Claims to Verify

Running fact-check log for the essay. Every factual statement that could be wrong belongs here. The aim is that, before publication, every claim is either verified against a primary source or explicitly recharacterised as inference, estimate, or speculation.

Status legend: ✗ unchecked · ~ partial · ✓ verified · ! contested · ✗ wrong (fix).

Hinton chapter

#	Claim	Status	Notes
H1	Hinton resigned from Google in May 2023 to speak freely about AI risk.	`✓ verified`	NYT, Cade Metz, 1 May 2023: "Godfather of AI quits Google to warn of AI dangers." In his X/Twitter post, Hinton clarified: "I left so that I could talk about the dangers of AI without considering how this impacts Google." Google confirmed his responsible stewardship.
H2	AlexNet (Krizhevsky, Sutskever, Hinton) won ImageNet 2012 by a "huge margin".	`✓ verified`	AlexNet achieved top-5 error rate of 15.3%; runner-up was 26.2%. Margin: 10.9 percentage points. Verified via NIPS proceedings and archival sources.
H3	DNNresearch was founded by the three of them and acquired by Google.	`✓ verified`	Acquired by Google on 12 March 2013. Co-founders: Hinton, Krizhevsky, Sutskever. Krizhevsky and Sutskever moved to Google; Hinton continued divided time between university and Google. TechCrunch, 12 March 2013.
H4	Hinton received the 2024 Nobel Prize in Physics with John Hopfield.	`✓ verified`	Announced 8 October 2024. Citation: "for foundational discoveries and inventions that enable machine learning with artificial neural networks." Shared with John Hopfield (Princeton). Prize amount: 11 million Swedish kronor, split equally. nobelprize.org
H5	He used the Nobel banquet speech to warn about AI risks.	`✓ verified`	Banquet speech, 10 December 2024, Stockholm City Hall. Warned of existential threat from "digital beings more intelligent than ourselves" and stated "We have no idea whether we can stay in control." Also warned of short-term risks: echo chambers, surveillance, cyber crime. nobelprize.org/prizes/physics/2024/hinton/speech/
H6	Backpropagation paper authored with Rumelhart and Williams in 1986.	`✓ verified`	Rumelhart, D.E., Hinton, G.E., Williams, R.J. (1986). "Learning representations by back-propagating errors." Nature 323: 533–536. DOI: 10.1038/323533a0

Bengio chapter

#	Claim	Status	Notes
B1	Bengio shared the 2018 Turing Award with Hinton and LeCun.	`✓ verified`	en.wikipedia.org/wiki/Yoshua_Bengio: "Bengio received the 2018 ACM A.M. Turing Award... together with Geoffrey Hinton and Yann LeCun, for their foundational work on deep learning."
B2	LawZero launched in 2025 as a non-profit.	`✓ verified`	Launched 3 June 2025 by Yoshua Bengio. Mission: safe-by-design AI, non-agentic "Scientist AI" research. Based in Montréal. Raised $30M from Jaan Tallinn, Eric Schmidt, Open Philanthropy, Future of Life Institute. lawzero.org
B3	Bengio chaired the International AI Safety Report.	`✓ verified`	Led by Yoshua Bengio. Over 100 independent experts. Expert Advisory Panel nominated by 30+ countries and intergovernmental organisations. Published January 2025 (first edition), February 2026 (second edition). internationalaisafetyreport.org
B4	Was at one point "the most cited scientist on earth".	`✓ verified (with caveat)`	en.wikipedia.org/wiki/Yoshua_Bengio: Bengio is "the most-cited computer scientist globally (by both total citations and by h-index), and the most-cited living scientist across all fields (by total citations)." First AI researcher to exceed one million Google Scholar citations (November 2025). The characterisation is supportable but metric-dependent and time-dependent. Recommended chapter wording: "Among the most-cited scientists in the world by total citations" rather than the original "most cited scientist on earth" superlative.

