Training AI on copyrighted content is not infringing, says Trump administration

However, the Trump administration has recognised that others hold a different view to it and said it is for the courts, not legislators, to “resolve this issue”.

“Although the administration believes that training of AI models on copyrighted material does not violate copyright laws, it acknowledges arguments to the contrary exist and therefore supports allowing the courts to resolve this issue,” the US government said in a new national policy framework it issued on AI (4-page / 234KB PDF).

“Similarly, Congress should not take any actions that would impact the judiciary’s resolution of whether training on copyrighted material constitutes fair use,” it added.

The ‘fair use’ concept that exists in US copyright law serves as a limitation on the rights and protections of copyright holders under the legislation. The concept has been interpreted broadly by the US courts in a wide range of contexts, but a definitive interpretation of whether, and if so to what extent, use of copyrighted material in AI training constitutes ‘fair use’, has yet to emerge from the AI-related copyright litigation that has arisen before the US courts.

In its policy paper, the US government did call on Congress to consider how it might enable licensing or collective rights systems to operate to allow rights holders to “collectively negotiate compensation from AI providers”. However, it said any new law in that area “should not address when or whether such licensing is required”.

Intellectual property law expert Alexander Bibi of Pinsent Masons in Frankfurt said: “The US government relies on the flexible fair use doctrine with respect to the use of copyrighted content for training AI models. This means that US copyright law follows a different approach than the European copyright acquis. European law, and thus the law of the member states, is tied to an enumerated and exhaustive list of exceptions and limitations. As a result, European law recognises only the text and data mining exception as a suitable basis for AI training, an exception which, however, is subject to a rights holder opt‑out in cases of commercial use. Consequently, AI training is, in any event, subject to the rightsholder’s control, as the rightsholder may object to such training in a machine‑readable manner.”

“In light of the lack of clear regulations and the flexibility of the fair use doctrine, several court proceedings in the US have already ended in settlements, and a licensing market appears to be emerging, a development that the US government now seeks to support through the use of collective rights‑management mechanisms. This should serve as a model for the EU. The EU and its member states have well‑developed systems of collective rights management and are accustomed, for example in the field of online music rights, to granting EU‑wide licenses. Similar mechanisms are conceivable and desirable for licensing AI training. Extended collective licenses could, for instance, be an appropriate instrument. However, the territoriality of copyright will remain an obstacle. If AI training does not take place within the EU, the only licensable subject matter will be regulation of the output,” he said.

Last week, the UK government confirmed that it will not pursue any immediate changes to UK copyright law relating to AI training and development, despite pressure on it to do so. At the time, the government acknowledged that there is no current consensus over how best to balance the respective interests of UK creatives and AI developers and added that it is uncertain over the impact that its potential interventions in the market might have. It plans to gather more evidence to better understand how copyright law is impacting on AI development and deployment and around the economic effects of reform.

London-based intellectual property law expert Gill Dennis of Pinsent Masons said: “Even though the US has also chosen not to legislate on the issue, it has taken a much more definitive step than the UK by signalling that, in most cases, training AI systems on copyrighted content will be lawful. In practice, the balance has shifted in the US in favour of tech over creativity. Only organisations with significant resources will be able to challenge in court whether their works are being lawfully used and even they can do so only if they know their material is being used – something the current policy paper does not address as it contains no transparency obligations on developers.”

Also in its policy paper, the Trump administration said new US laws concerning some AI outputs should be considered.

“Congress should consider establishing a federal framework protecting individuals from the unauthorised distribution or commercial use of AI-generated digital replicas of their voice, likeness, or other identifiable attributes, while providing clear exceptions for parody, satire, news reporting, and other expressive works protected by the First Amendment,” it said. “Congress should prevent persons from abusing such a framework to stifle free speech online.”