Who she is
Affiliation: Co-founder and Lead Researcher, AI Impacts — a research project "based at the Machine Intelligence Research Institute" (MIRI); donations to it are made to MIRI and "automatically flagged for AI Impacts" (aiimpacts.org/about, donate page). Role check run 2026-09-21: no job change found — she is introduced as "founder of AI Impacts" on a two-part podcast aired 2026-06-15 and 2026-06-22, and is first author on the project's newest paper, dated September 2026. She started working on AI risk "a lot" in 2014 after being introduced to the topic in 2008 (her words, AI and You #313). Team today: herself plus two support staff (Caitlin Elizondo, Jimmy Rintjema); the past-contributors list on the same page has 18 names, including Paul Christiano and Daniel Kokotajlo. TIME100 AI, 2024 (TIME, 2024-09-05). Personal blog: world spirit sock puppet. Grew up in rural Tasmania (New Yorker profile, as quoted back to her on #314).
One-line position (v1, carried): 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.
One-line position (2026, what the record actually shows): Two positions in one person. The instrument still says what v1 said — four iterations of the Expert Survey on Progress in AI (2016, 2022, 2023, 2024), the 50%-year for human-level AI falling 2061 → 2059 → 2047 → 2042 (ESPAI 2024 paper, Sep 2026). The person has been an advocate since December 2022 ("Let's think about slowing down AI"), signed the 2023 pause letter, put her own chance of catastrophe at 19% in 2023 and "maybe like 50 percent" in June 2026, and now argues AI unemployment and AI extinction "are often the same." v1 described only the instrument.
Discipline & technical bet
Survey methodologist and technology forecaster, not an ML engineer. The instrument: email every author published at six top AI venues (NeurIPS, ICML, ICLR, AAAI, JMLR, IJCAI), pay them $50, ask the same questions "with minimal modifications" every iteration, and aggregate the answers into probability curves. Two deliberately different definitions of human-level AI are asked side by side — HLMI ("when unaided machines can accomplish every task better and more cheaply than human workers") and FAOL (every occupation fully automatable) — because, as she put it in June 2026, "the different descriptions of it ended up a hundred years apart in people's predictions" (#313). The 2024 survey (fielded 2024-12-09 to 12-24; 19,874 emails; 2,052 responses, a 10% rate; 1,580 kept after an out-of-scope recruitment error was corrected) is the fourth in the series and was published as a 48-page paper in September 2026 — a 21-month lag the authors themselves flag: "The survey we report on in this paper was conducted in December 2024, over 18 months ago. The field of AI moves fast, and so the opinions of AI researchers at the time of publication are likely to be substantially different in some ways from what we report."
The technical wager beneath the surveys is older and less known: AI Impacts "was mostly focused on doing other things, like trying to figure out numbers more directly" — decomposing hardware vs. algorithmic contributions to progress, and studying how long other capabilities took to cross the human range. Her example: chess AI "took decades to cross the human range," which she reads as evidence against the fast-takeoff picture ("I think we've been vindicated already," #314). The 2024 survey adds a matching question — respondents estimated how much less progress they would expect with half the funding, half the researchers, or a five-fold instead of ten-fold hardware cost decline. There is no proprietary model; her bet is that a slow, boring, repeated instrument beats any one clever forecaster's curve.
Key claims (Says)
- Empirical (v1): AI researcher timelines are shortening dramatically — the 2023 survey found a 50% chance of human-level AI by 2047, 13 years earlier than the same cohort's estimate a year before. Held — and extended. The 2024 survey moved the HLMI 50%-year to 2042: "the year by which HLMI has a 50% chance of arriving moved up from 2061 in 2016 to 2059 (2022), 2047 (2023), and 2042 (2024). Over the course of eight years, the forecast horizon shrank from 45 years out to 18, getting about 3.4 years closer for each year that passed" (ESPAI 2024, p.1). The 2023 figures check against the primary: 2,778 respondents; 2047; 13 years earlier than 2022 (arXiv 2401.02843, v3 2025-10-08; published in JAIR 84, 2025). See the conflict logged under predictions: her own June-2026 preview of this result said the opposite.
