Methodology

The formula

AGI Proximity (30%)

Market probabilities that AGI arrives soon. “Near” resolves by 2028 (Manifold “AGI before 2028”, Kalshi “any company announces AGI before 2028”, Polymarket “OpenAI announces AGI before 2027”). “Far” means by 2030 (Manifold, Kalshi). Kalshi uses the bid/ask midpoint. The two buckets are averaged so no single venue dominates.

Expert p(doom) (25%)

Our curated list of self-stated p(doom) estimates, using each person's latest statement and the midpoint of ranges. The median is doubled and capped at 100, so a coin-flip median reads as maximum alarm. Every entry is checked against its primary source and carries the supporting quote; statements that aren't probabilities (such as a 1–10 threat rating) are shown but excluded. experts.json

Crowd Catastrophe Odds (15%)

Manifold markets on a massive AI-caused catastrophe before 2030 and AI killing ≥10% of humans before 2100. These are play-money markets, so they get a lower weight.

Rogue AI Activity (15%)

From the Alignment Incident Database. Each incident is weighted by reach (speculative 0.5, internal 1, breakout 2, deployed 2) × harm (none 1, minor 1.5, moderate 2, severe 3). The last 90 days are compared with the average 90-day window of the past year: a normal quarter scores 50 and double the usual pace scores 100.

Safety Deficit (15%)

The Future of Life Institute's AI Safety Index grades frontier labs on a 0–4 GPA scale. We take the mean across all graded labs and invert it.

History & backfill

A snapshot is stored every day in history.json. Before live tracking began, the first 90 days were reconstructed by re-running the same formula on each market's public price history, plus the statements and incidents dated on or before each day and today's lab grades. Reconstructed days are flagged and drawn dashed on the chart.

Failures

Every source is fetched independently. If one fails, its component is computed from the remaining inputs. If a whole component fails, yesterday's value is carried forward and flagged, or the weights are re-normalised.

Sources

Manifold, Kalshi and Polymarket public APIs; the Alignment Incident Database (CC BY 4.0); the AI Incident Database RSS (context only); the FLI AI Safety Index, Summer 2026; and individually linked expert statements. Metaculus is excluded because its API now requires an account token.

Limitations

Market definitions of “AGI” differ, and Manifold is play money. p(doom) statements use different definitions and horizons. Incident counts reflect what gets reported and catalogued. The weights are editorial judgement, published so you can disagree with them.

Global Market Fear Index (separate)

A second, independent index on the Global Market Fear page. It does not feed into the Doom Index and has its own method version (v1.0).

Percentiles make very different gauges comparable: Brazil's ETF VIX normally sits near 30 while Australia's A-VIX sits near 12. The weights are editorial and roughly track each region's share of world equity markets, capped so the US (VIX + VXN) is 30% rather than its ~60% market-cap share. Regions without a reliable keyless feed are listed on the page and get no weight; the weights of any index missing on a given day are re-normalised across the rest.

Data: the Yahoo Finance chart API (VIX, VXN, India VIX, S&P/ASX 200 VIX) and Cboe's public daily index history CSVs (VXEFA, VXFXI, VXEEM, VXEWZ), each with the other as fallback where available. The last good series is cached. If a source fails, its last close is kept and flagged; a close older than 4 calendar days is marked stale, and after 10 days the index drops out of the blend. History: one snapshot per day in volatility_history.json; the year before live tracking began was rebuilt by applying the same formula to the historical closes available on each date. Rebuilt days are flagged and drawn dashed.