On April 13, the Stanford Institute for Human-Centered AI published the 2026 AI Index and, inside 423 pages, included one chart that explains more of the week's politics than any release note: 73 percent of U.S. AI experts expect a positive impact on how people do their jobs. Twenty-three percent of the public agrees. [1] The gap is fifty points — wider than the index's trade-policy split, wider than its U.S.–China model-performance gap, wider than almost any dyad the report tracks.

Three names close a triangle around that number. Andrej Karpathy named the psychosis risk in knowledge work the paper wrote up on April 17. [2] Stanford has now quantified the legitimacy deficit the risk produces. And Janet Mills, on the same Wednesday Stanford published, signed the first state-level data center moratorium — a statute written, in effect, for the 77 percent of the public the experts are not speaking to. [3] Karpathy describes the mechanism; Stanford counts the house; Mills writes the law.

The index has done this before, with self-driving cars in 2019 and with generative models in 2023. [4] Each prior gap preceded, by roughly a year, a regulatory arrival. The difference this time is the speed and the size: a fifty-point chasm on the week after the Maine law, on the eve of Ohio's constitutional referral, with the Virginia local moratoria now drafting on county clerks' desks. Experts speak the industry's language. The public writes the bills. Stanford has told the experts what they already suspected — they are losing the vocabulary fight — and given the public's legislators a footnote to cite.

-- ANNA WEBER, Berlin