Circuit on 19 September 2024.

On 18 March 2025, the Court's Opinion (of Circuit Judge Millett) was delivered, affirming the denial of copyright, because the Copyright Act 1976 requires all eligible work to be authored in the first instance by a human author, and Dr Thaler's application had listed the Creativity Machine as the work's sole author. The Court's decision underlines that humanity is a necessary condition for authorship under the Copyright Act, whereas machines are tools used by humans in the creative process.  Adhering to the human-authorship requirement would not impede protection for works made with AI; any line-drawing disagreements as to how much AI had contributed to a particular human author's work were irrelevant here as Dr Thaler had listed the Creativity Machine as the sole author of the work. The Court did not therefore need to deal with the Copyright Office's argument that the Constitution itself requires human authorship of all copyrighted material.

The appellant's petition for rehearing or, in the alternative, rehearing en banc (on the basis that the decision conflicts with Supreme Court precedent on statutory interpretation) was denied.

Dr Thaler filed a petition with the US Supreme Court for a writ of certiorari for a review of the judgment of the US Court of Appeals for the District of Columbia. In the statement of case, he argues that the 1976 Copyright Act does not require a particular sort of traditional human contribution for a work to obtain copyright protection and that instead the Act explicitly allows non-human authorship. However, he argues, the USCO and the US Court of Appeal have "determined that an unwritten, and unclear, rule of human authorship should be read into the statute".  The petition argues that the USCO has created a "chilling effect on anyone else considering using AI creatively". The Supreme Court has however denied certiorari, leaving the D.C. Circuit's ruling in March 2025 intact.

The position on whether content created by AI generators is protectable differs from country to country (as noted below re the position in the UK as compared to the US). We have written about this here. 

This case relates to an image created by the artist Jason Allen titled "Théâtre D'opéra Spatial" by providing hundreds (around 600) iterative text prompts to the AI system Midjourney, which was rejected by the US Copyright Office for there not being a human that had "executed the traditional elements of authorship". Having filed a Complaint and Request for Declaratory Relief in the US District Court for the District of Colorado, Mr Allen has now filed a Motion for Summary Judgment as to the legal issue of whether a work generated by human creativity using AI as a tool is copyrightable. The Motion argues that the USCO's policy is inconsistent with the text and purpose of the Copyright Act but that, even if it did require that a person execute the traditional elements of human authorship, that standard has been met.

The Copyright Office has filed a response and its own Motion for Summary Judgment. It acknowledges for example that Mr Allen entered prompts "at least 624 times" but argues that whilst repeated prompting may demonstrate substantial effort, it also shows that prompters do not determine the expression generated by a system such as Midjourney and users have to keep trying unti it produces an image they deem 'acceptable', a process which it says does not equate to authorship.

This claim has been brought by Getty Images against AI image generator Stability AI in the UK High Court. Getty Images' claim (as summarised in its press release when commencing the claim) was that, through its Stable Diffusion model (under the name DreamStudio), Stability AI had "unlawfully copied and processed millions of images protected by copyright and the associated metadata owned or represented by Getty Images absent a license to benefit Stability AI's commercial interests and to the detriment of content creators". The claims related to copyright infringement, database right infringement, and trade mark infringement and passing off.

Following the trial in June 2025, the Court handed down its decision in November 2025. The High Court rejected Getty Images' secondary copyright infringement claim, finding that Stable Diffusion's model weights - though trained on datasets that included Getty's images - did not constitute 'infringing copies' as they do not store the original copyright works. However, the court did find limited and historic trade mark infringement relating to watermarks appearing in outputs from certain versions of Stable Diffusion. Getty had abandoned two significant claims during the trial: copyright infringement during the training and development process (which had taken place outside of the UK) and copyright infringement in outputs. This means certain important issues relating to GenAI and copyright remain yet to be considered by UK courts. 

We discuss the Court's judgment in more detail in our article: Getty Images v Stability AI: Unpacking the High Court's judgment. The judge has granted Getty Images permission to appeal on the secondary copyright infringement point but has refused permission to appeal to Stability AI on her findings on the trademark infringement claim. Stability will need to therefore seek permission from the Court of Appeal on that point. On the copyright issue, the court noted that the point was both novel and important, with potentially far-reaching ramifications, for AI models, but also other intangible articles such as software more generally.

The German collecting society, GEMA (which represents the interests of around 95,000 members in Germany), issued proceedings against OpenAI in relation to the reproduction of protected song lyrics by German authors. The proceedings were issued in the Munich Regional Court and argue that, when simple prompts are entered into ChatGPT, it reproduces the original song lyrics with which "the system has obviously been trained". GEMA had previously declared an opt-out from text and data mining on behalf of its members in accordance with the provisions in the Digital Single Market Copyright Directive.

