In May 2025, the U.S. Copyright Office published a groundbreaking, though non-binding, pre-publication report examining whether the use of copyrighted works to train generative artificial intelligence models falls within the doctrine of “fair use.” This is a critical and fundamental question for AI developers, copyright owners, and policymakers.

 

As is well known, AI models are trained on extensive data that often includes content protected by copyright. This raises a key legal question: Is the consent of rights holders required for the use of their works for machine learning purposes? While rights holders and creators argue that permission is necessary, proponents of open access compare machine learning to human learning and argue that free access to information is vital for technological advancement. The new U.S. report seeks to define a regulatory framework for this issue for the first time.

 

Despite being non-binding, the report sparked strong political reactions—most notably, the apparent dismissal of Shira Perlmutter, head of the U.S. Copyright Office, by President Donald Trump just two days after the report’s release. Many view this as a politically motivated act driven by pressure from the tech industry.

 

What Is the “Fair Use” Exception?

The fair use doctrine seeks to balance copyright protections with freedom of expression, innovation, and research. It allows certain uses of protected works without the rights holder’s permission. Fair use typically covers purposes such as self-education, research, criticism, commentary, and journalism. Courts evaluate several factors to determine whether a use is fair, including:

• The purpose and character of the use;
• The nature of the original work;
• The amount and substantiality of the portion used in relation to the work as a whole;
• The effect of the use on the market value of the original work.

 

As part of that analysis, courts may also consider whether the use is “transformative” that is, whether it adds new meaning, context, or purpose to the original work (e.g., parody, academic analysis, or critical commentary). If the new work serves a different function than the original, it may qualify as fair use.

 

Key Findings of the Report

The report addresses whether the extensive use of copyrighted works to train AI systems can be justified under the fair use exception.

 

Overall, the Copyright Office rejected the argument that training AI systems is inherently transformative. The report states that comparing machine processing to human learning is a flawed analogy. It explains that while non-commercial uses—such as academic research—may qualify as fair use, commercial uses aimed at generating AI outputs that imitate or resemble protected works likely do not.

 

The report also highlights the importance of the fourth fair use factor: the effect on the potential market. If AI-generated content substitutes for or diminishes the demand for the original, the use is highly problematic from a copyright perspective. This aligns with the U.S. court decision in Westlaw v. Ross.

 

Five Takeaways from the Report

1. AI Training Is Not Inherently Transformative
   The Office clarifies that training AI models on copyrighted works is not automatically considered transformative; each case must be assessed individually.
2. Commercial Use Is Particularly Problematic
   When training is done for commercial purposes, especially when outputs closely resemble the original works, the likelihood of qualifying as fair use is low.
3. Market Harm Is a Central Factor
   If training data undermines or replaces the market for the original work, this weighs heavily against fair use.
4. Unauthorized Use of Protected Works Is Not Neutral
   The report rejects the notion that large-scale data scraping for AI training is legally neutral, warning of potential copyright violations.
5. Need for Legislative and Judicial Clarification
   The Office calls on lawmakers and courts to define clear boundaries for the use of copyrighted content in AI training, rather than leaving the matter in regulatory uncertainty.

 

The Firing of the Copyright Office Director

Only two days after the report’s release, President Trump dismissed Shira Perlmutter from her position as Director of the U.S. Copyright Office. According to various reports, this move was seen as a direct response to her firm stance on AI and copyright protections.

 

It was claimed that tech industry leaders—reportedly including Elon Musk—lobbied heavily against regulation that would limit companies’ ability to use copyrighted content for training. Her dismissal was therefore viewed as a blow to the independence of the Office and to the system of regulatory checks and balances.

 

The Difference Between the U.S. and Israeli Legal Positions

In 2022, Israel’s Ministry of Justice published a legal opinion on the same issue.

The Israeli position contradicts that of the U.S. Copyright Office. It asserts that, aside from exceptional cases, the use of copyrighted content for training AI systems generally falls under the permitted use provisions of Israeli copyright law. Thus, in most cases, such training does not constitute copyright infringement.

However, the opinion leaves certain key questions unanswered:

• It does not address situations where training data consists of works by a small number of creators, potentially leading to systems that directly compete with those individuals.
• It does not engage with the fact that commercial companies typically use this content to provide paid products and services.
• The opinion analyzes the legality of the training process but remains silent on the legality of the outputs of such AI systems. It states that whether those outputs infringe copyright should be evaluated under existing infringement rules.

 

Conclusion

The copyright and AI debate illustrates the growing tension between creators and technology giants. On one side are artists, writers, and rights holders demanding protection from unauthorized use. On the other are developers seeking the freedom to advance innovation.

Although the U.S. Copyright Office report does not carry legal force, it is likely to influence future litigation, public policy, and legislative action. The message is clear: the era of unchecked AI training on copyrighted content is coming to an end.

The report marks a significant regulatory shift and suggests a new effort to balance technological progress with protection of creative rights. Whether this direction survives political and legal pressures remains to be seen—but for now, it signals a turning point in how copyright law may apply to the AI era.

 

***

 

Ran Vogel, a Partner in our IP group, is available to advise on questions regarding the intersection of IP and AI, as well as all other intellectual property matters.