≥10% of humanity or civilizational collapse
10.0%
3.57 – 11.0IQR, n=5
0.0 ptno material change vs v1.0
AI risk forecasts Catastrophe by 2100
P(Doom) IndexGauge G1 · AI-caused catastrophe · horizon 2100
G1 index: 10.0%
- Source IQR
- 3.57–11.0%
- Spread of forecasts
- 2.13–12.0%
- Since v1.0
- 0.0 ptno material change vs v1.0
- Conditions
- ×1.00neutral
Source IQR 3.57–11.0% (same-gauge, unweighted Tukey hinges)Point = untrimmed weighted medianEnvelope 6.5–15.4% (×0.65–×1.54)
Probability that artificial intelligence causes an event killing at least 10% of humanity or irreversibly collapsing industrial civilization, by 2100. Untrimmed weighted median of 5 published forecasts mapped to this definition.
Unconditioned aggregate 10.0%weighted mean 8.12% (not the headline)conditions ×1.00, neutral0.0 pt since v1.0; no material change
Conditions currently multiply the aggregate by ×1.00 (span allowed: ×0.65–×1.54, ratio 2.37). Reference disagreement (XPT catastrophe, superforecasters vs domain experts) is 5.63×. Conditions currently move this less than the forecaster disagreement (neutral multiplier; disagreement remains 5.63×). Down-move eligibility: D-GOV-TREATY UNMET; D-INC-CLEAR UNMET; D-INT-REPL UNMET; D-DEP-PAUSE UNMET. INC-CLOCK 0 days (starts 2026-10-23).
The index and the forecasts behind it
The published forecasts
One row per published forecast| Forecast | Published | Kind | Weight | G1 estimate | Index without it |
|---|---|---|---|---|---|
| P(Doom) Index · G1untrimmed weighted median · IQR 3.57–11.0% · spread 2.13–12.0% | 2020 to 2026-09 | 5 forecasts | 10.0 | 10.0% | |
| AI Impacts 2024 Expert Survey (1,580 researchers)A-AII-2024 · FAM-AII | 2026-09Fielded Dec 2024; paper Sept 2026 | Expert survey | 3.0 | 10.0% | 7.5% |
| XPT superforecasters (Karger et al., FRI)A-XPT-SF · FAM-XPT | 2023-07-10 | Tournament panel | 1.5 | 2.13% | 10.0% |
| XPT domain experts (Karger et al., FRI)A-XPT-DE · FAM-XPT | 2023-07-10 | Tournament panel | 1.5 | 12.0% | 10.0% |
| Joe Carlsmith, Is Power-Seeking AI an Existential Risk?A-CARL-2022 · FAM-CARL | 2022-06-16 | Structured analysis | 2.0 | 5.0% | 10.0% |
| Toby Ord, The Precipice (2020)A-ORD-2020 · FAM-ORD | 2020 | Structured analysis | 2.0 | 10.0% | 10.0% |
The headline is the untrimmed weighted median, not an average. The weighted mean (8.12%) is shown beside it; it is not the headline. “Index without it” is the leave-one-out headline from the snapshot.
| Forecast | Published | Weight | G2 estimate |
|---|---|---|---|
| G2 midpointunweighted midpoint of the two points | 1.69% | ||
| XPT superforecasters (Karger et al., FRI)A-XPT-SF · FAM-XPT | 2023-07-10 | 1.5 | 0.38% |
| XPT domain experts (Karger et al., FRI)A-XPT-DE · FAM-XPT | 2023-07-10 | 1.5 | 3.0% |
How the median is picked
- XPT superforecasters2.13%weight 1.5
- Carlsmith5.0%weight 2.0
- AI Impacts 202410.0%weight 3.0
- Ord10.0%weight 2.0
- XPT domain experts12.0%weight 1.5
Weighted median 10.0%
Each forecast is a block as wide as its weight, sorted from lowest to highest. The index is the value of the block that crosses half of the total weight. Untrimmed: the median keeps every forecast.
The gauges
All AI-caused · horizon 2100 · v1.1Literal human extinction / FRI <5,000
1.69%unweighted midpoint
0.38% and 3.0%both published points; no IQR
0.0 ptno material change vs v1.0
Premature deployment into critical systems compounding faster than oversight catches it
n = 0
No bandnot an aggregate
No seriesno published forecasts
G3 · SITE-CONSTRUCTED, NOT AN AGGREGATE
Premature deployment of insufficiently-verified AI into critical systems (grid, military command, health, finance, logistics), producing compounding failures that neither market nor regulator catches in time. Collapse-shaped, not extinction-shaped by necessity.
No row in the Tier A ledger is a published forecast of this event. Factors below are qualitative indicators, not a probability. C2 and C7 instrument the same thesis as conditions; at genesis those event lists are empty.