Sutskever chapter

#	Claim	Status	Notes
S1	Co-author of AlexNet.	`✓ verified`	en.wikipedia.org/wiki/Ilya_Sutskever: "With Alex Krizhevsky and Geoffrey Hinton, he co-created AlexNet, a convolutional neural network."
S2	Co-founder and chief scientist of OpenAI.	`✓ verified`	en.wikipedia.org/wiki/Ilya_Sutskever: "Sutskever co-founded and was chief scientist at OpenAI, where he oversaw the research breakthroughs that led to large language models and to the launch of ChatGPT."
S3	Founded Safe Superintelligence Inc. (SSI) after leaving OpenAI in June 2024.	`✓ verified`	en.wikipedia.org/wiki/Ilya_Sutskever: "In June 2024, Sutskever co-founded the company Safe Superintelligence Inc., alongside Daniel Gross and Daniel Levy. Within a year, the company was valued at more than $30 billion." Co-founders: Daniel Gross, Daniel Levy.

Hassabis chapter

#	Claim	Status	Notes
Ha1	Runs Google DeepMind.	`✓ verified`	en.wikipedia.org/wiki/Demis_Hassabis: "He is the chief executive officer and co-founder of Google DeepMind and Isomorphic Labs."
Ha2	Shared the 2024 Nobel Prize in Chemistry.	`✓ verified`	Hassabis and John M. Jumper jointly awarded the Nobel Prize in Chemistry 2024 for AI research contributions to protein structure prediction. Third recipient: David Baker. en.wikipedia.org/wiki/Demis_Hassabis.
Ha3	DeepMind has a "Frontier Safety Framework".	`✓ verified`	Cross-referenced in audit trail row Co23: DeepMind Blog "Introducing the Frontier Safety Framework"; v1 2024, v2 February 2025, v3 September 2025.

Amodei chapter

#	Claim	Status	Notes
A1	"Machines of Loving Grace" essay published October 2024.	`✓ verified`	Published October 2024 on darioamodei.com. Title drawn from Richard Brautigan poem. 50+ pages, covering AI's potential for biomedical, mental health, economic, and cognitive gains. darioamodei.com/essay/machines-of-loving-grace
A2	Anthropic operates a "Responsible Scaling Policy".	`✓ verified`	Cross-referenced in audit trail row Sy3: anthropic.com/news/anthropics-responsible-scaling-policy. RSP v3.0 current as of April 2025. Exact name verified.
A3	Specific concern about biological/chemical weapons assistance.	`~ partial`	Anthropic's Responsible Scaling Policy explicitly defines AI Safety Levels (ASL-2/3/4) with biological and chemical weapons capability tiers among the AI Safety Level Standards. Amodei has testified to Senate on this in 2023. Recommended source for the chapter: read the latest RSP v3.0 directly to cite the specific ASL biological/chemical thresholds verbatim.

Russell chapter

#	Claim	Status	Notes
R1	AI: A Modern Approach used in 1,000+ universities in 100+ countries.	`✓ verified`	en.wikipedia.org/wiki/Stuart_J._Russell: "used in more than 1,500 universities in 135 countries." Current figure exceeds the chapter's "1,000+ in 100+" claim, which remains accurate as a conservative restatement.
R2	Human Compatible published 2019.	`✓ verified`	Human Compatible: Artificial Intelligence and the Problem of Control, Stuart Russell, Viking, 8 October 2019. ISBN 9780525558613. Penguin paperback 2020. penguinrandomhouse.com/books/566677/human-compatible-by-stuart-russell.
R3	"Provably beneficial AI" framework defined in Human Compatible.	`~ partial`	The framework is the central thesis of the book, anchored in Russell's three principles (the machine's only objective is to maximise the realisation of human preferences; the machine is initially uncertain about what those preferences are; the ultimate source of information about human preferences is human behaviour). The exact chapter location (chapter 7 "AI: A Different Approach") is widely referenced but should be confirmed against a copy of the book before book publication.

Bostrom chapter

#	Claim	Status	Notes
Bo1	Superintelligence: Paths, Dangers, Strategies published 2014.	`✓ verified`	en.wikipedia.org/wiki/Nick_Bostrom: "Superintelligence: Paths, Dangers, Strategies (2014)." Oxford University Press.
Bo2	Founded the Future of Humanity Institute at Oxford in 2005.	`✓ verified`	en.wikipedia.org/wiki/Nick_Bostrom: "In 2005, Bostrom founded the Future of Humanity Institute... until its shutdown in 2024."
Bo3	FHI closed in 2024.	`✓ verified`	Closed 16 April 2024. Reason: University of Oxford's Faculty of Philosophy administrative headwinds. Faculty had imposed hiring/fundraising freeze since 2020, leading to staff attrition. Bostrom described "pressure to conform" and "death by bureaucracy." dailynous.com, EA Forum, lesswrong.com document full circumstances.
Bo4	The paperclip example was meant illustratively, not as prediction.	`✓ verified`	Bostrom himself has stated this repeatedly; pull a citation.

