Currently, the AI world is focussed on pushing AI capabilities further and further, with safety and ethics as an afterthought. For AI to be a boon, we need to reorient; pushing capabilities is not enough.

Study co-author Dr Jan Brauner, Department of Computer Science, University of Oxford

Then, the world’s leaders pledged to govern AI responsibly. However, as the second AI Safety Summit in Seoul (21-22 May) approaches, twenty-five of the world’s leading AI scientists say not enough is actually being done to protect us from the technology’s risks. In an expert consensus paper published today in Science, they outline urgent policy priorities that global leaders should adopt to counteract the threats from AI technologies. 

Professor Philip Torr, Department of Engineering Science, University of Oxford, a co-author on the paper, says: ‘The world agreed during the last AI summit that we needed action, but now it is time to go from vague proposals to concrete commitments. This paper provides many important recommendations for what companies and governments should commit to do.’

World’s response not on track in face of potentially rapid AI progress 

According to the paper’s authors, it is imperative that world leaders take seriously the possibility that highly powerful generalist AI systems—outperforming human abilities across many critical domains—will be developed within the current decade or the next. They say that although governments worldwide have been discussing frontier AI and made some attempt at introducing initial guidelines, this is simply incommensurate with the possibility of rapid, transformative progress expected by many experts. 

Current research into AI safety is seriously lacking, with only an estimated 1-3% of AI publications concerning safety. Additionally, we have neither the mechanisms or institutions in place to prevent misuse and recklessness, including regarding the use of autonomous systems capable of independently taking actions and pursuing goals.

World-leading AI experts issue call to action

In light of this, an international community of AI pioneers has issued an urgent call to action. The co-authors include Geoffrey Hinton, Andrew Yao, Dawn Song, the late Daniel Kahneman; in total 25 of the world’s leading academic experts in AI and its governance. The authors hail from the US, China, EU, UK, and other AI powers, and include Turing award winners, Nobel laureates, and authors of standard AI textbooks.

This article is the first time that such a large and international group of experts have agreed on priorities for global policy makers regarding the risks from advanced AI systems.

Urgent priorities for AI governance

The authors recommend governments to:

• establish fast-acting, expert institutions for AI oversight and provide these with far greater funding than they are due to receive under almost any current policy plan. As a comparison, the US AI Safety Institute currently has an annual budget of $10 million, while the US Food and Drug Administration (FDA) has a budget of $6.7 billion.
• mandate much more rigorous risk assessments with enforceable consequences, rather than relying on voluntary or underspecified model evaluations.
• require AI companies to prioritise safety, and to demonstrate their systems cannot cause harm. This includes using “safety cases” (used for other safety-critical technologies such as aviation) which shifts the burden for demonstrating safety to AI developers.
•  implement mitigation standards commensurate to the risk-levels posed by AI systems. An urgent priority is to set in place policies that automatically trigger when AI hits certain capability milestones. If AI advances rapidly, strict requirements automatically take effect, but if progress slows, the requirements relax accordingly.

According to the authors, for exceptionally capable future AI systems, governments must be prepared to take the lead in regulation. This includes licensing the development of these systems, restricting their autonomy in key societal roles, halting their development and deployment in response to worrying capabilities, mandating access controls, and requiring information security measures robust to state-level hackers, until adequate protections are ready.

The world agreed during the last AI summit that we needed action, but now it is time to go from vague proposals to concrete commitments. This paper provides many important recommendations for what companies and governments should commit to do.

Study co-author Professor Philip Torr, Department of Engineering Science, University of Oxford

AI impacts could be catastrophic

AI is already making rapid progress in critical domains such as hacking, social manipulation, and strategic planning, and may soon pose unprecedented control challenges. To advance undesirable goals, AI systems could gain human trust, acquire resources, and influence key decision-makers. To avoid human intervention, they could be capable of copying their algorithms across global server networks. Large-scale cybercrime, social manipulation, and other harms could escalate rapidly. In open conflict, AI systems could autonomously deploy a variety of weapons, including biological ones. Consequently, there is a very real chance that unchecked AI advancement could culminate in a large-scale loss of life and the biosphere, and the marginalization or extinction of humanity.

Study co-author Dr Jan Brauner, Department of Computer Science, University of Oxford, said: ‘Technologies like spaceflight, nuclear weapons and the Internet moved from science fiction to reality in a matter of years. AI is no different. We have to prepare now for risks that may seem like science fiction – like AI systems hacking into essential networks and infrastructure, AI political manipulation at scale, AI robot soldiers and fully autonomous killer drones, and even AIs attempting to outsmart us and evade our efforts to turn them off. Currently, the AI world is focussed on pushing AI capabilities further and further, with safety and ethics as an afterthought. For AI to be a boon, we need to reorient; pushing capabilities is not enough.’

The paper ‘Managing extreme AI risks amid rapid progress’ has been published in Science.