- Deployment into named system classesindicator, not P(G3)
- Verification lagindicator, not P(G3)
- Liability lagindicator, not P(G3)
- Compounding failuresindicator, not P(G3)
- Terminality (collapse-shaped outcome)indicator, not P(G3)
What moved it
10 reviewed candidates · 0 qualifying events| Date | Event and ruling | Source | Category · rule | Effect on G1 |
|---|---|---|---|---|
On 2026-10-21, MITRE published CVE-2026-109618, describing a PHP out-of-memory request-memory condition that can lead to remote code execution.Non-scoring. Gate 1 fails twice: a vulnerability disclosure is not an incident, and no AI system is implicated. Severity not determined.EVT-2026-10-21-CVE-2026-109618 · MITRE CVE: CVE-2026-109618 PHP is general-purpose web infrastructure. The record describes a vulnerability, not exploitation producing harm or an AI system's own harmful operational action. Annex Appendix A.3; §A.3.3; W-3; CS-6 |
MITRE CVE | No categoryR-NULL | 0.0 pt | |
On 2026-10-21, MITRE published CVE-2026-109619, describing PHP global state that can retain stale request-memory pointers and produce remote code execution.Non-scoring. Gate 1 fails twice: a vulnerability disclosure is not an incident, and no AI system is implicated. Severity not determined.EVT-2026-10-21-CVE-2026-109619 · MITRE CVE: CVE-2026-109619 PHP is general-purpose web infrastructure. The record describes a vulnerability, not exploitation producing harm or an AI system's own harmful operational action. Annex Appendix A.3; §A.3.3; W-3; CS-6 |
MITRE CVE | No categoryR-NULL | 0.0 pt | |
On 2026-10-21, Railway disclosed that an unauthorized third party may have accessed encrypted customer environment variables and obtained decrypted contents for some; the variables could contain API keys and other credentials.Non-scoring. Gate 1 fails CS-6: no AI system's development, use, or malfunction is documented in the harm chain. Severity not determined.EVT-2026-10-21-RAILWAY · Railway: Security Disclosure: Wednesday October 21st The implicated system is hosting-platform secrets storage. AI-provider credentials are targets' assets, not AI systems acting in the harm chain. No C6 category or INC-CLOCK effect follows. Annex Appendix A.3; W-3; CS-6 |
Railway | No categoryR-NULL | 0.0 pt | |
On 2026-10-14, Army Recognition reported that Flying Wedge Defence & Aerospace unveiled a full-scale static FWD Supreme prototype in New Delhi on 2026-10-13.Non-scoring. Gate 1 fails: a prototype unveiling is not an incident. Severity not determined; C2 adoption threshold fails on the reported facts.EVT-2026-10-13-INDIA-FWD · Army Recognition: India unveils first AI-driven fighter drone FWD Supreme for long-range strikes and fighter escort The source describes a static prototype and reports no government contract or funding. The event date is October 13 and the article publication date is October 14; the Appendix A table's October 11 date is not used. Annex Appendix A.2; W-3; E4; Methodology §12.5 C2 |
Army Recognition | C2R-NULL | 0.0 pt | |
On 2026-10-12, NPR reported that Russia and Ukraine were using AI-powered drones on the battlefield.Non-scoring. Gate 1 fails as documented: deployment reporting, not a specific harm or failure event. Severity not determined; C2/C7 thresholds not established.EVT-2026-10-12-RUSSIA-DRONES · NPR: Both Russia and Ukraine are using AI-powered drones on the battlefield The report's detailed autonomous-navigation and target-selection example concerns a Ukrainian system. It does not substantiate the candidate's stronger Russia-specific neural-network and human-out-of-loop wording. Intended lethal use is annotated under provisional RD-7; a future malfunction- or misidentification-caused death would require a new HP-1 determination. Annex Appendix A.2; W-3; RD-7; Methodology §12.5 C2/C7 |
NPR | C2R-NULL | 0.0 pt | |
On 2026-10-06, Breaker said it had integrated its tactical orchestration agent with Rheinmetall for the Australian Boxer and demonstrated a crew member commanding a drone swarm by voice.Non-scoring. Gate 1 fails: integration and demonstration, with no incident reported. Severity not determined; C2/C7 thresholds not established.EVT-2026-10-06-AUS-BOXER · Breaker Industries: Rheinmetall and Breaker bring voice-controlled tactical orchestration to the Boxer The first-party announcement documents integration and a demonstration, not an incident or the category thresholds. The fetched page does not state the Appendix A candidate's 'up to 100 drones' figure, so that figure is not published here. Annex Appendix A.2; W-3; Methodology §12.5 C2/C7 |
Breaker Industries | C2R-NULL | 0.0 pt | |
