UK starts review to train AI models on copyrighted content

The United Kingdom is developing measures to regulate the use of copyrighted content by technology companies to train their artificial intelligence models.

The British government on Tuesday launched a consultation aimed at increasing clarity for both the creative industries and AI developers when it comes to how intellectual property is obtained and how AI companies use it for training purposes.

Some artists and publishers are unhappy with the way companies like OpenAI and Google freely scrape their content to train their large language models: AI models trained with huge amounts of data to generate human-like responses.

Large language models are the fundamental technology behind today’s generative AI systems, including OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude.

Last year, the New York Times filed a lawsuit against microsoft and OpenAI accusing companies of infringing its copyright and abusing intellectual property to train large language models.

In response, OpenAI disputed the NYT’s allegations, stating that using open web data to train AI models should be considered “fair use” and that it provides an “opt-out” for rights holders “because it’s the right thing to do.” “

Separately, image distribution platform Getty Images sued another generative AI company, Stability AI, in the United Kingdom, accusing it of scraping millions of images from its websites without consent to train its Stable Diffusion AI model. Stability AI has contested the lawsuit, pointing out that the training and development of its model took place outside the United Kingdom.

Firstly, the consultation will consider making an exception to copyright law for AI training when it is used in the context of commercial purposes, but at the same time allow rights holders to reserve their rights to be able to control use of its content.

Secondly, the consultation will present proposed measures to help creators license and be remunerated for the use of their content by creators of AI models, as well as giving AI developers clarity on what material can be use to train your models.

The government said both the creative industries and technology companies must work harder to ensure that reservation of rights and transparency standards and requirements are effective, accessible and widely adopted.

The government is also considering proposals that would require AI model makers to be more transparent about their model training data sets and how they are obtained so that rights holders can understand when and how their content has been used to train AI. .

This could prove controversial: Tech companies are not particularly forthcoming when it comes to the data that feeds their coveted algorithms or how they train them, given the commercial sensitivities involved in revealing those secrets to potential competitors.

Previously, under former Prime Minister Rishi Sunak, the government attempted to agree a voluntary code of practice on AI copyright.

In a recent interview with CNBC, the head of app development software firm Appian said he believes the UK is well positioned to be the “global leader on this”.

“The U.K. has committed to declaring that it prioritizes personal intellectual property rights,” Matt Calkins, CEO of Appian, told CNBC. He cited the Data Protection Act 2018 as an example of how the UK is “closely associated with intellectual property rights”.

The UK is also not “subject to the same overwhelming lobbying bombardment from national AI leaders as the United States is,” Calkins added, meaning it might not be as likely to bow to pressure from tech giants as politicians. Americans.

“In the United States, anyone who writes AI legislation will hear from Amazon, Oracle, Microsoft or Google even before the bill reaches the floor,” Calkins said.

“That’s a powerful force preventing anyone from writing sensible legislation or protecting the rights of people whose intellectual property is being appropriated wholesale by these major AI players.”

The issue of potential copyright infringement by AI companies is becoming more notable as technology companies move toward a more “multimodal” form of AI—that is, AI systems that can understand and generate content in the form of images and videos, as well as text. .

Last week, OpenAI made its Sora AI video generation model available to the public in the US and “most countries internationally.” The tool allows the user to write a desired scene and produce a high definition video clip.