Panel verification, October 2026

Checked 2026-09-30 by claude-agent, under amendment A13 (three tiers with quotas) and methodology-v2 section 3. Re-marked 2026-10-01 under my rulings in amendment A14 (rulings 1 to 6); the section "Freeze candidates as of 2026-10-01" at the end is the current summary, re-marked again 2026-10-02 under amendment A16 (the looser newspaper rule; see "Re-marking under A16" below), and re-marked 2026-10-06 after my rulings of that day and the closing of the three unclear core cases (see "Closing the three unclear cases (2026-10-06)" and "Freeze candidates as of 2026-10-06" at the end, which is now the current summary). The panel was frozen by me on 2026-10-06 (amendment A18, "Freeze and merge") from the 2026-10-06 summary; this document is the evidence behind it. The frozen tiers and quota tags are in data/speakers.csv (25 core, 12 cohort); the wide list is data/wide_list_candidates.csv (308).

Full evidence (quotes, every link, every query run) is in panel-evidence-2026-10/. The tables below give one link per criterion; the evidence files give the rest.

The rules being checked

Core quotas (minimums among those frozen): 6 skeptic (pre-2017 statements reject or doubt AGI within decades), 3 non_us_uk (based outside the United States and United Kingdom when making their pre-2017 statements), 2 philosopher (philosopher or cognitive scientist), 3 frontier_lab (led, ran, co-founded or held a senior research role at a company building the most capable models of its time).

Verdicts: PASS; PASS (thin) = meets the rule only on weak evidence, named in the notes; UNCLEAR = depends on a ruling or a missing check; FAIL = does not meet the rule as searched. Nothing thin is rounded up to a plain pass.

How far to trust this

• Blocked outlets. The search and page-reading tools could not reach the New York Times, the Guardian, The New Yorker, Wired, the BBC, the Economist, the Financial Times, the Wall Street Journal, The Atlantic or New Scientist. Pieces from them are marked "seen in search result" or were read in copies elsewhere. Several C3 failures are probably the tools, not the record, especially for UK-based candidates (Shanahan, Wooldridge) and Pinker. A library database check would settle them.
• "Seen in search result" items count toward C3 here. They were not opened; the search showed the outlet, the date and the name.
• One copy carries four passes. The 2015 New Yorker profile "The Doomsday Invention" was read in full from a course-page printout. It names Kurzweil, Bostrom, Hassabis and Hinton, and counts toward all four.
• Page reading went through a summarising tool. Quotes should be checked against the live page before going into data/statements.csv.
• Earlier in the session one helper fetched on-site search pages after the search limit ran out; nothing in these tables depends on them.

Rulings applied (A14, 2026-10-01)

1. Self-written pieces count toward the three national pieces, but for at most one of the three. This turns Marcus, Russell and Brooks into passes, and confirms Tegmark.
2. Skeptic window is "2017 or earlier". Kai-Fu Lee gains the tag ("hundreds of years, if ever", 2017-06-24); LeCun's tag is strengthened by "still quite far from building artificial general intelligence" (2017, lecun-2017-01).
3. Co-written statements count for C2 when the person is a named author: Davis and Brynjolfsson meet C2 on co-written items.
4. Dennett is dropped from the core; he is added to the wide list.
5. Tags. frontier_lab applies if true at the time of any statement in the record, with dates in the justification. philosopher means primary role: Hinton and Hassabis are not tagged. Chalmers is non_us_uk. Suleyman is not a skeptic. Under A15 ruling 8, Kurzweil (Google engineering director from 2012), Chollet (Google senior engineer 2015 to 2024), Kai-Fu Lee (01.AI from 2023) and Karnofsky (OpenAI board 2017 to 2021, Anthropic staff from 2025) are tagged frontier_lab. Kokotajlo (trained in philosophy, works as an AI forecaster) is not tagged philosopher (A15 ruling 8).
6. Cohort prominence is information only.

Still open: archive checks through the Guardian and New York Times APIs (application programming interfaces, their official data services) once I have keys (ruling 7).

Core panel candidates

Abbreviations in the table: "co-wr." = co-written; "search" = seen in search result, not opened; "2nd" = secondhand.

Quota totals among core candidates who pass (as of 2026-09-30, superseded below)

Counted over the 16 with PASS or PASS (thin): Kurzweil, Bostrom, Hanson, Hinton, LeCun, Hassabis, Altman, Ng, Musk, Chalmers, Brynjolfsson, Schmidhuber, Kai-Fu Lee, Walsh, Tegmark, Suleyman.

