Closing the three unclear core cases, 2026-10-06

Checked 2026-10-06 by claude-agent, at no cost: web search, page fetching, the Internet Archive's Wayback Machine, and YouTube's own caption files (downloaded with yt-dlp, no video). No paid services were used. Rules applied are the ones in docs/methodology-v2.md section 3 and docs/panel-verification-2026-10.md, with amendments A14 and A16. Every quote below was checked against the raw text of the page or caption file, not only through a summarising tool, unless marked otherwise.

Verdict words as in the verification document: PASS; PASS (thin) = meets the rule only on weak evidence, named; UNCLEAR; FAIL.

Murray Shanahan: C2 (a statement on whether or when AGI arrives, dated 2024 or later)

Verdict on C2: PASS (thin). Overall: PASS (thin).

• Counted: 2025-04-24. Google DeepMind: The Podcast, "Consciousness, reasoning and the philosophy of AI | Murray Shanahan", hosted by Hannah Fry. Published on the official Google DeepMind YouTube channel on 2025-04-24 (upload date read from the video's own record): https://www.youtube.com/watch?v=v1Py_hWcmkU. Source: the channel's uploaded English captions (speaker-labelled, so not machine captions), at about 27:00. Primary (his own spoken words).

  Hannah Fry asks whether good embodied AI would bring "a big up-step in intelligence". He answers:

  "Well, I think it might make a big difference because the large language models we have at the moment, it's really difficult to discern, actually, to be honest, right now, where the limits are for how good they're going to get, whether we really are on the road to producing general intelligence that is comparable to human general intelligence. [...] But it may be that that sort of general ability to really get things on a deep level, on a deep, common sense level, maybe, that that does still require a bit of embodiment."

  Earlier in the same episode (about 2:50): "Now, we don't know quite how things are going to unfold in the next few years because maybe robotics will progress rapidly as well in the way that language has in AI."

• Why thin. It is a statement on whether human-level general intelligence is coming (he says it is hard to tell whether current systems are on the road to it, and suggests a physical body may be needed), but it gives no date, and it is partly about the route. Under the codebook (A15 ruling 4) a route claim alone would be coded unexpressed for endorsement; here he also says the outcome itself is uncertain, which is why it is counted. The same footing as Bostrom's 2026 paper and Chalmers's 2026 talk, which pass thin on statements without a dated forecast.

• Checked and not counted: 2024-07-14. Machine Learning Street Talk, "There are monsters in your LLM. (Murray Shanahan)", https://www.youtube.com/watch?v=ztNdagyT8po (upload date 2024-07-14 from the video's record; automatic captions). At about 2:02:00 he says today's language models "pass the spirit of the Turing test" because "they do really have human level conversational skills" and have "attained more or less human level language skills". That is about conversation, one capability, not general intelligence (the distinction A3 was written for), so it is not used for C2.

• C1 and C3 unchanged from 2026-10-02 (C1: 2014-07 secondhand, with the firmer Guardian 2015-06-26 quote; C3: Guardian 2015-06-26, 2015-12-27, 2016-03-05 under A16).

• Quota tags unchanged: frontier_lab (DeepMind from 2017).

Queries: Murray Shanahan Google DeepMind podcast Hannah Fry 2025 AGI transcript; Murray Shanahan Machine Learning Street Talk 2024 "AGI" transcript; "Murray Shanahan" "artificial general intelligence" OR "AGI" 2025 OR 2026 interview says; YouTube search Murray Shanahan Machine Learning Street Talk (through yt-dlp). The yt.chop.dev transcript page returned 403. A Global Player podcast lead ("Grading AI: The Hits and Misses") loaded without episode details and was not used.

Yoshua Bengio: C1 (a statement on whether or when AGI arrives, dated 2017 or earlier)

Verdict on C1: PASS (thin). Overall: PASS (thin).

• Counted: 2017-05 (reported 2017-06-02). Synced, "C2 Montreal Master Class with Yoshua Bengio: Know the Basics", by Meghan Han, 2017-06-02, reporting his master class at the C2 Montréal conference, 23 to 25 May 2017: https://syncedreview.com/2017/06/02/c2-montreal-master-class-with-yoshua-bengio-know-the-basics/ (fetched; quote checked in the raw page text).

  "We are very far from human level intelligence," said Professor Bengio. He stressed however that even if AI research were to stop today, previous discoveries would still provide benefits for another decade.

  Explicitly about human-level intelligence and dated inside the window. Secondhand (his words quoted by a technology news site), which is why the pass is thin, the same footing as Ng's secondhand C1.

