No need to destroy books in Singapore to legally use them for training AI, say experts

SINGAPORE – Singapore broadly allows lawfully accessed works to be used for artificial intelligence training, sparing AI developers from resorting to controversial workarounds that have attracted backlash overseas, according to legal experts.

Public outrage ensued after Anthropic and Amazon were reported to have allegedly bought, scanned and destroyed books – including rare editions – to feed their data-hungry AI models under the cover of “fair use” in the United States.

By explicitly permitting AI training using legally acquired material, Singapore gives businesses legal certainty, said Tony Yeo, vice-president of the International Association for the Protection of Intellectual Property’s Singapore Group.

Yeo, who is also director at law firm Drew and Napier’s intellectual property department, told The Straits Times: “That certainty puts us on the world map in AI and intellectual property law.”

Having certainty on using content for AI training is important for driving innovation in the technology, said Kent Walker, president of global affairs at search giant Google.

He was speaking at the Global Forum on Intellectual Property at the Marina Bay Sands Expo and Convention Centre on Aug 26.

The comments followed the Aug 26 launch of a public consultation by the Ministry of Law and Intellectual Property Office of Singapore on AI’s impact on copyright and patent regimes, including whether copyrighted works can be used to train AI models.

Due to uncertainty overseas over whether using copyrighted works to train AI violates copyrights, technology companies have been slicing pages off printed books, scanning the pages for AI training and shredding the pages.

So far, one US court has found such destructive scanning by Anthropic, the firm behind the Claude AI chatbot, to be transformative and fair use of the books that does not infringe on copyrights.

Technology companies, in their pursuit of new material to train AI, have turned to buying massive numbers of print books – millions in the case of Anthropic. Rare books are among the titles bought by firms.

Tech news site 404 Media reportedly used an Apple Airtag tracker and traced a shipment of rare books to an Amazon AI training facility that engages in destructive scanning.

Law experts said that amendments made to Singapore’s Copyright Act in 2021 are clear that copyrighted works can be used for AI training in the country.

But this is possible only if the works are legally obtained or accessed, said Yeo.

This means that if a book or photograph was bought at a store and used for AI training, that is allowed in Singapore.

Google’s Walker said that clarity is what the industry needs.

“A legal regime that required developers to get a commercial licence for every piece of publicly available data used to train (an AI) model would end AI innovation,” he said.

Instead, clear rules for training AI on publicly accessible content are needed, similar to what Singapore, Japan and the European Union have done.

The certainty in Singapore also means that works that can be freely viewed online, such as text and images on a website or social media pages, can be used for AI training, said Oh Pin-Ping, a partner at law firm Bird and Bird.

A legal expert who declined to be named added that rights owners cannot exclude their published works for AI training too, even if they state so.

Still, Yeo said that the Copyright Act does not allow AI models to produce responses that infringe on copyrights, such as prompting an AI to produce an image that is similar to or identical to a company’s copyrighted photo of a football star.

If rights owners do not want their works to be used for AI training, some ways include not letting netizens view them freely online, like putting them behind a paywall, or disabling the ability to download the works, said experts.

However, Oh said there is a cost involved. “If rights owners put their works behind lock and key, they lose eyeballs,” she said.

Oh said that it would likely be less attractive for AI companies to license copyrighted works from individual creators, such as authors and artists, and businesses with a small repertoire of works.

These rights owners could pool their works together and have an organisation represent them collectively and handle the licensing of their works for AI training to ensure each creator is properly compensated. This is similar to what is done in the US.

Giving rights owners more control is important, too.

Speaking to ST on the sidelines of the Global Forum on Intellectual Property, Google’s Walker said that it is key to allow people to opt their works out from AI training easily if they want to.

“If people can opt out, that will result in a market where people have choice,” he said.

“You can either be in the market or not, and we’ll see how that will evolve. But we think that’s the traditional approach that’s worked well on the internet and will work well across AI.”

When asked about the controversy over destructive scanning, Walker pointed to Google’s approach when scanning pages from borrowed library books to digitise them in the Google Books project many years ago.

“For books that were special, we had a technology that would automatically turn the pages one by one, very carefully, very laboriously, so as not to damage the books,” he said.

“Making sure that we are protecting rare books, making sure that the content is honoured in an appropriate way, and that we’re working with people who have the appropriate licences or rights (to the books), such as libraries – that’s the path we took with Google Books,” he added.