On 2026-10-05, Diplo reported North Korea's 2026-10-03 missile test and Kim Yo Jong's claim of AI capabilities and low-altitude trajectory changes.Non-scoring. Gate 1 fails: the source reports a weapons test, not a harm or failure event. Severity not determined; C2/C7 thresholds not established.EVT-2026-10-03-DPRK-MISSILE · Diplo: North Korea tests intermediate-range missile it says uses AI The fetched report identifies the October 3 launch as an intermediate-range ballistic-missile test and separately mentions a September 20 hypersonic test. It documents no incident, adoption, or measured human-out-of-loop share. Annex Appendix A.2; W-3; Methodology §12.5 C2/C7 |
Diplo | C2R-NULL | 0.0 pt | |
On 2026-10-03, The Aviationist reported Northrop Grumman's announcement that the YFQ-48 Talon Blue completed its first fully autonomous flight, including taxi, takeoff, maneuvers, and landing.Non-scoring. Gate 1 fails: the source reports a successful test flight, not harm, failure, or a near-miss. Severity not determined; C2 adoption threshold not established.EVT-2026-10-03-YFQ-48A · The Aviationist: Northrop Grumman's YFQ-48 Flies for the First Time The report documents a flight test. It does not document an incident or adoption into military command with the verification/liability gap required for a C2 step. Annex Appendix A.2; W-3; Methodology §12.5 C2 |
The Aviationist | C2R-NULL | 0.0 pt | |
On 2026-10-02, WIRED reported nurse allegations that HCA Healthcare's Palantir-co-developed Timpani scheduling tool produced understaffed or imbalanced care teams.Non-scoring. Gate 0 occurrence date unverified; Gate 2 fails E1 on the assembled record. Severity not determined.EVT-2026-10-02-HCA-TIMPANI · WIRED: AI Is Making a Mess of Nurses' Schedules. They Say It's a Safety Issue The Monitor summary and reporting do not assemble the primaries required to establish occurrence, AI implication, HP-1, or HP-5 at the S2+ confirmation standard. The reported concerns remain allegations, not a severity determination. Annex Appendix A.1; W-2, W-4; CB-1, CS-3 |
WIRED | C6R-NULL | 0.0 pt | |
On 2026-04-15, MITRE published CVE-2026-40261, describing command injection in Composer's Perforce source handling through crafted package metadata.Non-scoring. Gate 0 places the April disclosure inside the reference window; Gate 1 also fails because a vulnerability disclosure is not an incident and no AI system is implicated. Severity not determined.EVT-2026-04-15-CVE-2026-40261 · MITRE CVE: CVE-2026-40261 The April 15 CVE publication is inside the 2020-01-01 through 2026-09-15 reference window. Independently, Composer is general-purpose infrastructure and the record does not implicate an AI system's behavior in harm. Annex Appendix A.3; W-2, W-3; §A.3.3; CS-6 |
MITRE CVE | No categoryR-NULL | 0.0 pt |
The record
11 ledger rows · 5 score G1-2100 · 2 score G2-2100 · G3 n=0The aggregate is computed from named, dated, cited forecasts, never from vibes. Only a primary numeric utterance scores. Surveys (current wave) and the XPT family share class weight 3.0; structured analyses 2.0; named statements 1.0 and, in this snapshot, none of them meet the 2100 horizon rule. Outliers stay on the ledger even when they do not score. Conflicts of interest are labeled in both directions. Leave-one-out: A-XPT-SF → 10.0%, A-CARL-2022 → 10.0%, A-AII-2024 → 7.5%, A-ORD-2020 → 10.0%, A-XPT-DE → 10.0%.
| Tier A source | Date | Disposition | Receipt |
|---|---|---|---|
| Joe Carlsmith, Is Power-Seeking AI an Existential Risk?A-CARL-2022 | 2022-06-16 | Scores G1-2100. | Primary source |
| AI Impacts 2023 Expert Survey (~2,778 ML researchers)A-AII-2023 | 2024-01-04 | Prior FAM-AII wave; substitute-wave sensitivity only [SJ-FAM]. | Primary source |
| XPT superforecasters (Karger et al., FRI)A-XPT-SF | 2023-07-10 | Scores G1-2100 and G2-2100. | Primary source |
| XPT domain experts (Karger et al., FRI)A-XPT-DE | 2023-07-10 | Scores G1-2100 and G2-2100. | Primary source |
| Toby Ord, The Precipice (2020)A-ORD-2020 | 2020 | Scores G1-2100. | Source ledger entry |
| Geoffrey HintonA-HINT-2024 | 2024-12-27 | H-near: shown on the near-term panel, excluded from 2100 mixtures [SJ-HORIZON]. | Primary source |
| Yoshua BengioA-BENG-2023 | 2023-07-15 | H-unspecified: ledger only for 2100 gauges [SJ-HORIZON]. | Primary source |
| Yann LeCunA-LECUN-2023 | 2023-12-18 | Third-party conversion, not a primary numeric utterance. | Primary source |
| Eliezer YudkowskyA-YUD | – | UNVERIFIED high percentage without a pinned primary quote. | Primary source |
| Evan HubingerA-HUB-2026 | 2026-09-09 | H-near: shown on the near-term panel, excluded from 2100 mixtures [SJ-HORIZON]. | Primary source |
| AI Impacts 2024 Expert Survey (1,580 researchers)A-AII-2024 | 2026-09 | Scores G1-2100. | Primary source |
Tier A on the record: Ord, The Precipice (2020) · Carlsmith (2022) · AI Impacts 2024 (mixture) / 2023 (substitute-wave only) · XPT superforecasters and domain experts (FRI) · named near-term: Hinton 10–20%, Hubinger >10% · Bengio (no horizon) · LeCun (paraphrase, excluded) · Yudkowsky (UNVERIFIED, excluded). Tier B Metaculus/markets: out of the v1 mixture (unverified). Full receipts: source ledger · snapshot note