What this shows. The dated prominence rule does what A13 predicted: it rewards people who were loud early, mostly lab founders and singularity forecasters, so the skeptic quota is the one at risk. Six of the skeptic candidates fail on C3 (Pinker, Floridi, Davis, Etzioni on C2, Wooldridge, Mitchell on C1), several probably because the tools could not reach UK and US newspapers. The core reaches 19 with the byline ruling, and 23 if the four unclear cases (Yudkowsky, Bengio, Dennett, Cowen) are all resolved in their favour, short of about 25; an archive check on the C3 failures (Pinker, Shanahan, Wooldridge, Goertzel, Floridi, Legg) is the cheapest way to grow it, and would mostly add skeptics.

2020s cohort candidates

Cohort count: 8 pass (Amodei, Aschenbrenner, Kokotajlo, Cotra, Huang, Leike, Mollick, Chollet) and Sutskever is unclear, against a target of 10 to 12. Three of the passes are thin. Updated 2026-10-06: 12 pass (Sutskever by my corrected ruling; Liang, Mensch and Narayanan from the search I asked for). See "Freeze candidates as of 2026-10-06".

Note on C3 for the cohort. Information only, confirmed by me (A14 ruling 6).

Wide list

Built by the enumerable rule and saved as data/wide_list_candidates.csv (name, source list, link, notes): 184 names by the rule after removing the 18 who are core or cohort candidates, plus Dennett (A14 ruling 4): 185.

time.com refuses scripted requests, so the TIME lists could only be read through a page-reading tool, which returned about half of each list (mostly the "Leaders" and "Innovators" sections). The missing names were not filled in from memory; the gap is recorded on each row. The full lists need a manual export before the wide list is used. Closed 2026-10-06: the full lists were read from Wayback Machine copies (details and links in panel-evidence-2026-10/gaps-2026-10-06.md); every earlier name was confirmed, 126 were added, and each row's notes now cite the copy. The target is now whatever the rule yields (A14 ruling 8). Daniel Dennett was added on 2026-10-01 when he was dropped from the core (A14 ruling 4), so the file holds 185 names.

Gaps to close before freezing

Guardian and NYT archive checks were run on 2026-10-01 and finished on 2026-10-02 (see below); still open: a 2024-or-later statement from Etzioni (GeekWire 2026 lead) and Shanahan (2024 and 2025 podcasts); the original 2016 Bengio interview; a transcript of Hassabis's 2010 talk; opening the three Cowen reviews; the full TIME100 AI lists.

Status 2026-10-06. Closed: Shanahan (2025 podcast), Bengio (original 2016 interview found, plus an explicit 2017 quote), Floridi (2017 Financial Times piece confirmed), the full TIME100 AI lists. Partly closed: one of the three Cowen reviews opened (Daily Beast); the Wall Street Journal and Economist reviews are behind a paywall and a security check. Still open: Etzioni's 2024-or-later statement (not found again); a transcript of Hassabis's 2010 talk (not attempted). Neither open item changes a verdict that the freeze depends on: Etzioni fails either way unless a statement is found, and Hassabis already passes thin.

Newspaper archive checks (2026-10-01)

Run with scripts/archive_check.py through the Guardian Open Platform and the New York Times Article Search API (A14 ruling 7, A15). For each person: the name with AI-future terms, pre-2018 (for C3) and 2024 onward (leads for C2). Every hit is in panel-evidence-2026-10/archive-checks.csv (1,184 rows). A hit counts toward C3 only if the person's full name is in the headline or abstract and the headline or abstract carries an AI-future term. Name in the headline with no abstract: flagged headline_only, not counted (3 hits). Name only in the article text: flagged named_in_text_only, not counted (1,037 hits).

The New York Times closed the connection twice. After the first, the run was resumed once; after the second (while checking Brynjolfsson's 2024-onward results) it was stopped, per the overnight rule. Six people were therefore not fully checked: Brynjolfsson (NYT 2024 onward only), Schmidhuber, Kai-Fu Lee, Walsh, Tegmark, Suleyman. All six already pass, so the gap cannot lower a verdict, but their newspaper evidence is not on the same footing yet.

Archive hits add evidence; their absence does not cancel pieces verified elsewhere. Three counted hits are not about AI futures on reading the headline (Goertzel 1999, Pearl 2012, Ng 2017 resignation) and were not used.

Rerun, 2026-10-02. The six people the stopped run missed were checked in full, with a 20-second pause between New York Times requests; no connection was dropped. No verdict changes. Schmidhuber's two unopened newspaper pieces are now confirmed by the archives (his pass stays thin only on the 2024 statement); Tegmark no longer relies on his own op-ed for the three pieces; Kai-Fu Lee gains nothing from these two archives and still rests on CNBC and MIT Technology Review. The file now has 1,362 rows.

What changed: LeCun moves from thin to full pass (NYT 2015 interview). Brooks no longer relies on his own essay (NYT 1999, 2002). Seven candidates who were waiting on these archives found nothing that counts and are now plain FAIL: Pinker, Floridi, Legg, Sutskever, Shanahan, Wooldridge, Goertzel. Kurzweil, Bostrom, Hassabis, Musk and Hinton gain archive pieces but were already passes.

