Meta’s ‘fair use’ defence for ‘training AI with published books won’t work’ in UK, says PA
April 4, 2025 in News by RBN Staff
Source: TheBookseller.com
The “fair use” legal defence used by Meta and Anthropic last week against separate allegations of copyright infringement for training Artificial Intelligence (AI) tools with published books won’t work in UK courts, according to the Publishers Association and Society of Authors.
Last week, legal representatives for Meta told a federal judge in San Francisco that it made “fair use” of books in developing its chatbot Llama, Reuters reported. Meta, which owns Facebook, WhatsApp and Instagram, said that the lawsuit should be thrown out because AI training is protected by the legal doctrine of “fair use”, which it claims allows for the unauthorised use of copyrighted material under certain circumstances.
Four days later, Anthropic asked a Californian federal court to dismiss a case brought by several authors who claimed the business violated their copyrights by using their books to train its chatbot Claude, also on the basis that ingesting books was “fair use”.
Anthropic accepted it had used the authors’ books “as a miniscule part of the corpus of data to train Claude”, but argued that it wasn’t a copyright infringement because its use was “transformative”.
This follows the publication by the Atlantic last week of a searchable database revealing the contents of LibraryGenesis (LibGen), a database of pirated books, which it revealed had been allegedly used to train Meta’s AI LLM Llama 3. The revelations have resulted in a planned protest led by the Society of Authors and a Change.org petition signed by more than 8,000 writers.
The legal principle of “fair use” argues that the use of copyrighted work by AIs is fair because it “transforms” the work and won’t show copyrighted works to end users. “Fair use” sounds similar but is fundamentally different from the defence of “fair dealing”, which is commonly used in UK courts but has a more limited scope. However, the arrival of two successive and high profile “fair use” arguments has led to questions over whether tech firms would be able to use a similar approach in UK courts if British authors were to take legal action.
Catriona MacLeod Stevenson, general counsel and deputy CEO at the Publishers Association, told The Bookseller: “We are watching what happens in these US lawsuits with interest. While ‘fair use’ under US law is broader in nature than the concept of ‘fair dealing’ under UK law, neither amounts to a defence in respect of the massive infringement of copyright-protected works that has taken place in the training of AI models.”
Macleod Stevenson, who is co-chair of the International Publishers Association’s Copyright Committee and is in regular contact with her counterparts in other key territories, added: “Our top priority in the UK is transparency in relation to the works tech companies have used and wish to use. The government must implement transparency mechanisms as a matter of urgency. This would enable the licensing market to further establish itself. Ultimately, it will allow authors and publishers to have control over how their work is used and to be appropriately remunerated.”
Ambre Morvan, senior public policy manager and contracts advisor at the Society of Authors, told The Bookseller: “Under UK law, the scraping of copyright works by tech companies is, to put it simply, unlawful. The UK has a copyright exception for strictly limited extracts of a copyright work if the use is ‘fair dealing’ and for a small handful of specific uses. The fair-dealing criteria include: is the use commercial (rather than for private research, or teaching purposes, for example), does using the new work affect the market for the original work, and is the amount of the original which is being used reasonable and appropriate? If use of the new work acts as a substitute for the original or causes the owner to lose revenue, it is unlikely to be fair.”
She goes on: “The US ‘fair use’ exception is much broader in scope in that it is not limited to a list of permitted acts [as is the case in the UK regarding ‘fair dealing’]. An important factor to assess whether the use is ‘fair use’ is that the use has to be ‘transformative’, i.e., the new use adds something new, with a further purpose or different character. US tech companies are presumably relying on this provision alone, arguing that the use of copyright-protected works by their AI models is transformative. However, there are other legal requirements for fair use to apply and a judge will look at the amount and substantiality of the work used and whether the use could hurt the current market for the original work, and whether the new use is of a commercial nature.”
Morvan added: “In our view and that of the US Authors Guild, and very many others on both sides of the Atlantic, what tech companies are doing is neither ‘fair use’ nor ‘fair dealing’. It is a blatant infringement of copyright.”
Baroness Beeban Kidron, a member of the House of Lords, who recently spearheaded opposition to the UK government’s consultation on allowing a copyright exception to train AIs, told The Bookseller: “The short version is that the US tech companies are trying to ‘stretch’ fair use, which was intended for quoting, inspiration or education, into ingesting. So it is immaterial that they don’t sell books, because they sell/share/use that data to create products based on them.”
She added: “If you ask an AI model/product to create a nursery rhyme in the style of Paul Muldoon, it could not do it, if it had not learnt on the back of ingesting Paul Muldoon’s poetry. Or if you think about it another way: if it was only the fragments that they claim, then the AI companies would only need to ingest the dictionary. If it was only what is in their own products then Meta could get information from its own services. Why they want the copyright content is because it is high quality, has thought and meaning, and that is worth paying for.”
The Bookseller has attempted to contact Anthropic and Meta for comment.
