They trained AI on content without the consent of the authors. Now they will pay $ 1.5 billion. Anthropic contains a settlement with writers
If the judge approves the conditions, It will – as the plaintiffs emphasize – the largest publicly disclosed compensation for copyright violations in history. It is also the first settlement in a series of loud disputes against companies developing generative AI (including OpenAI, Microsoft and Meta) regarding the use of protected content in model training.
Last year, the authors of Andrea Bartz, Charles Graeber and Kirk Wallace Johnson filed a lawsuit. They claimed that Anthropic, supported by capital, among others by Amazon and Alphabet, He illegally used millions of pirate copies of books to teach Claude to answer user questions. Their allegations were part of a wider wave of matters brought by writers, editorial offices and artists who indicate that technology companies stole their work for the needs of AI training.
The website's arguments were predictable: Companies refer to the allowed use (Fair Use), claiming that the models are creating new, transformational content. The creators answer that mass copying their works without consent breaks copyright law.
Check also: Distillation in artificial intelligence. This is what billionaires are afraid of today
Which the court has already determined
In June, Judge William Alsup decided that Anthropic's use of authors' works in the training process itself was within the limits of the permitted use. At the same time he stated violation of rights due to the creation and storage of a “central library” containing over 7 million pirate books – a collection that did not have to be necessary for training purposes.
This decision opened the way to the trial for the amount of compensation, which was to start in December, and potential claims could reach hundreds of billions of dollars.
See also: Ai does not go for your work – he is heading for your whole company. He is increasingly undermining the business models of even large corporations
What contains a settlement
Details were not made public except for $ 1.5 billion. and the fact that the agreement requires approval by the court. Usually, such settlements are determined by the rules of payment for members of the collective claim and possible repair activities (e.g. license procedures or changes in data collection practices), but in this case, apart from the sum of the party, they did not disclose other provisions.
Wider context: what about other matters
The dispute about whether the training of models on protected content without a license is fair use remains open. Shortly after the June Alsupa ruling, another judge in San Francisco, conducting a similar process against the finish line, He suggested that the use of protected materials without consent in many circumstances would be unlawful.
This shows that the case -law is not uniform, and each case depends on the facts: data sources, copying scale, purpose and how to use them.
The signal for the market is simple. The first large settlement in disputes about AI training indicates that companies will be more willing to pay than risking an unpredictable process and potentially draconian compensation.
Read also: Ai will not allow you to retire. “In a decade will double the length of human life”
It also means the growing cost of data. The construction of models may increasingly require a license for large content sets, which will increase the entry barrier and strengthen the position of catalog owners.
Organizations training models They will also have to better document the origin of the data, methods of storage and legal grounds for use – Especially after the criticism of “central libraries” of pirate files.
The court must now approve the settlement. Until then, there is no formal closure of the case. Regardless of the result, other proceedings against model creators (including OpenAI, Microsoft and finish) will go their path – and their decisions can assess the limits of the permitted use differently.