Re-marking under A16 (2026-10-02)

A16 (decided 2026-10-02): a piece counts toward the three national pieces if the person is named anywhere in its text and its headline or abstract shows the piece is about where AI is going. The strict rule (name in the headline or abstract) had excluded 1,037 of the 1,184 archive hits.

How it was applied: from archive-checks.csv, pre-2018 rows flagged named_in_text_only with an AI term in the headline or abstract; then each headline read, and business news, obituaries and pieces about something else left out. The Guardian pieces were opened through the Guardian's data service to confirm the name is in the text and to see how the person is used. New York Times text is not available that way, so NYT pieces rest on the archive's own full-text match.

Agent readings, for me to confirm or overturn: counting Legg although he is named only in passing; counting jobs-and-automation pieces for Cowen as "where AI is going"; passing Yudkowsky on two edge outlets. Each is marked thin.

Freeze candidates as of 2026-10-02 (superseded by the 2026-10-06 summary below)

For my freeze. Nothing has been frozen. Updated 2026-10-02 after A16.

Core: pass now (22)

Kurzweil, Bostrom, Hanson, Hinton, LeCun, Hassabis, Altman, Marcus, Ng, Russell, Brooks, Musk, Chalmers, Brynjolfsson, Schmidhuber, Kai-Fu Lee, Walsh, Tegmark, Suleyman, and under A16 Legg, Cowen, Yudkowsky. Twelve pass on thin evidence (named in the table): Hanson, Hassabis, Ng, Musk, Chalmers, Brynjolfsson, Schmidhuber, Kai-Fu Lee, Tegmark, Legg, Cowen, Yudkowsky.

All four minimums are met. The core is three short of "about 25". The three added under A16 bring no skeptics: the skeptic candidates still fail or are unclear.

Core: unclear (3)

If all three passed: 25 people; skeptic 8, non_us_uk 6, philosopher 4, frontier_lab 11.

Core: fail (11)

Pinker (C3), Sutskever (C3; unclear for the cohort), Wooldridge (C3), Goertzel (C3), Chollet (C3; passes the cohort), Mitchell (C1), Etzioni (C2), Davis (C3), Pearl (C3), Karnofsky (C3; also fails the cohort rule). Dropped by ruling: Dennett.

2020s cohort: pass now (8)

Amodei, Aschenbrenner, Kokotajlo, Cotra, Huang, Leike, Mollick, Chollet. Unclear: Sutskever (whether a co-signed "hard to predict when" counts as a timeline statement). Quota tags among the 8: frontier_lab 3 (Amodei, Leike, Chollet). The target is ten to twelve.

Wide list

185 names (184 by the rule, plus Dennett). The TIME100 AI lists are still only partly retrieved.

Closing the three unclear cases (2026-10-06)

Checked 2026-10-06 by claude-agent at no cost (web search, page fetching, the Wayback Machine, and YouTube's caption files; no paid services). Full quotes, links and queries: panel-evidence-2026-10/gaps-2026-10-06.md.

my rulings applied (2026-10-06, amendment A17)

1. Legg, Cowen and Yudkowsky stay in the core, marked thin. This confirms the three agent readings listed under "Re-marking under A16": Legg counted although named only in passing; jobs-and-automation pieces counted for Cowen; Yudkowsky passed on two pieces from edge outlets.
2. Sutskever passes the 2020s cohort: "Not a timeline statement". my corrected ruling: the co-signed 2015 "hard to predict when" is not a timeline statement from before 2018, so he meets the cohort's first test. An earlier answer that day said the opposite; it was given to a question the coordinating session had framed backwards, and this is my ruling on the question framed correctly. Cohort C2 checked: the 2025-11-25 Dwarkesh Patel interview gives "I think like 5 to 20" (years) for a system that learns as well as a human, in his own words.
3. Freeze after the three gaps are closed, not before. They are now closed.
4. Both thin readings count (on the coordinating session's recommendation): Shanahan's 2025 remark on "whether we really are on the road" to human-level general intelligence is his 2024-or-later statement on whether AGI arrives, and Floridi's 2017 Financial Times column on robot rights is a piece about where AI is going under A16. Both stay thin. The core stays at 25.
5. "Search first, then freeze": before freezing, search for one to three more cohort members at no cost, with the rules exactly as written. Done the same day (below); the freeze is drafted as a proposed amendment (A18 in docs/amendments.md) for me's yes. Nothing has been frozen and data/speakers.csv is unchanged.

The three verdicts

Not proposed: a skeptic tag for Bengio. "Very far from" gives no timescale, so it does not meet "reject or doubt AGI within decades" as written. He stays a near miss.