This was the first lawsuit filed by a collecting society worldwide against a provider of a genAI system and has therefore been watched very closely. GEMA presented a generative AI licensing model in September 2024 calling for a responsible approach to genAI, including protection of IP, fair participation of creative professionals in value creation, sustainability, and transparency and responsibility from AI providers.

On 11 November 2025, the Court handed down it's decision in favour of GEMA based on its argument of reproduction of song texts in OpenAI's LLMs and also their reproduction in outputs. The Court concluded that memorization meant that the lyrics were reproducibly contained in two OpenAI LLMs (finding that indirect perceptibility is sufficient for reproduction) and that the reproduction of the song lyrics in outputs also encroached on the rights holders' rights. The Court rejected OpenAI's reliance on the text and data mining exception as the training not only extracted information from training data, but also reproduced the works, which did not amount to text and data mining. The Court rejected making a reference to the CJEU but it is likely that OpenAI will appeal. Gema's CEO has said: “The internet is not some kind of self-service buffet, and creative achievements by human beings are not simply templates for use free of charge. Today, we have set a precedent that both protects and clarifies the rights of creative copyright holders: operators of AI tools such as ChatGPT must also comply with copyright law. Today, we successfully defended the livelihoods of music creators.”

Following its claim against OpenAI in relation to reproduction of protected song lyrics, Gema has filed a complaint against Suno Inc in the Munich Regional Court in relation to its core area, licensing of playable music titles. In its complaint, GEMA argues that using simple prompts, the system outputs "obviously infringe copyright, in terms of melody, harmony and rhythm", providing examples such as Mambo No. 5 by Lou Bega and Daddy Cool by Boney M. It has provided sound files in its press release which it argues demonstrates the similarities between the original songs and those produced using Suno.

In its FAQ, GEMA notes that the aim of the lawsuits is to obtain a licence fee for the authors and music publishers who works have been trained (and has drawn up a licence model for these purposes). However, it is not seeking to prevent the use of GEMA works by AI systems in general. Neighbouring rights of performers and producers of sound recordings are also not the subject of the lawsuit.

The case was heard in March 2026, with the judgment due to be delivered on 12 June 2026.

The District Court of Hamburg in Germany was asked to consider infringement arising out of the use of images taken by photographer Robert Kneschke (which had been downloaded from Shutterstock which had terms and conditions prohibiting scraping etc) against LAION, during the creation of its LAION 5B dataset of image-text pairs made available free of charge (LAION is a not for profit organisation). The claim specifically does not cover further acts of training or development of AI models using the data set (by companies such as Stability AI, for example).

The District Court delivered its decision on 27 September 2024. It found that there was an infringement of the Plaintiff's copyright work by reproduction in the creation of the dataset. The Defendant was not entitled to rely upon the defence of temporary reproduction as the act of reproduction was not transient or incidental. However, as a research organisation, the Defendant could rely upon the exception for text and data mining for non-commercial scientific research purposes (as provided for in Article 3 of the Digital Single Market (DSM) Copyright Directive, and implemented in German law) in relation to its acts of scraping and analysis in the creation of the data set. The data set had been published free of charge and made available to researchers in the field of artificial neural networks. It was irrelevant in the assessment of the creation of the data set that it was also used by commercial companies for training and further developing their AI systems.

The Court therefore did not need to decide whether the Defendant could also rely on the general text and data mining exception provided for in Article 4 of the DSM Copyright Directive. Unlike the exception in Art.3, a right holder can opt out of the TDM exception in Art.4 provided that its reservation of rights is in a machine-readable format. Whilst the Court did not need to decide on this issue, it suggested that a reservation of rights written solely in 'natural language' would be 'machine understandable' but this would need to be assessed depending on the technical development at the relevant time of use of the work.

The Higher Regional Court of Hamburg has upheld the lower court's decision but also went further, finding that LAION could also rely upon the TDM exception in Article 4 of the Digital Single Market Copyright Directive, because the Plaintiff was unable to rely upon a reservation of rights expressed in natural language as this did not meet the requirement of machine-readability as at the relevant date in 2021. It is likely this case will be further appealed to the Federal Court of Justice, with potential scope also for questions to be referred to the CJEU.

Three associations acting on behalf of authors and publishers have brought proceedings against Meta in the 3rd Chamber of the Paris Judicial Court arising out of alleged use of copyrighted works, without authorisation of their authors and publishers, in order to train its GenAI model. This is the first action brought in France by rights holders in relation to the training of GenAI models. The plaintiffs demand copyright enforcement and the complete removal of the data repositories used to train the GenAI model.

This claim, brought by a range of leading Canadian media companies and news publishers, has been issued against OpenAI in the Ontario Superior Court of Justice. The claim is for a declaration that the various OpenAI defendants are jointly and severally liable for (i) infringing, authorizing and/or inducing infringement of copyright in various works published on the media companies' websites (ii) engaging in prohibited circumvention of technological protection measures; (iii) breaching the terms of use of the plaintiffs' various websites; and (iv) unjust enrichment at the expense of the plaintiffs.

