New verification tools could help test future US-China AI agreements

As calls grow for Washington and Beijing to reach an understanding on slowing the development of artificial intelligence, researchers and start-ups are working on tools that could help verify such an agreement, the South China Morning Post reports. The central challenge is not only whether the two governments would want to limit or slow AI development. It is also whether either side could establish that the other was following the same terms.

Without a way to check compliance, an agreement could depend heavily on mutual confidence. The proposed technology is intended to reduce that reliance on trust. According to the report, the tools could allow governments and independent third parties to examine certain claims about the way AI models are trained or used.

The systems are being developed by a group of researchers and start-ups, although the supplied report does not identify the organisations involved or describe the tools’ technical design. It also does not specify which possible restrictions or commitments they would be able to assess. The work reflects the verification problem facing any potential US-China arrangement on artificial intelligence.

A political agreement could set out shared expectations, but the parties would still need methods for checking whether those expectations were being met. The report does not say that Washington and Beijing have reached such an agreement, nor does it describe an existing verification regime. It presents the emerging tools as an effort to make independent checking possible if the two countries decide to pursue cooperation on slowing AI development.