Other gaps

• Etzioni (C2): still FAIL. The GeekWire 2026 interview still cannot be opened; the line quoted from it ("We haven't achieved artificial reliability") is about how dependable AI agents are, not about AGI.
• Cowen's three reviews: the Daily Beast review (2013-09-17) was opened and is about "genius machines" guiding our lives, which strengthens his pass; the Wall Street Journal and Economist reviews could not be opened.
• TIME100 AI lists: closed. Full lists read from Wayback Machine copies; 126 names added to the wide list, which then had 311 names (308 after the cohort search below).
• Hassabis's 2010 talk transcript: not attempted (it would not change his thin pass).

Cohort search (2026-10-06)

I asked for one to three more cohort members before the freeze, found at no cost with the rules as written (no public AGI timeline statement found before 2018; one dated 2024 or later), preferring people who broaden the cohort. Full evidence: panel-evidence-2026-10/cohort-search-2026-10-06.md.

Near misses: Sayash Kapoor (meets the rule on the same co-written essay, but adds no independent statement and would take the cohort to 13); Abeba Birhane (no 2024-or-later statement on AGI found, only criticism of hype); Katja Grace (probably fails the before-2018 test through the 2017 expert survey she led). Not checked: Yi Zeng, Andrew Yao, Xue Lan, Emily M. Bender.

The three new members were taken off the wide list, as the rule requires. The near misses stay on it: they were looked at only in a search for additions, and the wide list is built by its rule, not by who was considered.

Freeze candidates as of 2026-10-06

Frozen by me on 2026-10-06 (amendment A18, "Freeze and merge") exactly as listed here, with three readings chosen at the freeze: Liang tagged frontier_lab and Mensch not; Mensch's "I don't believe in AGI" counts as his 2024-or-later statement; Narayanan, not Kapoor. This replaces the 2026-10-02 summary.

Core: pass now (25)

Kurzweil, Bostrom, Hanson, Hinton, LeCun, Hassabis, Altman, Marcus, Ng, Russell, Brooks, Musk, Chalmers, Brynjolfsson, Schmidhuber, Kai-Fu Lee, Walsh, Tegmark, Suleyman, Legg, Cowen, Yudkowsky, and from 2026-10-06 Shanahan, Bengio, Floridi.

Fifteen pass on thin evidence (the reason for each is in the main table): Hanson, Hassabis, Ng, Musk, Chalmers, Brynjolfsson, Schmidhuber, Kai-Fu Lee, Tegmark, Legg, Cowen, Yudkowsky, Shanahan, Bengio, Floridi. Ten are full passes: Kurzweil, Bostrom, Hinton, LeCun, Altman, Marcus, Russell, Brooks, Suleyman, Walsh.

All four minimums are met, and the core reaches the "about 25" target. The skeptic quota has two to spare; if Ng's medium-confidence tag were dropped it would still have one.

Core: unclear (0)

None.

Core: fail (11)

Pinker (C3), Sutskever (C3; passes the cohort), Wooldridge (C3), Goertzel (C3), Chollet (C3; passes the cohort), Mitchell (C1), Etzioni (C2), Davis (C3), Pearl (C3), Karnofsky (C3; also fails the cohort rule). Dropped by ruling: Dennett.

2020s cohort: pass now (12)

Amodei, Aschenbrenner, Kokotajlo, Cotra, Huang, Leike, Mollick, Chollet, Sutskever (my corrected ruling, 2026-10-06), and from the search of 2026-10-06 Liang Wenfeng, Arthur Mensch, Arvind Narayanan. The cohort is at the top of its range of ten to twelve. Six are thin: Huang, Leike, Mollick, Liang, Mensch, Narayanan.

Wide list

308 names: 307 by the rule (the MIRI and AI Impacts predictions dataset and the full TIME100 AI lists 2023 to 2025, less core and cohort members), plus Dennett. The stream M media sample has not run, so names quoted there are not yet added.

Notes after the full collection (2026-10-12, ruled by me; amendment A28)

The full collection (pull request 4) read each member year by year and found earlier statements than this record had. None changes a verdict or a tier. Recorded here so the tables above are read with them.

Chollet's 2017 essay is not a public timeline statement. His 2017-11-27 essay "The implausibility of intelligence explosion" (recorded as chollet-2017-02b) says civilization "will create superhuman AI ... over timescales we can barely comprehend". Decided 2026-10-12: that is not a public AGI timeline statement, so the cohort's first test (no AGI timeline statement before 2018) still holds and Chollet stays in the 2020s group. His first timeline statement remains the 2025 "five-ish" years remark.

Earlier first statements found (the dates in the tables above stay as they were when the verdicts were given; the collection's rows are the record now):

Rendered from docs/panel-verification-2026-10.md at version a7bd17a.