Sahai urges investment in people who can scrutinise AI discoveries

Amit Sahai is calling for more mathematically skilled researchers to study and assess ideas generated by artificial intelligence, arguing that human communities need the capacity to understand discoveries that could shape consequential technologies.

In a guest post on Terence Tao’s blog, Sahai says AI systems he has worked with are producing ideas that go beyond calculations or arguments familiar to strong human researchers. He warns that researchers may feel unable to keep pace, but says abandoning the work of understanding would be a collective abdication of responsibility.

Sahai proposes sustained support for research groups to spend extended periods examining unfamiliar AI-generated ideas, with help from AI tools. He also calls for a significant expansion of the pool of mathematically sophisticated researchers, able to contribute across fields.

His example is hypothetical: a future AI system proposes a one-terawatt fusion power plant based on principles people have not conceived of or tested. Before considering construction, Sahai says, communities would need to understand the design, the evidence behind it and the uncertainties involved, including how failures might be contained.

He says human involvement does not automatically improve technical decisions, and that people need not manually repeat work an AI can do more reliably. But he argues that understanding a guarantee requires examining the model it rests on, the evidence for that model and its limitations. Independent expertise, he adds, should extend beyond the organisation proposing a technology.

Sahai describes the wider pool of experts as a “deployable intellectual reserve” and says expanding human capacity to understand unfamiliar work could help preserve meaningful agency over decisions with major consequences. He says GPT-6 Astra helped draft the post and that its views are his own.