According to TechCrunch, United States law still does not have a clear conclusion as to whether the AI model can be trained in the use of copyright-protected books. Existing decisions show that the courts do not necessarily consider “training” as a violation per se, but the companies concerned may still face substantial compensation if the training materials come from pirated sources.

Anthropic releases complex signals.

The article mentions that last year United States judge William Alsup, in a collective action against writers, demanded that Anthropic pay $1.5 billion in royalties and settlements. However, this decision does not directly negate AI ' s training conduct itself. The focus of the judge's punishment was that Anthropic obtained books from the illegal “shadow library”.

Along the lines of the decision in this case, the model learns a large volume of text, which is closer to “reading and absorbing content” rather than simply reproducing work. The article quotes the intellectual property lawyer, Cathy Gellis, who claims that this finding is more favourable to AI as a whole, since copyright law primarily regulates reproduction and does not directly prohibit reading, use or access to the contents of the work.

The dispute focused on reasonable use.

It was reported that the current copyright legal system in the United States was scheduled to be in 1976 and that, in the face of today ' s generation of AI, the courts could only interpret new issues with old rules. The key to most current cases is whether AI training is “reasonable use”, especially if it is sufficiently “transformative”.

The courts usually consider several factors, including the purpose of the use, the nature of the work, the percentage of use, and the impact on the original market. The article cites counsel Jason Henderson ' s argument that if a company uses a model of content training for others in order to create a directly competitive product, the courts are usually more cautious; if it does not constitute direct competition, there will be more room for support.

Whether or not to make competition is still the focus

The article cites, for example, Thomson Reuters v. Ross Intelligence, where the Court found that the latter ' s use of the former ' s content for the development of the AI legal platform was not justified, inter alia, because of the direct competition between the two products.

By contrast, while writers can argue that chat robots use their work to generate new content, thereby crushing the author ' s income, this argument has not yet won a decisive victory in court. In other words, whether AI exports are sufficient to replace the original author ' s market remains an important argument in subsequent cases.

There's another level of problem with AI generation.

The article also mentions that the question of whether AI training is legal and whether the content generated by AI is copyrighted is in fact two different sets of questions. Previously, in the Thaler v. Perlmutter case, the United States court held that works produced exclusively by AI could not be protected by copyright.

This also raises new realities: if a piece is partially completed by AI, how to define the proportion of human creation may continue to be controversial in the future. TechCrunch argued that, while it would be difficult for United States courts to provide uniform answers in the short term while a large number of cases were still at trial stage, existing decisions had begun to affect AI ' s training and data use.