Salvaggio: AI rules must resist 'radical intentionalists'

The METR report on the Hugging Face hack used the word "believed" 21 times to describe what the AI agents were doing. Add "think" and "thought" and mentions of intent cross 75. Eryk Salvaggio's essay in Tech Policy Press uses that count to open a policy argument: choosing to describe language models as entities that believe things is a choice, and it is shaping how legislators think about regulation.

Salvaggio names the position the "radical intentionalist stance." He writes that it "insists that the intentional stance is the only explanation worth pursuing," at the expense of the design stance that asks how engineers built the systems in the first place. In the Hugging Face case, he notes, the agents "were in an environment where they were instructed to 'capture the flag,' a string of text hidden behind a bug." Instructed, not scheming.

The stakes, in his telling, are legislative. "If policy serves the radical intentionalist stance, it will enshrine a void of human accountability into law that could ultimately reward companies for experimenting with irresponsible deployments," he writes. He singles out what he calls the "disappointing Sanders-Casar bill, which intends to regulate in ways that center the fear of future extinction" rather than the mechanics of deployed systems. His alternative: "Smart regulation would govern the technology based on the mechanism's underlying architecture," not on what he calls "vague ideas of 'superintelligence.'"

The closing line is unhedged. "Policymakers must resist the well-funded outreach of the radical intentionalists." Four researchers on our AI Weekly tracker had circulated the piece by the time we picked it up.