- Empirical (v1): 10% chance of human-level machine performance by 2027, from the 2023 survey. Held — re-confirmed unchanged. The 2024 survey's comparison table: "HLMI: 10% probability 2027 2027 0 yr" — the tail did not move while the median moved five years (ESPAI 2024, Table 1). Narrow-task version: "for all but one task, the forecast reached a 10% probability of becoming feasible by 2027 at the latest" (the exception, an unaided solution to a long-standing math problem, reached 10% in 2028).
- Empirical (v1): Expert surveys are the empirical backbone of timeline discourse — her series is the most-cited source for "what do AI researchers actually expect." Graded in two halves, per the mechanism/valence rule. Held on citation share: Google Scholar shows 1,973 total citations, 1,418 for the 2018 JAIR paper and 262 for the 2025 JAIR paper (Scholar, read 2026-09-21); the 2024 paper calls the 2023 iteration "the largest peer-reviewed survey of AI researchers ever" and lists TIME, Nature Briefing, the NYT and The Economist among citers. Drifted on currency: a 21-month fielding-to-publication lag, the authors' own "substantially different" caveat, and a faster rival — the Forecasting Research Institute's LEAP panel, fielded 2026-04-20 to 05-11 and published 2026-06-02 (expert median for AGI 2050; superforecasters 2047; LEAP Wave 8). Most-cited is no longer the same as most-current.
- Empirical (v1): Median estimates mask wide disagreement. Held. The 2024 abstract: "the mean aggregate forecasts for both have assigned non-negligible probabilities to them occurring within ten years, and to them not having occurred within two centuries." On risk: median 10%, mean 18.2%, and 51.1% of respondents at ≥10% for extinction or similarly severe disempowerment (Table 4). An outside reanalysis of the 2023 data: "For automation by 2048, the middle half of experts assigned it a probability between 1% and a 60%" (Grilo citing Adamczewski, 2025-01-09).
- Empirical (v1): Methodology matters — framing ("human-level AI" vs "full automation of labor") changes the answer significantly. Held — strengthened by her own data and her own words. Same survey, same respondents, 2024: HLMI 50% by 2042, FAOL 50% by 2098 — a 56-year gap from wording alone (2023: 2047 vs 2115, 68 years; the gap narrowed by 12 years in one iteration, which the file will watch). She now says it plainly: definitions "that I thought should be basically at the same time" landed "a hundred years apart" (#313). Note the reporting critique: "task automation (HLMI) numbers have been far more widely reported than occupation automation (FAOL)" (Adamczewski via Grilo) — and Tom Adamczewski is now a co-author of the 2024 paper.
- Empirical → normative (new, 2026-04-27): "AI unemployment and AI extinction are often the same" — "If powerful AI agents have their own agendas, those agendas will win out. They might win out in one crushing step, or win out in a trillion small familiar ways" (WSSP). On the podcast: humans losing control of resources "doesn't look like science fiction particularly, it looks like starvation" (#313). Ungraded — a framing, not a forecast; the observable is whether displacement data and control-loss incidents start being reported as one story.
- Normative (2022-12-22, restated 2026): slowing AI should be seriously considered — "the naive first-pass guess should be a strong 'probably'" (Let's think about slowing down AI); April 2026: "Common wisdom says that it is incredibly hard to coordinate to not build more dangerous AI… I don't feel like they are bringing their A game to this" (An easy coordination problem?); June 2026, on regulation: "it would be much better for someone else to step in and do that" (#314). Held — still endorsed, behavior matches (signature-framing check passed: the 2022 essay, the 2023 letter signature, and the 2026 posts all point one way).
- Normative (new, 2026): "AI is not an arms race but a Trojan horse race" — "the thing at the end of the race is creating new entities that are plausibly going to destroy the people creating it"; the race framing "gives them too much credit" (#313). Ungraded — rhetoric with a testable corollary: if leading labs coordinate a pause, the arms-race model was the wrong one. Check attached below.