This is the first case brought against OpenAI in Canada and represents a fresh jurisdiction where it is now facing allegations of copyright infringement and related claims. Proceedings have also been brought in Canada at the British Columbia Supreme Court by the Canadian Legal Information Institute against Caseway AI.

On 12 July 2024, the EU AI Act was published in the Official Journal of the EU.  The Act entered into force on 1 August 2024 and will be fully applicable 24 months after its entry into force, i.e., on 2 August 2026 (though certain provisions will be applicable sooner, and others at 36 months). There are staggered dates for when different parts of the Act will take effect (see our EU AI Act implementation tracker for more information) :

• 6 months after coming into force, provisions concerning banned AI practices take effect (2 February 2025)
• 1 year after coming into force, provisions on penalties, confidentiality obligations and general-purpose AI take effect (2 August 2025)
• 2 years after coming into force, the remaining provisions take effect (2 August 2026)
• 3 years after coming into force, obligations for high-risk AI systems forming a product (or safety component of a product) regulated by EU product safety legislation apply (2 August 2027)

In relation to copyright, the Act contains provisions relating to obligations on general-purpose AI systems around compliance with EU copyright law (including relating to text and data mining and opt-outs under the EU Digital Single Market Copyright Directive) and transparency around content used to train such models in the form of sufficiently detailed summaries. This form is now available and is mandatory for completion for GPAI models put on the market from 2 August 2025 (with grandfathering provisions in place).  We have written about the template, the information to be provided, and how useful this will be for rights holders here. 

The GPAI Code of Practice requires signatories to: 

• Draw up, keep up-to-date and implement a copyright policy
• Reproduce and extract only lawfully accessible copyright-protected content when crawling the World Wide Web
• Identify and comply with rights reservations when crawling the World Wide Web
• Mitigate the risk of copyright-infringing outputs
• Designate a point of contact and enabling the lodging of complaints

Signatories are also required to provide a general description of how the model was trained, including a breakdown of each training stage and key technical/design choices and assumptions, as well as how its training data was sourced.

The Code is voluntary and non-binding, and is intended to assist GPAI providers in compliance with the relevant provisions of the AI Act. It does not supersede or otherwise affect existing requirements to comply with other applicable laws, directives and regulatory frameworks, such as copyright. OpenAI, Anthropic, Google and others have signed the Code, whereas Meta is the outlier stating that Europe is going down "the wrong path" on AI regulation.

The Commission has published a consultation on protocols for reserving rights from TDM under the EU AI Act and the GPAI Code of Practice which seeks to gather evidence on the use of existing rights reservations solutions, and to contribute to discussions to help identify and agree upon opt-out solutions.

The Commission has published the second draft of a Code of Practice on Transparency relating to the obligations under Article 50. The draft covers rules for making and detecting AI-generated and manipulated content, applicable to providers of AI systems, and rules for labelling deepfakes and AI-generated and manipulated text, applicable to deployers of AI systems

The UK Government has published its report and impact assessment on generative AI and copyright. The report follows its consultation at the end of 2024, discussed in our article: UK Government consultation on copyright and AI: A 'win-win'? (publication of the report was also a statutory requirement under the Data (Use and Access) Act 2025).

We have written more about the report in this article: Copyright and AI: The UK Government's report

The main headline in the report is that there will be no immediate reform to UK copyright law, and the Government no longer sees implementation of a text and data mining (TDM) exception with rights holder opt-out as its preferred option (as initially stated in its consultation). Instead, it plans to gather more evidence on how copyright laws are impacting the development of AI in the UK, and engage with industry on issues such as input transparency and technical tools and standards, all whilst keeping an eye on what happens internationally. Whilst rights holders will be pleased that a TDM exception with opt-out is no longer the preferred option, it does not appear to have been entirely abandoned. If the further evidence the Government intends to gather, alongside international developments, suggest that intervention is needed, it seems that a TDM exception with opt out may still be an option on the table. For now, however, the status quo will continue. One issue the Government was particularly required to consider in its report was the treatment of AI models that are trained overseas and put on the market in the UK. The Government seems inclined to let the litigation between Getty Images and Stability AI run its course before giving any answer as to what it intends to do in relation to overseas-trained models (the case will be heard by the Court of Appeal later this year).

On licensing, rights holders and AI companies are largely aligned: licensing is a commercial negotiation between private parties, and the Government should not intervene. Indeed, direct licensing deals and collective frameworks are beginning to emerge though, as always, there is a risk that some will be left behind. The Government therefore sees its role as putting market conditions in place to enable licensing to flourish, including through, as noted above, best practices on input transparency and technical tools and standards.

