Frontier AI: Regulatory Challenges

The two-day international artificial intelligence (AI) safety summit that was held in the United Kingdom on November 1^st and 2^nd at Bletchley Park in Buckinghamshire—the pile that is well known for British mathematician Alan Turing's decoding of the Enigma code and as the birthplace of the first programmable digital computer, Colossus—can perhaps be termed as a first big step towards arriving at a cooperative approach to control the threats posed by technology.  

Around 150 representatives from across the globe including government leaders from the US, EU, China, India, Brazil, Indonesia, etc., and industry, academia and civic society leaders, participated in the summit, which in itself is a “remarkable achievement” for the UK in diplomatic terms.

Under a joint commitment signed by 28 governments including the US, China and the EU, and leading AI companies had reached a consensus on the need for sustained international cooperation to combat both the short-term and long-term risks posed by ‘frontier AI’. Accordingly, all the advanced AI models shall be subjected to a battery of safety tests before they are released. It also emphasized the need to share an evidence-based understanding of the risks posed by frontier AI and the safety measures thereagainst across countries.  

More importantly, participants agreed to develop a ‘State of the Science’ report on the capabilities and risks of frontier AI by an international panel of scientists assembled under the leadership of AI luminary, Yoshua Bengio. The report developed by such an eminent group of scientists shall be an invaluable document for educators, employers, policymakers and scientists. Interestingly, the United Nations confirmed its support for the creation of an expert AI panel akin to the Intergovernmental Panel on Climate Change.

Another interesting outcome is an announcement of the UK for the creation of AI safety Institute meant for researching the most advanced AI capabilities and testing their safety. It proposes to collaborate with its international counterparts and like-minded governments. The US also announced the formation of its own AI Safety Institute.

Said that, we must also appreciate that mere summit agreements will not be enough to achieve a balance between risk management and innovation. Indeed, Prof Robert Trager, Director of Oxford Martin AI Governance Initiative, observed that the summit failed to arrive at a “consensus path forward in establishing international standards and oversight of advanced AI.” In fact, challenges are aplenty: designing ‘tripwires’ that would subject certain models to heightened scrutiny and constraints, developing AI safety research that incorporates the complexities of human interaction with AI systems, understanding how frontier AI technology is likely to behave when it is incorporated into billions of automated problem-solving software ‘agents’, etc., are some of the pressing challenges that need to be addressed. There is thus, a need for pragmatic design of institutional mechanisms such as the international aviation safety process, to counter these global challenges.

No doubt, the summit facilitated a global conversation on AI safety, and the need for international collaboration on AI regulation, however, there existed divergent views on the type of regulation required. Indeed, different processes are already running in parallel: The US government has issued executive orders on safe, secure, and trustworthy AI. The European Union is finalizing introduction of its regulatory mechanism for AI. China has already announced its regulatory framework. These national regulations may be able to deal with simpler AI applications and LLMs but the most powerful frontier models—models feared to be capable of creating harmful pathogens or cyber weapons and might lead to “artificial general intelligence” which could even threaten humanity’s survival—call for global rules and an international body to regulate them. Thus, the basic issue remains unanswered: How to engage all the countries, including China, in arriving at an acceptable global regulatory framework and an institute that is fit enough to verify the models to be introduced and certify them as trustworthy for their continuation? Indeed, as some experts opined, it may even call for the creation of a range of institutions. 

In conclusion, it merits to bear in mind the sane observation of Dr Heloise Stevance, Schmidt AI in Science Fellow: “Historically, technological advances have not benefited all tranches of society and all countries equally— if we want a better and safer future for everyone, we must ensure that the fruits of the AI revolution are not only ‘safe to eat’ but also shared fairly with humanity.”