Notable predictions — with falsifiable checks
- (2024-01, v1) Researchers' median for human-level AI: 2047 — "the estimate itself is the data point." Drifted — superseded within one iteration. The next survey said 2042. Check on the underlying event: HLMI as defined, credibly claimed and independently measured by 2042-12-31. Interim check on the series: by the next published ESPAI, has the median moved below 2037 (the 3.4-years-per-year trend continuing) or stalled near 2042 (her June-2026 "trend did not continue")?
- (2024-01, v1) 10% chance of human-level AI by 2027. Held — restated unchanged in the 2024 survey. Check: 2027-12-31. Only the 10% branch is gradeable: a credible HLMI claim by then makes the 2042 median badly late; absence teaches nothing about a 10% call.
- (2022, v1) Previous survey median ~2060; the 13-year shift in one year is itself a finding and "raises question of whether estimates are stable or reactive to recent demos." Held. 2022 = 2059 (738 respondents, 17% response; AI Impacts wiki; reported as 2060 in 2023, restated 2059 after data-cleaning changes). The reactive-vs-forecasting question got partial evidence in 2024: the headline HLMI year fell five more years while the 39 narrow-task forecasts barely moved ("33 of the 39 moved by less than two years in either direction"). The aggregate reacts; the task-level answers are sticky. Check: the next iteration's task-level vs headline movement.
- (2026-06-15, hers) "We actually have one more survey after the 2023 one that's not out yet, but I can tell you that the many years per year trend there did not continue… the newest forecast for the 50 probability of all the tasks definition is 2048" (#313, read live: "Let me just look at the data here"). Wrong on the published number. The paper she first-authored three months later reports 2042, a five-year drop, and is titled "sooner than we thought." Two sources conflict; the file trusts the published primary — later, peer-authored, and explicit that "small changes to code and data cleaning between surveys shifted the aggregate slightly" — over a live read of an unfinished analysis. But the gap is six years and the sign flipped; whether she was reading a different aggregate (median-of-medians vs mean CDF) is unknown. Flagged for Rich.
- (2026-06-22, hers) "quite soon we're hoping to have out the newest survey results" (#314). Held — PDF dated September 2026, live on aiimpacts.org by 2026-09-21.
- (2026-06-22, hers) Personal p(doom): "Well, it varies. I'd say maybe like 50 percent" (#314); June 2023: "I came up with like 19% chance of the world being destroyed" (Hear This Idea, 2023-06-10). Unfalsifiable as stated — no horizon. Tracked as a trend (19% → ~50% in three years) and graded as a (number, horizon) pair when she gives one. Note her own gloss: she wants people to state "the default p(doom) and also… how much you expect you can change it," and she is "pretty optimistic about changing the situation."
- (2026-04-07, hers, implicit) Coordination among the handful of leaders who profess concern is "fairly tractable." Live. Check: by 2028-12-31, any negotiated frontier-pace agreement or verified mutual slowdown among top labs or between the US and China (Kokotajlo's AI 2040: Plan A is the vehicle on the bench). Absence through 2028 grades the "easy" framing Wrong; a signed deal grades the arms-race model — and her critics — Wrong.
- (2024-12 survey, gradeable narrow tasks) Respondents put 50% on several tasks within two to three years — e.g. "Win World Series of Poker" 50% by 2026; "Answer Googleable factoid questions" and "Group unseen objects into classes" 50% by 2027 (ESPAI 2024, task table). Live. Check 2027-12-31: count of the 39 tasks credibly achieved vs the aggregate's 50%-years; a systematic early/late bias here is the cleanest calibration test the instrument offers.
Revealed behavior (Does)
- Time went into one more iteration and a very slow write-up. Fielded the 2024 survey in a two-week window in December 2024, corrected a recruitment error post hoc (2,052 → 1,580), and published 21 months later with eight co-authors, several of them outside collaborators — including a former critic (Adamczewski) and David Krueger — rather than the staff researchers of earlier years. AI Impacts' own Substack posted once in 2025 (the methodology FAQ, 2025-10-31; archive).
