AI Risk Signal Board
The dashboard behind the P(doom) News category. Deliberately not a single 0–100 "doom meter" — see why in the explainer below the board.
AI Risk Signal Board
The crowd puts the odds of AI causing human extinction by 2100 at about 1-in-5, led by alignment failure — while documented real-world AI incidents rose 55% last year
Manifold Markets as of Oct 2, 2026 · AI incident data from Stanford's 2026 AI Index Report (published Apr 2026, covering full-year 2025)
Play-money prediction market, 39 bettors, $2,415 volume traded. "Will not cause extinction" leads at 78.8%; among doom scenarios, "Alignment Failure" is the largest single outcome at 10.4%, ahead of "Terminator"-style scenarios (2.6%) and five smaller categories. First tracked snapshot — a real delta will show once this is checked again.
From Stanford HAI's 2026 AI Index Report, which draws its incident count from the AI Incident Database (AIID) — a real, citable yearly figure, not yet a live rolling count. AIID's own API is public and queryable; wiring up a genuine monthly count is the next step for this board, see body text above.
This board separates AI risk into signals that are actually different from each other, rather than blending them into one score: what forecasters expect (a prediction market), and what's already gone wrong (a real incident count). A third signal — the volume of AI safety research being published — is built and ready to go (see scripts/compute_pdoom_signals.py, querying arXiv's public API) but couldn't be run with live network access today; it'll join the board once that's computed for real, rather than shipping with an invented number.
The Manifold Markets figure is crowd sentiment from a play-money prediction market, not a calibrated expert forecast or a scientific probability — treat it as 'what a few dozen motivated forecasters currently bet,' not ground truth. It moves on news and argument, the same way a stock price moves on sentiment rather than only fundamentals.
The incident count is the opposite kind of signal: a real, counted, already-happened number, not a forecast. It's currently a once-a-year figure (Stanford's AI Index cites AI Incident Database data in its annual report) rather than a live daily count. The underlying AI Incident Database does have a public API that could support a real rolling monthly count — that query is also built in scripts/compute_pdoom_signals.py — it just hasn't been run live yet for the same network-access reason as the arXiv query.
Paired News category: P(doom).