There are a number of other important topics upon which the Government has not been able to reach a firm position. On these, it says it will work with industry and other experts to develop best practices, whilst also monitoring international developments and keeping the need for regulation under review. These include in relation to input and output transparency. Meanwhile, development of technical tools and standards is seen as largely a matter for the market to progress, with the Government playing a supporting role. It does appear that the current protection that is available for computer-generated works (CGWs) will be removed but, for now, the Government will again monitor its use and impact. Finally, the Government is clearly alive to concerns about problematic uses of individuals' likenesses. It therefore wishes to understand the case for giving individuals, from public figures and creative professionals to the general public, better control over how their likeness, voice or personality can, and cannot, be digitally replicated, whilst also recognising that AI innovation can bring opportunities. As such, it will introduce a further consultation this summer on how to address deepfake harms whilst protecting legitimate innovation. This will include whether to create a new personality right, raising questions that extend beyond IP law into areas such as data protection and privacy. Individual performers will wish to engage closely with that consultation process.

In March 2023, the US Copyright Office published a Statement of Policy setting out its approach to registration of works containing material generated by AI.

The guidance states that only the human created parts of a generative AI work are protected by copyright. Accordingly, only where a human author arranges AI-generated material in a sufficiently creative way that ‘the resulting work as a whole constitutes an original work of authorship’ or modifies AI-generated content ‘to such a degree that the modifications meet the standard for copyright protection,’ will the human-authored aspects of such works be potentially protected by copyright. 

This statement follows a decision by the USCO on copyright registration for Zarya of the Dawn ('the Work'), an 18-page graphic novel featuring text alongside images created using the AI platform Midjourney. Originally, the USCO issued a copyright registration for the graphic novel before undertaking investigations which showed that the artist had used Midjourney to create the images. Following this investigation (which included viewing the artist’s social media), the USCO cancelled the original certificate and issued a new one covering only the text as well as the selection, coordination, and arrangement of the Work’s written and visual elements. In reaching this conclusion, the USCO deemed that the artist’s editing of some of the images was not sufficiently creative to be entitled to copyright as a derivative work.

As part of its study of the copyright law and policy issues raised by AI systems, in August 2023, the USCO sought written comments from stakeholders on a number of questions:

1. The use of copyrighted works to train AI models 
2. The copyrightability of material generated using AI systems
3. Potential liability for infringing works generated using AI systems
4. Issues related to copyright 

In July 2024, the USCO published Part 1 of its Report on Copyright and Artificial Intelligence, focusing on Digital Replicas (also called 'deepfakes'). Based on the input received, the USCO concluded that a new federal law is needed to deal with unauthorised digital replicas, as existing laws do not provide sufficient legal redress. This would cover all individuals, not just celebrities. However, whilst the paper also notes that creators have concerns over AI outputs that deliberately imitate an artist's style, it does not recommend including style in the coverage of the new legislation at this time.    

In January 2025, the USCO published Part 2 of its report, focused on copyrightability of outputs from using generative AI. The report concludes that outputs can only be protected by copyright where a human author has determined sufficient expressive elements. This can include situations where a human-authored work is perceptible in an output, or a human makes creative arrangements or modifications of the output. However, it will not apply in the case of mere provision of prompts. The report also confirms that the use of AI to assist in the process of creating/including AI-generated material in a larger human-generated work may be protected by copyright.

In May 2025, the USCO published a 'pre-publication' version of Part 3 of its Report, focusing on generative AI training. The report considers the steps involved in creating and deploying a generative AI system which involve using copyrighted works in ways that implicate the right of reproduction including: data collection and creation; training; RAG; and production of outputs.  In relation to fair use, the Office notes that the responses it had received to its Notice of Inquiry were 'sharply divided'.  Given that generative AI involves a spectrum of uses and impacts, the Office notes that it is not possible to prejudge litigation outcomes but does offer the following analysis:

• On the first factor, the Office expresses the view that training a generative AI foundation model on a large and diverse dataset will often be transformative but this will depend on the functionality of the model and how it is deployed. Meanwhile, the use of RAG is less likely to be transformative  where the purpose is to generate outputs that summarise/provide abridged versions of copyrighted works, as opposed to hyperlinks. The USCO forms the view that the knowing use of a dataset consisting of pirated or illegally accessed works should weigh against fair use without being determinative.
• On the fourth factor, the effect of the use upon the potential market for or value of the copyrighted work (the most important factor), the Office identifies that where a model can produce substantially similar outputs that directly substitute for works in the training data, it can lead to lost sales. Even where outputs are not substantially similar, they can dilute the market for similar works in the training data, including by generating material stylistically similar to those works. The assessment of market harm will also depend on the extent to which copyright works can be licensed for AI training.