- Tried to delegate the survey's email collection to AI and regretted it: "there were a lot of errors — and then it was hard to figure out what had happened… I feel uneasy turning things over to AI, partly because I don't want to contribute to this whole situation" (#313). Behavior consistent with the stated view.
- Signed the FLI pause letter (2023): "I did sign it, yeah… if it fails but gets a bunch of attention, I think that would be helpful for the next iteration" (Hear This Idea, 2023-06-10). FLI is also a named funder of the 2024 survey.
- Ran a persuasion experiment, not just a survey: with Nathan Young, showed ~570 people arguments that AI is risky; baseline p(doom) "24% average, 11% median"; arguments alone raised p(doom) "by about 1.3%"; with counterarguments, three times as many moved down as up (2024-09-04), after compiling "Ten arguments that AI is an existential risk" (2024-08-13). Measuring whether advocacy works is advocacy's R&D.
- Publishes daily, mostly not about AI. Twelve personal posts between 2026-06-22 and 07-27 (elephant seals, "Therapy notes 10 June 2026," fatness, past lives); three of them on AI (WSSP archive). Column-inches are personal-essayist; the AI work is the day job, not the writing habit.
- Money: no commercial position found. The project is funded by "Coefficient Giving, Jaan Tallinn, Future of Life Institute (FLI) and others" (ESPAI 2024 acknowledgements); the donate page names Open Philanthropy (now Coefficient Giving) and the Survival and Flourishing Fund. Dollar amounts were not verifiable this cycle (the Open Phil grants database now redirects to coefficientgiving.org and returned 403).
- Hires: none visible; the team is one researcher plus two support roles, down from a bench of contributors that once included Kokotajlo, Stein-Perlman, Korzekwa and Heninger.
Feels
A resource-conflict fear with a plain-language ending: "if there are much more powerful agents in the world who want those resources for some other goals that are not us, then I think we're in trouble" — and it "looks like starvation." Impatience with the room ("I'm not in the habit of reading the room. I'm more like, I'll calculate it for myself and then be like, 'People need to know about this'"). Wants to be true: that "the world will wake up more," that coordination is easy, and that the survey — which she "doesn't enjoy running" (TIME, 2024) — keeps people from being "more wrong about what AI researchers think."
Hears
The Berkeley EA / rationalist safety cluster end to end: MIRI (institutional host), Christiano (early AI Impacts collaborator, per Wikipedia), Kokotajlo (former contributor), Yudkowsky (whose fast-takeoff picture she argues against), Nathan Young and Tom Adamczewski (co-authors), David Krueger; funders Coefficient Giving, Jaan Tallinn and FLI; and, through the survey, ~1,500–2,800 AI-conference authors a cycle — the one input channel on the bench that is not the cluster.
Sees
The only minimally-modified longitudinal record of what published AI researchers expect, 2016–2024, plus the non-response and demographic data that lets her test her own selection bias (respondents who had thought least about AI's social impact had the same 10% median as everyone but the least attentive — "reassuring evidence against the median being affected by heavily involved people being more likely to take the survey," ESPAI 2024 §3.9). She also sees the framing effect from the inside, at full size: same people, same week, 2042 vs 2098.