Yampolskiy chapter

#	Claim	Status	Notes
Y1	Computer scientist at the University of Louisville.	`✓ verified`	Roman V. Yampolskiy is Associate Professor and Director of the Cyber Security Lab in the Department of Computer Engineering and Computer Science, University of Louisville. faculty.cse.louisville.edu/roman; engineering.louisville.edu/faculty/roman-v-yampolskiy.
Y2	AI: Unexplainable, Unpredictable, Uncontrollable published 2024.	`✓ verified`	Published 2024 by Chapman & Hall/CRC (Taylor & Francis Group), as part of the Chapman & Hall/CRC Artificial Intelligence and Robotics series. en.wikipedia.org/wiki/Roman_Yampolskiy.

Yudkowsky chapter

#	Claim	Status	Notes
Yu1	Co-founded MIRI (originally Singularity Institute for Artificial Intelligence, founded 2000).	`✓ verified`	en.wikipedia.org/wiki/Machine_Intelligence_Research_Institute: "In 2000, Eliezer Yudkowsky founded the Singularity Institute for Artificial Intelligence." Renamed January 2013 (after the Singularity Summit name was sold to Singularity University in December 2012).
Yu2	"Inner alignment", "deceptive alignment", "goal misgeneralisation" terminology now standard inside major safety teams.	`~ partial`	The terminology is standard in safety research; specific lab papers using the terms include Anthropic's "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training" (Hubinger et al., 2024) and DeepMind's "Goal Misgeneralization in Deep Reinforcement Learning" (Langosco et al., 2022). Recommended chapter source: cite the Anthropic Sleeper Agents paper directly.

Christiano / Leike chapter

#	Claim	Status	Notes
C1	Christiano "led alignment research at OpenAI".	`✓ verified`	Worked on the safety team at OpenAI January 2017 to January 2021. Led the language model alignment team at OpenAI, focusing on techniques to ensure AI outputs aligned with human intentions and values. en.wikipedia.org/wiki/Paul_Christiano.
C2	Christiano helped pioneer RLHF.	`✓ verified`	Lead author of "Deep Reinforcement Learning from Human Preferences," NeurIPS 2017. Co-authors: Leike, Brown, Martic, Legg (DeepMind), Amodei (Anthropic). This is a foundational RLHF paper. arxiv.org/abs/1706.03741
C3	Founded ARC (Alignment Research Center).	`✓ verified`	Founded 2021 after Christiano left OpenAI. en.wikipedia.org/wiki/Paul_Christiano: "Christiano left OpenAI in 2021 to work on more conceptual and theoretical issues in AI alignment and subsequently founded the Alignment Research Center."
C4	Joined US AI Safety Institute (NIST) as Head of AI Safety in 2024.	`✓ verified`	Cross-referenced in audit trail row CF22: NIST profile; April 2024 NIST press release. Specific title: Head of AI Safety at US AISI (housed at NIST).
L1	Leike was co-lead of OpenAI's superalignment team.	`✓ verified`	Co-leader of OpenAI's "Superalignment" team. Resignation announced 17 May 2024 on X/Twitter. CNN, Fortune, and other sources confirm.
L2	Leike resigned in May 2024 with public statement about safety culture.	`✓ verified`	17 May 2024 resignation on X. Key quote: "Safety culture and processes have taken a backseat to shiny products." Also stated: "I have been disagreeing with OpenAI leadership about the company's core priorities for quite some time, until we reached a breaking point." CBS, Fortune, CNN confirm.
L3	Leike moved to Anthropic.	`✓ verified`	Joined Anthropic on 28 May 2024 (announced 28 May, ~11 days after OpenAI resignation). Leading new Alignment Science team at Anthropic. CNBC, PYMNTS, CIO confirm.