• Also found: the original 2016 interview. MIT Technology Review, "Will Machines Eliminate Us?", by Will Knight, dated January 29, 2016. The old address was found in the Wayback Machine's index (technologyreview.com/s/546301/will-machines-eliminate-us/, snapshots from 2016-02-01) and read in full: https://web.archive.org/web/20160201194619/https://www.technologyreview.com/s/546301/will-machines-eliminate-us/. Primary (question-and-answer interview). Standfirst: "People who worry that we're on course to invent dangerously intelligent machines are misunderstanding the state of computer science." The introduction frames it around "the progress AI is making toward matching, or perhaps surpassing, human intelligence". His words:

  "companies want to tell a nice little story that 'Oh, we have this revolutionary technology that's going to change the world—AI is almost here, and we are the company that's going to deliver it.' That's not at all how it works."

  Asked "What approaches beyond deep learning will be needed to create a true machine intelligence?": "Traditional endeavors, including reasoning and logic—we need to marry these things with deep learning in order to move toward AI."

  "The thing I'm more worried about, in a foreseeable future, is not computers taking over the world. I'm more worried about misuse of AI."

  On its own this interview is close to explicit (it rejects "AI is almost here" in an interview about machines matching human intelligence) but never uses "human-level" or "general" in his own words. It is kept as the primary backing; the 2017 quote is the explicit one.

• Skeptic tag: still a near miss, not proposed. "Very far from" gives no timescale, so it does not meet "reject or doubt AGI within decades" as written. Recorded for me.

• Not usable: an interview on a Harvard blog with the famous "It could be ten years, it could be 100 years" line is dated 2019-06-17 (interview 2019-06-07), too late: https://archive.blogs.harvard.edu/toshietakahashi/?p=1291 (fetched). A RE•WORK video interview of 2016-08-11 has no transcript on the page.

• C2 and C3 unchanged (C2: 2025-06-03 LawZero post; C3: NYT 2016-07-17, 2017-04-09, 2017-10-22 under A16). Quota tag non_us_uk (Montréal) unchanged.

Queries: Yoshua Bengio 2016 OR 2017 "human-level" AI interview decades away; Bengio 2017 "human-level intelligence" OR "human-level AI" "far from" Montreal interview; Wayback Machine index lookup for technologyreview.com/s/5463*.

Luciano Floridi: C3 (named in three national-outlet pieces about where AI is going, dated 2017 or earlier)

Verdict on C3: PASS (thin). Overall: PASS (thin).

The three pieces, under A16 (named anywhere in the text; headline or summary about where AI is going) and A14 ruling 1 (his own piece counts for at most one of the three):

1. The Guardian, 2016-03-18, "How much should we fear the rise of artificial intelligence?", by Tom Chatfield (comment section): https://www.theguardian.com/commentisfree/2016/mar/18/artificial-intelligence-humans-computers (fetched 2026-10-06). Summary line: "From the games program AlphaGo to the movie 2001, we are often warned of the threats posed by computers. But there is a way to live alongside technology". Names him in the text: "I've borrowed this example from the philosopher of technology, Luciano Floridi, who in his book The Fourth Revolution explores…".
2. The Guardian, 2017-01-27, "AI watchdog needed to regulate automated decision-making, say experts", by Ian Sample: https://www.theguardian.com/technology/2017/jan/27/ai-artificial-intelligence-watchdog-needed-to-prevent-discriminatory-automated-decisions (fetched 2026-10-06). Written by a reporter, not by him. Quotes him: "We are already too dependent on algorithms to give up the right to question their decisions," said Floridi.
3. Financial Times, 2017-02-22, "Roman law offers a better guide to robot rights than sci-fi", by Luciano Floridi (his own piece, the one allowed): https://www.ft.com/content/99d60326-f85d-11e6-bd4e-68d53499ed71. Not opened: ft.com returns a security-check page to scripted requests. Confirmed two ways:
   • Full citation, with author, date, headline and this address, in Harris and Anthis, "The History of AI Rights Research" (arXiv 2208.04714, read from the PDF): "Floridi, L. (2017b, February 22). Roman law offers a better guide to robot rights than sci-fi. Financial Times." The paper quotes the piece's close: "[W]e can adapt rules as old as Roman law, in which the owner of enslaved persons is responsible for any damage. [...] We need less science fiction and more philosophy."
   • A link post on Robohub, 2017-02-23, titled "Roman law offers a better guide to robot rights than sci-fi | Financial Times", giving the piece's opening: "The European Parliament this month voted in favour of a resolution to create an ethical legal framework for robots." https://robohub.org/?p=72417 (fetched).

Why thin. The FT piece was not opened, and it is his own. Its headline is about legal rights for robots in reply to the European Parliament's vote, which is read here as being about where AI is going (the same reading that let the 2017 Guardian piece on an AI watchdog count). I may read either headline more narrowly; if he does, Floridi falls back to one piece short.

Quota tags unchanged: skeptic ("utterly implausible", 2016-05-09), philosopher.

Queries: Luciano Floridi Financial Times 2017 robots Roman law slaves legal personhood; "Roman law offers a better guide to robot rights than sci-fi".