Incentive map
Re-derived this cycle; v1 carried none. Host: AI Impacts is legally housed at MIRI — donations go to MIRI, her Scholar profile is verified at intelligence.org — and MIRI's 2024 position is that "absent an international government effort to suspend frontier AI research, an extinction-level catastrophe is extremely likely" (MIRI, 2024-12-02). The field's "neutral instrument" is fiscally inside its most pause-committed organization. Funders: Coefficient Giving (formerly Open Philanthropy — the field's largest safety funder and Cotra's former employer), Jaan Tallinn (SFF's principal donor), FLI (organizer of the pause letter she signed). Selling: survey credibility. The numbers that get quoted — 18% mean, "51% assign at least a 10% chance," "sooner than we thought" — are the ones that raise the survey's value to these funders; the paper's own caveats ("unclear how valuable their predictions are for accurately forecasting outcomes"; opinions "likely to be substantially different" by publication) are the counterweight, and they are hers. Cannot say without cost: that a 10% response rate (down from 17% in 2022 and 15% in 2023) makes the timeline numbers soft; that the instrument is too slow to inform a 2026 decision; that her ~50% p(doom) sits on top of a survey whose median respondent says 10%. Household: none found. Rule-25 discount: low on published survey numbers (methods, question text and code are public; a 2024 reanalysis open-sourced the codebase — 2024-12-16); moderate on interpretation and titling ("sooner than we thought" vs her own June "trend did not continue"); moderate on the advocacy claims, which are argued from priors, not from the instrument.
Theories aligned with
- Empirical forecasting methodology (v1) — the survey branch, distinct from Cotra's compute anchors and time horizons and from Kokotajlo's scenario models
- AI alignment / x-risk — v1 said "adjacent… primarily a measurement voice, not an advocate." Wrong: a slowdown advocate since 2022-12, pause-letter signatory, personal p(doom) ~50%. She is inside the theory, holding its measuring stick.
- Automation / displacement — new: her 2026 "unemployment and extinction are often the same" collapses the labor and control-loss stories into one resource-conflict mechanism
- Slow-takeoff / continuous-progress view — the chess-crossing-the-human-range finding; she reads the LLM era as confirmation that human-level is "spread out," not a bright line
What she's reacting against
- Gut-feel timelines that aren't anchored to any systematic evidence (v1)
- Tribal polarization of AI discourse (doom vs. hype) — insists on empirical measurement (v1)
- Overconfidence in any single forecasting methodology — surveys are one input, not the answer (v1); her 2024 paper: "it is unclear how valuable their predictions are for accurately forecasting outcomes"
- The assumption that expert opinion is stable — her data shows it's shifting fast (v1)
- NEW: the arms-race frame — a "Trojan horse race" in which "companies are bringing the trojan horses into their own compounds" (#313)
- NEW: the assumption that coordination is impossible or "miraculous world government" (2022) and that adversaries cannot act in their own interest: "They seem to take for granted that certain entities—especially Trump and China—would be acting against their own interests, were they to cooperate" (2026-06-26)
- NEW: fatalism among people with high p(doom) — "they're very pessimistic about changing it"; and the objection to quoting p(doom) at all: "if someone shouts 'fire'… 'Could you maybe say fire more likely than not?'" (#314)
- NEW: treating unemployment and extinction as separate policy problems
Where she overlaps / splits (with Rich)
- Overlaps with Cotra on trying to ground timelines in data; splits on method — Cotra uses compute anchors (now measured time horizons), Grace uses expert surveys (v1). Sibling-run lead checked: Cotra's file cites no AI Impacts survey data, so no shared dataset; the connection is personnel and money — Cotra's husband Christiano was an early AI Impacts collaborator, and Coefficient Giving (Cotra's employer for ~8 years) funds AI Impacts. Both now argue coordination/transparency is tractable; Cotra's instrument reads fast and slow-biased, Grace's reads slow and lags 21 months
- Overlaps with Tetlock-style forecasting community on methodology; splits by being domain-specific to AI (v1). Now a live rivalry: FRI's LEAP panel runs monthly with superforecasters in the sample; its 2026 expert median (2050) is eight years later than her 2024 survey's (2042)
