The AI Safety Summit and Frontier AI: regulating potentially dangerous AI

On 1 November the UK will host a global political summit on AI safety, bringing together politicians, academics, lawmakers and industry experts to discuss the most pressing and dangerous risks that prominent voices in the machine learning community fear are likely to be posed by the next-generation of the most capable large language models.


These particular kinds of potential harms associated with AI are increasingly being referred to under the umbrella term “Frontier AI”.

It is expected that over 100 people from 28 nations will attend, including US Vice President Kamala Harris, Elon Musk and DeepMind CEO, Demis Hassabis among others.

The Summit has certainly set itself ambitious aims: to develop a shared global understanding of the risks posed by Frontier AI, to agree areas for collaboration on AI safety research and to look to build a process for international collaboration on Frontier AI safety, including how best to support national and international frameworks.

We have previously discussed how the UK is positioning itself as a leader in AI safety through the Summit (here). Given that Frontier AI is a relatively new concept, and has emerged recently as a focus of the UK government’s AI policy efforts, this article provides some additional context to help explain Frontier AI. It then considers the focus of other multilateral policy initiatives to tackle potential risks posed by AI.

What is Frontier AI?

In its guidance to the Summit, the government defines Frontier AI as “highly capable general-purpose AI models that can perform a wide variety of tasks and match or exceed the capabilities present in today’s most advanced model.”

Given that the most advanced general-purpose large language models today are OpenAI’s GPT-4, and Google’s PaLM 2, then it follows that the focus of the Frontier AI risk debate is on the successors to these models and others of equivalent scale in terms of parameters, pre-training data and capabilities. The particular characteristics of these models that prompt concern are not just that they can be easily adapted for a whole range of tasks, but also their unpredictable capacity to develop emergent capabilities, either of which may (in the wrong hands) pose significant risks to society.

The Summit will focus primarily on two types of risks posed by such AI:

  • First, it will examine misuse risks, where bad actors use Frontier AI capabilities as a tool to facilitate plans to cause significant harm or loss of life, for example, through assistance in developing biological weapons, or the enabling of sophisticated cyber-attacks.
  • Second, it will examine the risks of what is referred to as “loss of control”, associated with the hypothetical emergence of Artificial General Intelligence (AGI), where AI systems develop a form of self-awareness and/or hyper-intelligence, and in the most extreme “doomer” scenarios, promptly exterminate the entire human race.

The origins of Frontier AI

Much of the Summit’s focus draws on a joint research paper authored by researchers from leading AI developers and policy organisations including OpenAI, Deepmind and Microsoft, published in July this year. “Frontier AI Regulation: Managing emerging risks and public safety” defines Frontier AI, as “highly capable foundation models that could possess dangerous capabilities sufficient to pose severe risks to public safety”.

The paper sets out several examples of such dangerous capabilities, including enabling non-experts to design new biochemical weapons, mount mass disinformation campaigns with minimal effort and harness offensive cyber capabilities to cause unprecedented harm. The definition of Frontier AI, however, is specific to general purpose models and the risks they pose and does not incorporate narrow models (such as for example, the risks inherent in a model specifically designed to optimize the toxicity of a compound).

The paper notes that the crucial difference between narrow models and general purpose models is that the latter are often made available through “broad deployment” via sector-agnostic platforms such as APIs, chatbots or open sourcing, and as such “can be integrated into a large number of diverse downstream applications possibly including safety critical sectors.” The paper identifies risks specific to Frontier AI, such as the difficulty in preventing deployed models from being misused and proliferating broadly, and suggests regulatory building blocks to address these risks.

The Summit’s focus on Frontier AI risks has been criticized by voices in the machine learning community who have long argued that the focus on somewhat hypothetical catastrophic and even existential risks is a distraction from the more immediate and proximate concerns posed by AI technologies in use today, including bias, intellectual property infringement and the production of misinformation and deceptively realistic content.