Ilya Sutskever, 2020s cohort: checking C2 as recorded

My ruling, 2026-10-06, as corrected the same day: "Not a timeline statement". The co-signed 2015 "hard to predict when" is not an AGI timeline statement from before 2018, so Sutskever meets the cohort's first test. (An earlier answer said the opposite; it was given to a question the coordinating session had framed backwards. The agent had flagged that the earlier wording would fail the first test.)

C2 for the cohort is a statement dated 2024 or later. Confirmed. Dwarkesh Patel podcast, published 2025-11-25 (the page's own publication time, 2025-11-25T17:04), https://www.dwarkesh.com/p/ilya-sutskever-2, transcript checked in the raw page text:

Dwarkesh Patel 01:22:14 Speaking of forecasts, what are your forecasts to this system you're describing, which can learn as well as a human and subsequently, as a result, become superhuman? Ilya Sutskever 01:22:26 I think like 5 to 20. Dwarkesh Patel 01:22:28 5 to 20 years? Ilya Sutskever 01:22:29 Mhm.

Primary, solo, and a dated range. He meets cohort C2 as recorded.

Other gaps (quick checks)

Oren Etzioni, C2: still not found. FAIL unchanged. The GeekWire 2026 interview ("Agents, OpenAI, deepfakes and the messy reality of the AI boom") still refuses scripted requests (403 and a Cloudflare check). The search result for it quotes him only on reliability: "We haven't achieved artificial reliability. That's still a ways off." That is about how dependable today's AI agents are, not about whether or when AGI arrives, so it is not counted. A Johns Hopkins conversation of 2025-01-31 is listed as an event only. Queries: Oren Etzioni AGI 2025 "artificial general intelligence" said interview; Etzioni GeekWire 2026 "artificial reliability" AGI; "Etzioni" 2024 AGI "not around the corner" OR "far from" OR "decades" superintelligence quote.

Tyler Cowen, the three book reviews of Average Is Over (2013): one of three opened.

• The Daily Beast, 2013-09-17, "Welcome to Tyler Cowen's Future of Genius Machines", by Robert Herritt: https://www.thedailybeast.com/welcome-to-tyler-cowens-future-of-genius-machines/ (fetched). Summary line: "In the future, genius machines will guide our lives, and wages for those not in the top 10 percent will stagnate—or so predicts Tyler Cowen in his new book." The text calls the book "a digression-filled speculation about how ever-smarter machines will affect our way of life in the near future." Clearly about where AI is going and names him. Whether The Daily Beast is a national outlet is debatable (a national online magazine; not on the agent's 2026-09-30 list either way).
• The Wall Street Journal, 2013-10-02: paywall (401). Not opened.
• The Economist, 2013-09-21: security-check page (403). Not opened. Cowen already passes thin on three Guardian pieces (confirmed by me 2026-10-06); the Daily Beast piece is a fourth that is about machine intelligence rather than only jobs.

TIME100 AI lists 2023 to 2025: closed. time.com still refuses scripted requests (406 on all three list pages), but the Wayback Machine holds full copies, read on 2026-10-06:

Names were taken from the page's own markup (headings and the list of profile links), not through a summarising tool. For 49 of the 2025 entries the copy holds only the link to the profile, so the name was taken from the profile's web address (for example /7305856/mark-zuckerberg-ai/), with capitalisation and hyphens set by the agent; those rows say so in their notes. Pairs were split into two people.

Result in data/wide_list_candidates.csv:

• 126 names added (everyone on the three lists not already in the file and not a core or cohort candidate). Among them public figures who are not AI specialists (for example Scarlett Johansson, Pope Leo XIV, Marques Brownlee), because the rule is the list, not a judgment about who belongs.
• 3 rows gained a second list: Fei-Fei Li (2023 as well as 2025), Nat Friedman and Steve Huffman (2024 as well as 2025).
• Every one of the 128 names transcribed on 2026-09-30 is on the full lists (matched ignoring accents and punctuation), so no earlier row was wrong. Their notes now cite the full copy instead of "partial retrieval".
• 21 people on the lists are core or cohort candidates and were left out, as the rule says: Dario Amodei, Altman, Hassabis, Musk, Kai-Fu Lee, Ng, Huang, Suleyman, Yudkowsky, Hinton, Legg, Sutskever, LeCun, Tegmark, Bengio, Leike, Russell, Kurzweil, Mollick, Kokotajlo, Chollet.
• The file then held 311 names (310 by the rule, plus Dennett). Later the same day three of them (Liang Wenfeng, Arthur Mensch, Arvind Narayanan) became cohort members and were taken off, leaving 308; see cohort-search-2026-10-06.md.

Rendered from docs/panel-evidence-2026-10/gaps-2026-10-06.md at version a7bd17a.