- Splits with confident short-timelines voices (Aschenbrenner, Kokotajlo) — her data shows wide uncertainty, not convergence on "soon" (v1). Update: Kokotajlo came out of her shop (past-contributors list), and his current medians (automated coder Nov 2027, superintelligence Mar 2029) sit 13–15 years inside her survey's 2042. Yet her personal line — a "substantial risk… some time quite soon perhaps," p(doom) ~50% — is closer to his than to her respondents'. The split is instrument vs person
- Splits with dismissive voices (Marcus) — the data shows researchers moving timelines forward, not standing pat (v1). Update: Marcus now calls for "Pause OpenAI" (2026-09) — converging on her 2022 slowdown position from the opposite capability premise; she has the survey he lacks, he has the megaphone she doesn't use
- Splits with Yudkowsky on takeoff shape — she cites his "idiots and Einstein right next to each other" picture as one AI Impacts set out to test and "been vindicated" against; overlaps on the value of stating numbers (she is "in favor of people quoting those numbers around" the year after he retired his)
- Overlaps with Christiano historically (early AI Impacts collaboration on brain-vs-supercomputer comparisons); the split now is institutional — he sits on the OpenAI Foundation board, she signs pause letters
- For Rich: use the survey as the floor of what the field admits, not as a forecast — and read her blog separately, as one more ~50%-doom advocate with unusually good manners about her own uncertainty
Track record
- Built AI Impacts into the most-cited empirical forecasting resource in the AI timeline space (v1) — 1,973 Scholar citations; the 2018 JAIR paper alone 1,418
- Survey series is referenced by virtually every serious timeline discussion — genuine field infrastructure (v1); the 2024 paper documents TIME, Nature Briefing, NYT and Economist coverage, and four iterations with "minimal modifications" is the longest such series that exists
- Methodology has been critiqued (survey framing effects, selection bias) but remains the best available systematic data (v1) — and she has absorbed the critics: Adamczewski's HLMI-vs-FAOL reporting critique (2025-01) became a 2024-paper co-authorship; the 2023 data was reanalyzed and open-sourced (2024-12-16); non-response checks are in Appendix A of every iteration
- Does not make strong personal timeline predictions — positions herself as the measurer, not the forecaster (v1). Wrong. She has stated a personal p(doom) on the record twice (19%, 2023; ~50%, 2026), calls the arrival of dangerous AI "quite soon perhaps," and has argued for slowing AI since 2022. She is a forecaster-advocate who also runs the field's survey
- Early slow-crossing study, undated in the transcript ("a while ago"; chess and other capabilities took years to decades to traverse the human range) — she claims vindication as LLMs sit "in the middle of" human-level for "a stretch" (#313, #314); a real, early, unfashionable call
- TIME100 AI 2024; the 2017/2018 paper was among the most-discussed scientific papers of its year (her bio)
- Misses in the window: a 21-month publication lag on the 2024 survey; a live preview number (2048) contradicted by her own published headline (2042); an AI-delegated recruitment step that "erroneously included people outside of the intended group"; response rate down to 10%
- Pattern: strongest at building and defending a boring instrument; weakest at speed; personal views have moved faster than the instrument she runs
Empirical vs normative
- Empirical: survey design, data collection, methodological analysis (v1) — plus hardware-vs-algorithm decomposition and capability-crossing studies
- Normative — v1 said "minimal… she rarely advocates specific policies." Wrong. Dated record: "Let's think about slowing down AI" (2022-12-22); pause-letter signature (2023); "Ten arguments that AI is an existential risk" (2024-08-13); "An easy coordination problem?" (2026-04-07); "it would be much better for someone else to step in" on regulation (2026-06-22). v1's "this makes her unusual in the field" was built from her papers, not her blog
- What is still unusual: she keeps the ledgers separate — the survey papers carry no policy ask, and the advocacy never cites her own p(doom) as data
- Conflict note (CONVENTIONS rule 25): MIRI-housed, funded by the pause-side philanthropies; low discount on published numbers, moderate on titling and on advocacy
