Affiliation: Founder, AI Impacts; researcher focused on AI forecasting methodology and expert surveys One-line position: The most useful thing we can do for AI forecasting is measure what researchers actually believe, track how those beliefs shift, and ground the discourse in data rather than vibes.
What she's reacting against
- Gut-feel timelines that aren't anchored to any systematic evidence
- Tribal polarization of AI discourse (doom vs. hype) — insists on empirical measurement
- Overconfidence in any single forecasting methodology — surveys are one input, not the answer
- The assumption that expert opinion is stable — her data shows it's shifting fast
Key claims
- AI researcher timelines are shortening dramatically: the 2023 survey found 50% probability of human-level AI by 2047 — 13 years earlier than the same cohort's estimate just one year prior
- 10% chance of human-level machine performance by 2027: from the 2023 survey of 2,778 AI researchers (published 2024)
- Expert surveys are the empirical backbone of timeline discourse: her survey series (2016, 2022, 2023) is the most-cited source for "what do AI researchers actually expect"
- Median estimates mask wide disagreement: the distribution of researcher forecasts is very wide, meaning apparent consensus hides deep uncertainty
- Methodology matters: how you frame the question ("human-level AI" vs. "full automation of labor" vs. "AGI") changes the answer significantly
Theories aligned with
- Empirical forecasting methodology
- Adjacent to AI alignment community but primarily a measurement voice, not an advocate
Where she overlaps / splits
- Overlaps with Cotra on trying to ground timelines in data; splits on method — Cotra uses compute anchors, Grace uses expert surveys
- Overlaps with Tetlock-style forecasting community on methodology; splits by being domain-specific to AI
- Splits with confident short-timelines voices (Aschenbrenner, Kokotajlo) — her data shows wide uncertainty, not convergence on "soon"
- Splits with dismissive voices (Marcus) — the data shows researchers are moving timelines forward, not standing pat
Notable predictions
- (2024-01, published) AI researchers' median for human-level AI: 2047 — outcome: TBD; the estimate itself is the data point
- (2024-01) 10% chance of human-level AI by 2027 — outcome: TBD; live test
- (2022) Previous survey had median ~2060; the 13-year shift in one year is itself a finding — outcome: documented; raises question of whether estimates are stable or reactive to recent demos
Track record
- Built AI Impacts into the most-cited empirical forecasting resource in the AI timeline space
- Survey series is referenced by virtually every serious timeline discussion — genuine field infrastructure
- Methodology has been critiqued (survey framing effects, selection bias) but remains the best available systematic data
- Does not make strong personal timeline predictions — positions herself as the measurer, not the forecaster
Empirical vs normative
- Almost entirely empirical: survey design, data collection, methodological analysis
- Normative content is minimal: implicit position is that better forecasting improves decision-making, but she rarely advocates specific policies
- This makes her unusual in the field — most AI timeline voices have strong normative commitments
Sources
- Survey papers: "Thousands of AI Authors on the Future of AI" (2024, with others); prior surveys 2016, 2022
- Website: aiimpacts.org
- Personal site: katjagrace.com
- Google Scholar: multiple publications on AI forecasting methodology
Weak spots / open questions
- Expert surveys are susceptible to anchoring effects, social desirability bias, and framing sensitivity — Grace acknowledges this but the data is still used as if it's harder than it is
- The 13-year shift in one year (2022 → 2023) raises questions about whether researchers are forecasting or reacting to recent headlines
- AI Impacts is a small organization with limited resources — how robust is the institutional infrastructure?
- Measurement without advocacy means her work gets used by all sides — useful for discourse but means she doesn't steer it
- "Human-level AI" remains ambiguously defined even within the surveys — different respondents may be answering different questions
Rich's take
- (your synthesis here)
converts-from: personas/katja-grace.md · schema v1 · AI & Society domain