AI models fall short of draft EU rules, researchers say

Receive free Artificial intelligence updates

We’ll send you a myFT Daily Digest email rounding up the latest Artificial intelligence news every morning.

Companies building new artificial intelligence models, including ChatGPT creator OpenAI, Google and Facebook owner Meta, risk falling foul of draft EU rules governing the technology, US research has warned.

The Stanford University paper points to a looming clash between companies spending billions of dollars developing sophisticated AI models, often with the support of politicians who view the technology as central to national security, and global regulators intent on curbing its risks.

“Companies are falling short [of the draft rules], most notably on the topic of copyright,” said Rishi Bommasani, an AI researcher at the Stanford Center for Research on Foundation Models. 

“If foundation models are generating content then they need to summarise which data they trained on is copyrighted,” Bommasani said. “At the moment most providers are doing especially poorly on this.”

The launch of ChatGPT in November prompted the release of a wave of generative AI tools — software trained on massive data sets to produce humanlike text, images and code.

EU lawmakers, spurred on by this breakneck pace of development, recently agreed a tough set of rules governing the use of AI. Under the proposals of the AI Act, developers of generative AI tools such as ChatGPT, Bard and Midjourney would have to disclose content that was generated by AI and publish summaries of copyrighted data used for training purposes, so that creators can be remunerated for the use of their work.

The Stanford study, led by Bommasani, ranked 10 AI models against the EU’s draft rules on describing data sources and summarising copyrighted data, disclosure of the technology’s energy consumption and computing requirements, and reports of evaluations, testing and foreseeable risks associated with it.

Each model fell short in a number of key areas, with six of 10 providers scoring less than 50 per cent. Closed models, such as OpenAI’s ChatGPT or Google’s PaLM 2, suffered from a lack of transparency around copyrighted data, while open-source rivals, or those publicly accessible, were more transparent but harder to control, the researchers found. Ranking bottom on the study’s 48-point scale were Germany’s Aleph Alpha and California-based Anthropic, while the open source BLOOM model ranked top.

“AI is not inherently neutral, trustworthy nor beneficial,” Rumman Chowdhury of Harvard University told a US Congress science, space and technology committee hearing on AI on Thursday.

“Concerted and directed effort is needed to ensure this technology is used appropriately,” she added. “Building the most robust AI industry isn’t just about processors and microchips. The real competitive advantage is trustworthiness.”

The findings from Bommasani’s research, which were cited at Thursday’s hearing, will help regulators globally as they grapple with technology that is expected to shake up industries ranging from professional and financial services to pharmaceuticals and media.

But they also highlighted the tension between rapid and responsible development.

“Our adversaries are catching up” on AI, Frank Lucas, the committee’s Republican chair, said on Thursday. “We cannot and should not try to copy China’s playbook, but we can maintain our leadership role in AI, and we can ensure its development with our values of trustworthiness, fairness, and transparency.”

The US is gearing up to advance legislation in coming months, but the EU’s draft AI Act is further along in terms of adopting specific rules. 

Bommasani said that greater transparency in the sector would enable policymakers to regulate AI more effectively than they have in the past.

“From social media it was clear we didn’t have a good understanding of how platforms were being used, which compromised our ability to govern them,” he said. 

But companies’ non-compliance with the draft AI Act shows that enforcing the laws will be difficult. 

It is not “immediately clear” what it means to summarise the copyrighted portion of these huge data sets, said Bommasani, who expects lobbying efforts in Brussels and Washington to be stepped up as the regulations are finalised.