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 (v1). Held: 2042 vs 2098 in one survey is the framing effect measured; her own title ("sooner than we thought") uses the shorter one
- The 13-year shift in one year (2022 → 2023) raises questions about whether researchers are forecasting or reacting to recent headlines (v1). Drifted — partly answered: the next shift was five years, and the 39 task-level forecasts barely moved while the headline did; reaction lives in the aggregate question, not the task questions
- AI Impacts is a small organization with limited resources — how robust is the institutional infrastructure? (v1) Held — worse than v1 knew: one researcher, two support staff, a 21-month lag, and a host (MIRI) that has pivoted to advocacy
- Measurement without advocacy means her work gets used by all sides — useful for discourse but means she doesn't steer it (v1). Wrong premise. She advocates. The real weak spot is the inverse: an advocate with a ~50% p(doom) runs the instrument the whole field cites as neutral, from inside MIRI. The mitigations — public code, published question text, appendix-level bias checks — are strong; the optics are not
- "Human-level AI" remains ambiguously defined even within the surveys — different respondents may be answering different questions (v1). Held; she says so herself
- NEW: response rate 17% → 15% → 10% across 2022–2024, with a recruitment error corrected after fielding; the non-response appendix is reassuring on the risk questions, less so on timelines
- NEW: the June-2026 "2048 / trend did not continue" vs September-2026 "2042 / sooner than we thought" discrepancy is unexplained; until it is, the headline year carries a six-year error bar of its own
- NEW: p(doom) ~50% with no horizon, from someone whose median respondent says 10% — the two numbers are never reconciled in public
- NEW: no 2025 survey was fielded (the 2024 one is "one more survey after the 2023 one"); if the series goes biennial with two-year lags, its role as "backbone" passes to LEAP by default
Rich's take
- (your synthesis here)
Delta log
2026-09-21 — v1→v2 migration + validation (bulk run)
- Grades: 11 Held / 3 Drifted / 5 Wrong (plus 1 unfalsifiable item tracked, not graded; 2 live predictions with dated checks; the 10%-by-2027 item sits under both Key claims and Predictions and is counted once). Four of the five Wrongs are one miss restated in four v1 places — "measurer, not advocate/forecaster" — graded where each lives, counted separately because each line carried its own weight in the file.
- Wrong: "measurer, not forecaster" (Track record), "normative content is minimal" (Empirical vs normative), "measurement without advocacy" (Weak spots), "not an advocate" (Theories) — Let's think about slowing down AI, 2022-12-22; Hear This Idea, 2023-06-10 (19%; "I did sign it"); 2026-04-07; AI and You #314, 2026-06-22 ("maybe like 50 percent"; "much better for someone else to step in"). Fifth Wrong: her June-2026 preview "2048… the many years per year trend there did not continue" (#313, 2026-06-15) vs the published 2042 (ESPAI 2024, Sep 2026).
- Drifted: 2047 median (superseded by 2042 — ESPAI 2024, Table 1); "empirical backbone" on currency (21-month lag; LEAP Wave 8, 2026-06-02); the reactive-vs-forecasting weak spot (task-level stability vs headline movement, ESPAI 2024 §3.1–3.2).
- Held: timelines shortening (2042; 3.4 years/year over eight years); 10% by 2027 (unchanged in 2024); "backbone" on citation share (Scholar); wide disagreement (two centuries of spread; 51.1% at ≥10% risk); framing matters (2042 vs 2098); slowdown still endorsed with matching behavior; "results quite soon" (published Sep 2026); 2022→2023 shift finding; framing-sensitivity weak spot; small-org weak spot; HLMI-ambiguity weak spot.
- Most surprising delta: the bench's "neutral measurer" put her own p(doom) at ~50% on air in June 2026, up from 19% in 2023, while her survey's median respondent says 10% — and v1 had described her as having almost no normative content.
- Employer/role check (mandatory): no change — Co-founder and Lead Researcher, AI Impacts, based at MIRI (about; donate); introduced as founder on 2026-06-15/22; first author on the September-2026 paper. Hats counted: one, plus the personal blog. Incentive map built from scratch: MIRI host, Coefficient Giving / Tallinn / FLI funders (paper acknowledgements), no commercial or household conflict found.