The UK government has defended its approach in its own Summit Guidance, noting that as an international forum the Summit aims to focus on the most pressing concerns internationally while leaving the more immediate risks to national regulation and internationally processes currently underway. This includes the implementation of the EU AI Act, which we have previously discussed here and as further discussed below.

Irrespective of the criticisms levied at the focus of the Summit, it is clear that it will offer the UK the opportunity to place itself at the centre of the conversation on AI safety and build upon the already considerable international effort to regulate AI and create international standards.

Other international efforts relating to AI risks

The UK Summit is not the only attempt to focus the minds of the international community on the risks posed by AI. Several international organisations are already driving further collaboration, including the United Nations (UN), Organization for Economic Co-operation and Development (OECD), Global Partnership on Artificial Intelligence (GPAI), Council of Europe, G7 and G20.

Echoing the UK position, these organisations focus on international risks threatening to humanity as a whole, while mostly leaving it to individual countries to develop national strategies on the more nuanced risks. However, how far each organisation is willing to go and the options for future regulation considered by each organisation largely depends on their membership and whether it will be possible to align the members’ priorities and values.

United Nations

On one end of the spectrum is the UN, which convened the first ever UN Security Council debate on AI in July (also initiated by the UK). Reaching any kind of consensus between the 193 members of the UN with widely diverging values and priorities will be a challenge. Any efforts to regulate AI will therefore likely focus on risks to human rights, peace and security. With these aims in mind, the UN Secretary General António Guterres has proposed the following:

  • By the end of 2023: a High-Level Advisory Board for AI will report on the options for global AI governance.
  • By the end of 2026: conclude negotiations on a legally binding instrument to prohibit autonomous weapons systems that function without human control or oversight and establish a UN body to govern AI.

These proposals are inspired by existing UN treaties and oversight bodies brought into force to address international risks that are perceived to threaten humanity. The comparison shows that the risks of AI should not be underestimated and are being treated with the same urgency as:

  • nuclear weapons (see the Treaty on the Prohibition of Nuclear Weapons / the International Atomic Energy Agency);
  • international civil aviation (see the Chicago Convention on International Civil Aviation / International Civil Aviation Organization); and
  • climate change (see the Paris Climate Accords / Intergovernmental Panel on Climate Change).

G7 – Hiroshima Process

An example of an intergovernmental forum that is more closely aligned in the values of its members is the G7.

In a ministerial forum dubbed the “Hiroshima Process on Generative AI”, the leaders of the G7 countries are seeking to address a slightly broader set of risks than the UN or the UK Summit.

A G7 report published in September highlights the priority risks determined by the G7 countries as including intellectual property infringement, disinformation and manipulation of opinions, privacy, security and human rights.

This reflects the G7 members’ common values, such as freedom, democracy and human rights. According to the report, any output will also strive for fairness, accountability, transparency.

On 30 October 2023, the G7 published a set of International Guiding Principles on AI, and a voluntary Code of Conduct for AI developers. Further work is underway on a Comprehensive Policy Framework. That these documents were published on the eve of the UK’s summit is unlikely to be a coincidence.

Concluding remarks

The UK government has (arguably) adroitly pivoted its international AI policy efforts to focus on Frontier AI risks. These do not include more proximate topics such as bias, discrimination, privacy and intellectual property infringement. These will be left to local national or regional regulation such as the EU’s AI Act. This suggests that while consistency may be eventually found internationally in relation to Frontier AI risk, there is a very real potential for considerable regional divergence when it comes to regulation and policy-making in relation to the more proximate risks. Whether the EU’s AI Act, currently in an intense final stage of internal negotiation between the EU institutions in Brussels, becomes a benchmark in this area, remains to be seen.

Perhaps one thing we can all can take away from these international efforts, whether you are trying to come to grips with AI on a global, national or even company level, it is the UN Secretary General’s call “to approach this technology with a sense of urgency, a global lens, and a learner’s mindset.”

Charlie Hawes


Cosima Lumley

Elisa Lindemann


Related Articles