Stephen Casper called for stronger safeguards against severe AI misuse in a Harvard interview published on October 19. The computer scientist sees room to act before a major harmful event and advocates greater transparency around developers’ safety decisions and stronger government control over the pace of development.
The conversation concerns an October 15 scenario analysis with Max Kamachee and Jordi Weinstock at the Belfer Center. It examines hypothetical major cyber, biological or chemical attacks assisted by AI. Those scenarios do not describe attacks that have already happened.
Casper recommends layers of protection rather than a single safeguard: limiting capabilities, training against harmful requests and monitoring usage. External controls are much harder to enforce on downloadable models, he argues. His judgments about effectiveness are expert assessments, not guarantees.
He considers serious risks plausible in the near future while emphasizing uncertainty about precise timing. Open research and reliable comparisons are particularly limited for biological and chemical risks, he says. The analysis establishes neither a particular imminent attack nor its probability.