Saudi Arabia’s new copyright law could be a boon for Arabic-language models

A new law in Saudi Arabia explicitly allows using copyrighted materials for AI training, giving the Kingdom an edge as a developer of its own AI models. Saudi Arabia joins Japan, the European Union, and Singapore in having adopted specific copyright carveouts for AI training.

The legal development comes as the Kingdom is vying to become a major player in AI — both as an exporter of computing power and a developer of its own AI models and tools — and competing with the UAE to be the world’s go-to source for Arabic-language AI tools.

The law doesn’t solve all the potential legal problems for Saudi AI models — AI models are increasingly global and copyright regimes vary from country to country, creating potential legal problems for models no matter where they’re trained. But it solves the most pressing legal issues and, just as importantly, it signals that Saudi Arabia is doing serious legal legwork to become a major player in the field.

“It’s not sufficient, but it’s a prerequisite — to at least have a copyright regime that gives you a reasonable argument that you’re not just rampantly infringing bns of people’s rights,” Matthew Sag (LinkedIn), a law professor focused on both AI and copyright issues at Emory University, tells EnterpriseAM.

The move is also likely to be a boon for Arabic-language models and tools. In addition to allowing training with copyrighted Arabic-language works, the law’s implementing regulations lay out a system for Saudi entities to translate works from other languages into Arabic without the explicit consent of the copyright holder — if no Arabic translation has appeared within three years or another specified period. The system isn’t explicitly aimed at model training, but could create a method for producing large quantities of Arabic training content in different dialects.

“Arabic models will be developing at a faster rate because of these developments. This paves the way for Arabic-language models to prosper,” Bachir A. Chakra (LinkedIn), a Riyadh-based attorney with Gowling WLG focused on technology and intellectual property, tells EnterpriseAM. Humain, Saudi’s state-backed AI champion, which is most poised to benefit from the law, did not respond to a request for comment.

Details: The law, which took effect in August, allows developers in Saudi Arabia to train AI models using copyrighted works provided they obtained the work lawfully and don’t use it for anything other than model training. Developers must also log the type, source, purpose, and date of each use.

That limitation also means models can’t reproduce the work after they’re trained on it — something models generally aren’t supposed to do, but sometimes do in error. (As a general matter, AI training is supposed to operate similarly to human learning. When a human reads a novel, it may change how she thinks for years to come, but most humans can’t recite text from novels they’ve read once they’re done.)

International comparison: Laws in the EU, Japan, and Singapore make similar exceptions. In the US, AI companies argue that using copyrighted material to train AI models falls under the “fair use” exception to copyright law — an exception that also allows reproducing portions of copyrighted works in news articles for the purposes of teaching or to satirize them. Some publishers, including The New York Times, disagree and have sued OpenAI, seeking payments for training that uses their copyrighted works.

The trend seems to be that more nations are adopting carveouts or interpreting existing laws to allow training on copyrighted works. “At an international level, what is emerging is a standard of permitted, but always accountable,” Chakra says.

What about the UAE? Current UAE copyright law doesn’t have a carveout for AI model training or a “fair use” doctrine. That hasn’t resulted in any legal problems for the country’s burgeoning AI ecosystem so far, but may not be sustainable — significantly, the nation’s flagship Arabic model, Jais, was trained on servers in California, so the training would fall under US copyright protections.

UAE next? Given the benefits of a specific carveout for model training, Sag predicted the AI-hungry nation would likely update its copyright regime in the next 18 months. The UAE’s Ministry of Economy and Tourism, which oversees copyright protections, did not respond to a request for comment.

Some copyright holders still want a cut. The New York Times sued OpenAI and Microsoft in 2023, claiming the companies violated its copyright by training ChatGPT with its articles. The case hasn’t yet reached trial. In the UK, Getty Images lost a case claiming that Stability AI infringed its copyrights, though its case was weakened because the training happened abroad. The image service is appealing the ruling. “There’s a tension between IP and AI innovation that won’t disappear,” Chakra notes.

There’s also no guarantee that Saudi-trained models won’t run into legal troubles in other nations with different copyright laws. Copyright watchers have warned specifically about a provision of the EU’s AI Act, which states that any company offering a general-purpose AI model in the region must comply with EU copyright law regardless of where the model was trained. Saudi models might fall afoul of the EU act because the Saudi law doesn’t contain the same requirements for allowing copyright holders to opt out of data mining.

And yet: AI is developing at such a rapid pace that it makes little sense for AI companies now to worry about the possibility of country-by-country legal fights over copyright issues years down the road — at a point when some AI companies are likely to be so wealthy that any judgment against them would be meaningless and others are likely to be out of business.

Chakra still urges Saudi tech companies to impose safeguards, including ensuring that their models don’t store data after training and allowing copyright holders to expressly opt out from their content being used for AI training.

Priorities: “If I was a Saudi Arabian AI developer, I would just be so much more worried about, ‘Can I make a good model that’s actually competitive?’” rather than focusing on potential future copyright fights, Sag says, adding, “The next box is: Do you have the people? Do you have the talent? Do you have the electricity? Can you get the chips?”