The Doom Predictions, Graded
A decade of AI executives warning the world about their own product, checked one by one against what actually happened. Most of it is unfalsifiable marketing. Three predictions were specific enough to test — and the newest grade just came in, from Anthropic's own numbers.
What was checked, and what could not be.
Anchored to a named, dated primary source for every quote in the piece: Amodei's Senate testimony (Washington Post, 7/25/2023), Musk's 2017 remark (Fortune) and 2024 restatements (Deadline), Altman's Senate remarks (Forbes, 5/16/2023), the CAIS statement (safe.ai, 5/30/2023), OpenAI's GPT-2 release note, Amodei's 2025 jobs prediction (Axios), and Amodei's own essay "Machines of Loving Grace."
The pivotal claim, that the bioweapons warning has "moved from prediction to something closer to confirmed," rests entirely on Anthropic's own September 2026 threat-intelligence report. That report is Anthropic grading its own product on the exact question its CEO raised in 2023. No independent lab, academic study or government body is cited confirming the "five cases" figure or the "no longer a certainty" language; it comes straight from Anthropic's self-published safety document. That is disclosed in the text but it means the central grade in this piece is a company checking its own homework, not a third party checking it.
Musk's 10-20% figure is explicitly attributed to Geoffrey Hinton, not to independent Musk analysis, and the piece is careful to say so. The claim that Musk repeated it at three named events (Abundance Summit, Cannes Lions, FII) is sourced to Deadline for the Cannes appearance; the other two are asserted without a distinct citation for each individual restatement, only Deadline's account referencing the pattern.
The Axios-sourced jobs prediction is graded "pending, trending toward confirmed" on the strength of "early labor-market data" that is not linked or named anywhere in the piece, no BLS series, no named study. That is an estimate dressed as an observation and should be read as an editorial judgment call, not a sourced figure.
The excluded item, a screenshot-based claim from an earlier draft, was correctly left out for lack of a primary source. That is good practice, not a gap.
One structural note: OpenAI's GPT-2 "didn't happen" grade also rests on OpenAI's own November 2019 release note, the company self-reporting the absence of harm. Same pattern as the Anthropic case above, just with a grade that runs against the company's original claim rather than for it.