Evidence library

Aggregated crowd forecasts for weakly general and general AI. Community median for strong AGI is around the early 2030s.

Employment of 22-25-year-olds in AI-exposed occupations trails less-exposed peers by 19 points (Jun 2026), driven by reduced hiring rather than separations.

Play-money prediction market resolving to the year an AI passes a high-quality adversarial Turing test. Drives the live median-year signal.

EEG study over 4 months: LLM-assisted writers showed up to 55% reduced brain connectivity and 83% could not quote their own just-written essay. Preprint; small n in the final session.

Contrary evidence, catalogued for balance: meta-analysis finding positive effects of ChatGPT tutoring on measured learning performance. The cognition pillar weighs both directions.

319 knowledge workers, 936 real tasks: higher confidence in AI predicts lower perceived critical-thinking effort (beta = -0.69, p < 0.001); 55-79% report reduced effort per task type.

Projects global AI water withdrawal of 4.2-6.6 billion m3 by 2027; ~80% of datacenter water is lost to evaporation.

Employer survey covering 14M workers: 92M jobs displaced and 170M created by 2030 (net +78M) — structural churn of 22% of today's jobs. Balance evidence for the labor pillar.

Global datacenter electricity: 415 TWh in 2024 (1.5% of world demand) to ~945 TWh by 2030 in the base case — growth 4x faster than every other sector.

Largest survey of AI researchers to date. Median respondent assigns ~5% probability to extinction-level outcomes from advanced AI within 100 years; median HLMI arrival ~2047.

Quasi-experiment: Stack Overflow activity fell 25% within six months of ChatGPT relative to control platforms, across all experience levels. Anchors the live offloading signal.

~40% of global employment exposed to AI (60% in advanced economies), with about half of exposed jobs at risk of negative effects. Sets the labor-pillar anchors.

First comprehensive binding AI regulation, including obligations for general-purpose and systemic-risk models. A mitigation signal.

Open dataset tracking training compute, parameters and data of notable ML models. Basis for the compute-growth signal.

Crowdsourced repository of real-world AI incidents and harms, used to track the frequency and severity of failures.

US datacenters used 176 TWh in 2023 (4.4% of US electricity), projected to 325-580 TWh (6.7-12%) by 2028.

Curated technical research on aligning advanced AI systems with human intent, the core open problem behind the risk signals.

One-sentence statement that mitigating AI extinction risk should be a global priority, signed by leading researchers and lab CEOs.

Open letter calling for a public, verifiable pause on training systems more powerful than GPT-4, signed by tens of thousands.

OECD-average mathematics fell 15 points vs 2018 (~0.75 school-year), triple any prior cycle change. COVID-confounded; PISA 2025 is the first genuinely post-LLM cycle.

169 superforecasters and domain experts. AI-caused catastrophe by 2100: 2.13% (supers) vs ~12% (experts); AI-caused extinction: 0.38% vs ~3%. The AI-specific medians calibrate the IRCA hazard model; catastrophe = death of >=10% of humanity within 5 years.

Six-premise multiplicative decomposition of AI existential risk (~5%, later revised >10% by 2070). Methodological basis for auditable conditional chains.

Assigns ~1-in-6 total existential risk this century (unaligned AI ~1/10). Used as the upper tail of the anchor uncertainty range.

Foundational analysis of existential risk from machine superintelligence and the control problem.

Coupling between networks converts graceful degradation into abrupt first-order collapse. Basis for the IRCA cascade-coupling stage.

The canonical reference for composite indices: compensability of aggregation rules, power means, and mandatory uncertainty analysis. Basis for the IRCA aggregation stage.