Seoul AI summit will showcase cooperation, but need for global standards is pressing

21 May 2024
By Frank Hersey

Six months is a long time to talk up artificial intelligence's opportunities and risks, not so long when it comes to establishing hard rules and standards.

After last November's inaugural AI Safety Summit in the UK attracted global leaders and broad publicity, a smaller follow-up event gets under way today in South Korea — but little of regulatory note has changed in the intervening time, bar the formal passing of the EU's AI Act.

There has been a flurry of bipartisan talks around the globe ahead of the meeting, and national safety institutes are starting to test AI models, but is consensus emerging on the technology's more immediate risks? How far away are the standards needed as the building blocks for future cooperation and control?

Tech experts and ministers have converged at the AI Seoul Summit, a halfway stage on the road to a second annual summit to be hosted by France early next year, where global leaders are expected to reconvene.

With less media attention, though, Seoul may lead to more meaningful talks on issues such as standards, while AI risks are likely to be treated as less of a hot potato than at the UK summit, which took a doom-laden view that AI posed imminent risks of catastrophe.

British Prime Minister Rishi Sunak and South Korean President Yoon Suk Yeol are co-hosting, and in a joint newspaper opinion piece this week they pledged to “take the next steps on shaping the global standards that will avoid a race to the bottom.” They are calling, too, on AI companies to “do more to show how they assess and respond to risk.” 

Recent activity

Despite little in the way of regulation, the wheels have been turning on AI safety.

One of the outcomes of November’s summit was to begin a global report into understanding the risks of AI. The first draft of the International Scientific Report on the Safety of Advanced AI was published on Friday.

Experts agree that they don’t yet agree on the major questions of what AI can actually do, what it might be able to do, and how dangerous it is. Many of the scientists involved who have spoken about the report remain unfazed by AI's current trajectory, but most agree much more research is needed.

AI safety centers have popped up around the world since the UK summit, spurring international cooperation. The US, UK, EU, Singapore and Japan all have their own, and South Korea joined the club this week.

The UK and US have made firm commitments to collaborate, with the head of the UK’s AI Safety Institute, Ian Hogarth, describing them as “soldered together.” The UK yesterday announced it was opening a San Francisco offshoot for its institute to attract talent, be near the biggest western AI companies and collaborate more closely with the US. It also stated an agreement to work with Canada as it creates a safety center.

China was an active participant at the UK summit, seen as something of a coup for the organizers. Last week, US and Chinese officials met in Geneva to discuss AI issues including standards. This week, China announced that it is building a regular negotiation mechanism with Russia to cooperate on strengthening AI technologies, and both said they opposed any moves to hinder AI development by other countries.

Top EU officials continued their AI tour to Washington last week, after visiting the UK earlier in the month.

International organizations have all continued to work on AI, from the Group of Seven to the United Nations. Most recently, the Council of Europe adopted the first legally-binding international treaty on AI last week, with an emphasis on human rights and democracy. It had input from observer countries beyond the 46 European nations, including the US and Japan. Signing can start in September, and signatory nations must each ratify it.

The UK's AI safety institute published its first technical report yesterday: The news was not all good. Its testing of four major large-language models — not named — found that they all returned answers with harmful content, and even more so when “jailbreaking” attack techniques were used.

The tests also found that models displayed high levels of accuracy when answering science questions, but struggled with cybersecurity and other research projects made up of multiple tasks. The concept of “agentic AI,” where AI programs autonomously get on with work, is one such threat to both employment and security.

Testing itself is a topic for collaboration among national institutes. Working independently on methodology for testing AI models or analyzing data sources would be “incredibly stupid” in part due to a lack of staff, said EU lawmaker Dragos Tudorache, architect of the EU's AI Act, last month.

Meanwhile, governments are talking more and more about AI and joining dots. For example, existing regulation in areas such as antitrust is kicking in where the possibility of AI cartels looms.

Big AI

The trend for institutional cooperation, however, is perhaps overshadowed by tech companies' evolving products as large AI developers appear to be releasing new models with more risks. OpenAI’s latest version of its ChatGPT app, released last week, is seemingly at odds with the EU AI Act as demonstrations showed it using emotion recognition, a no-no in the EU.

They also don’t seem to have been forthcoming in sharing their models for testing, although the director of the European Commission’s technology arm, Roberto Viola, has cautioned that governments can't expect companies to share entire models due to concerns of industrial secrets and security.

“The more the algorithm is proprietary, the more, of course, you need the deep cooperation of the people that develop this model,” Viola said while in London this month. Companies might share parts of the model, unless ways are developed that one day allow an entire model to be opened up without being left vulnerable to hacking or leaks, he said. 

“At least some companies are making an update to make certain parts of the algorithm available for testing,” Viola said. Unlike with international conformity in areas such as aviation, when it comes to AI models, “you are not yet there because we don't have the standards, we don't have the cooperation of the companies to be willing to do that.”

At the summit today, more major AI companies committed to publishing their safety frameworks, with requirements to determine risk thresholds where they may not develop or deploy a model if they can't avoid the risks it would bring. China's Zhipu.ai and France's Mistral AI joined the existing seven to form a new group of 16 developers.

Conflicting concerns

To be sure, AI-specific legislation and regulation is being drafted around the world, though the UK government, for one, plays down the need and the risk of denting innovation. In their joint article, Sunak and Yoon shifted the emphasis to nurturing small companies, wanting to ensure that “we do not overregulate the start-ups that could produce the next big idea.”

Indeed, catastrophizing over AI's existential risks, prevalent last year, appears to have faded as each software iteration raises concerns around more everyday risks, notably the threat of job losses — itself becoming thought of as a looming disaster. And a slew of major software releases is expected soon. 

Expectations for France’s hosting of the next full safety summit, with the EU's AI Act in the foreground, are for a much stricter line for developers and governments alike. Pressure is building to establish the standards needed to underpin international collaboration and make protections universal, but a lot could happen before France.

For the latest developments in AI regulation and the tech sector, privacy and cybersecurity, online safety, content moderation and more, activate your instant trial of MLex today.