- Numbers re-verified against primaries: 2,778 / 2047 / 13 years (arXiv 2401.02843); 738 / 17% / 2059 / 5% (2022 wiki); 45 years / 120 years (arXiv 1705.08807); 1,580 / 10% / 2042 / 2098 / 18% / 10% / 51% / 72% (ESPAI 2024 text extracted locally from the PDF). v1's "~2060" for 2022 stands (2059 after re-cleaning).
- Sibling-run lead (Cotra shared data source): not confirmed as a data link — Cotra's file cites METR and her own posts, not ESPAI. Confirmed instead as a personnel/funding link: Christiano collaborated on AI Impacts (Wikipedia); Coefficient Giving funds both. Logged in Overlaps and Incentive map.
- Post-review sources found: yes — the September-2026 paper, two June-2026 podcast transcripts, three 2026 blog posts (Apr 7, Apr 27, Jun 26), the 2025-10-31 methodology FAQ, and LEAP Wave 8 — all after the file's 2026-06-16 review date or missed by it.
- Not verified this cycle: grant dollar amounts (Coefficient Giving pages returned 403); the exact September-2026 publication day; whether "2048" on #313 was a different aggregate; whether a 2025 or 2026 ESPAI is planned.
- Tier: semiannual — the instrument updates yearly at best with a 21-month lag and her personal positions moved over years, not months; wake triggers catch the discrete events (new survey, host/funder change, restated p(doom), a pause proposal). Confidence unchanged at medium: grades rest on primaries, but one unexplained six-year discrepancy sits on the headline number.
Sources
- 2026 primary — survey: "Advanced AI according to 1,580 researchers: uncertain, unsafe, and sooner than we thought" — Grace, Simonelli, Weinstein-Raun, Krueger, Young, Duffy, Adamczewski, Reeve, Beck (AI Impacts, September 2026; text extracted from the PDF) · aiimpacts.org (lists the paper as newest work, read 2026-09-21)
- 2026 primary — her own words: AI and You #313 transcript (aired 2026-06-15) · AI and You #314 transcript (aired 2026-06-22) · episode page · her post on the episodes (2026-06-26) · An easy coordination problem? (2026-04-07) · AI unemployment and AI extinction are often the same (2026-04-27) · Why are adversaries assumed to be incapable of responding to AI risk? (2026-06-26) · WSSP archive
- Survey series (primary): Thousands of AI Authors on the Future of AI (2023 survey; arXiv v1 2024-01-05, v3 2025-10-08; JAIR 84, 2025) · 2022 Expert Survey on Progress in AI (AI Impacts wiki) · When Will AI Exceed Human Performance? (2016 survey; JAIR 62, 2018) · FAQ: ESPAI methodology (2025-10-31) · Reanalyzing the 2023 survey — open-source codebase (2024-12-16) · AI Impacts blog archive
- Advocacy and experiments: Let's think about slowing down AI (2022-12-22) · Hear This Idea — slowing down AI and whether the x-risk case holds up (2023-06-10) · Ten arguments that AI is an existential risk (2024-08-13) · What happens if you present 500 people with an argument that AI is risky? (2024-09-04) · FLI podcast (2024-03-14)
- Role, org, funding: AI Impacts — About (team, past contributors, "based at MIRI") · AI Impacts — Donate (MIRI receives; funders named; page text stale, references 2023) · MIRI 2024 end-of-year update · TIME100 AI 2024 profile (2024-09-05) · katjagrace.com · media list (2024–2025 press) · Google Scholar · Wikipedia — Paul Christiano (AI Impacts collaboration)
- Critique and rivals: Grilo / Adamczewski — Experts' AI timelines are longer than you have been told? (2025-01-09) · Forecasting Research Institute — LEAP Wave 8 (2026-06-02)
- v1 sources carried: survey papers (2024, with others; prior surveys 2016, 2022) · aiimpacts.org · katjagrace.com · Google Scholar
migrated v1→v2 2026-09-21 · backup: _archives/katja-grace.html.bak-20260921 · converts-from: people/katja-grace.md · schema v1 → v2 · AI & Society domain