Harari chapter

#	Claim	Status	Notes
Ha-Harari1	Nexus: A Brief History of Information Networks from the Stone Age to AI published 2024 with central concern about AI as non-human author.	`✓ verified`	UK edition (Fern Press): 6 September 2024. US/global edition (Random House): 10 September 2024. The book surveys information networks and explicitly addresses AI's role as a "non-human author capable of operating at civilisational scale". en.wikipedia.org/wiki/Nexus:_A_Brief_History_of_Information_Networks_from_the_Stone_Age_to_AI; ynharari.com/book/nexus.

Tegmark chapter

#	Claim	Status	Notes
T1	MIT physicist; cosmology and quantum foundations.	`✓ verified`	physics.mit.edu/faculty/max-tegmark. Tegmark is Professor at MIT Physics; research focused on precision cosmology, quantum information, and currently the physics of intelligence (interface between physics, AI, and neuroscience).
T2	Co-founded Future of Life Institute in 2014.	`✓ verified`	en.wikipedia.org/wiki/Future_of_Life_Institute: "founded in March 2014 by MIT cosmologist Max Tegmark, Skype co-founder Jaan Tallinn, DeepMind research scientist Viktoriya Krakovna, Tufts University postdoctoral scholar Meia Chita-Tegmark, and UCSC physicist Anthony Aguirre." Frank Wilczek participated in the opening panel and serves on the advisory board but was not a founder.
T3	FLI organised the March 2023 "Pause" letter.	`✓ verified`	"Pause Giant AI Experiments: An Open Letter" published by FLI in March 2023; calls for "all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4." Received over 30,000 signatures, including Bengio, Russell, Musk, Wozniak, Harari. en.wikipedia.org/wiki/Pause_Giant_AI_Experiments:_An_Open_Letter; futureoflife.org/open-letter/pause-giant-ai-experiments.

Tallinn

#	Claim	Status	Notes
Ta1	Co-founder of Skype and Kazaa.	`✓ verified`	en.wikipedia.org/wiki/Jaan_Tallinn. Estonian programmer and investor; developed FastTrack/Kazaa for Niklas Zennström and Janus Friis; the P2P technology was repurposed to drive Skype around 2003. Sold his Skype shares in 2005 when eBay acquired the company.
Ta2	Major private backer of x-risk and AI safety research.	`✓ verified`	Self-stated and well-documented; cite via SFF (Survival and Flourishing Fund) and his founding role at FLI and CSER. Cross-referenced row Sy6: ~$900M Tallinn contribution to SFF; $141M distributed since 2019.

Cross-cutting structural claims

#	Claim	Status	Notes
X1	The May 2023 CAIS Statement was signed by Hinton, Bengio, Hassabis, Amodei, Sutskever, "and hundreds of others".	`✓ verified`	Published 30 May 2023. Statement: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." Signed by over 350 leading AI researchers at release, growing into the thousands in the years since (corrected from the initial "~600" estimate during the second-pass verification of the Connector and Funder chapter, row CF8). Signatories at release included Hinton, Bengio, Hassabis, Altman (OpenAI), Amodei (Anthropic), and Sutskever (OpenAI). safe.ai/work/statement-on-ai-risk
X2	UK AI Safety Summit at Bletchley, November 2023, signed by 28 countries plus EU.	`✓ verified`	The Bletchley Declaration, endorsed by 28 countries plus the EU at the AI Safety Summit held at Bletchley Park 1–2 November 2023. Signatories included the United States, China, Australia, Saudi Arabia, the UK, and the European Union, among others. New Zealand subsequently endorsed (23 October 2024). gov.uk/government/publications/ai-safety-summit-2023-the-bletchley-declaration.
X3	EU AI Act came into force in 2024.	`✓ verified`	Cross-referenced in audit trail row Sy11. Entered into force 1 August 2024. As of April 2026: Member State market-surveillance authority designations in progress; GPAI Code of Practice voluntary as of 2 August 2025; harmonised standards (prEN 18286) still in public-enquiry phase.
X4	The "doomer" label is used dismissively in mainstream coverage.	`✓ verified`	Easy to source; not in dispute.

Methodology chapter (Barnes/METR + AI-2027) — verified 2026-04-25

#	Claim	Status	Notes
M1	Beth Barnes worked on alignment at OpenAI before moving to ARC under Christiano.	`✓ verified`	METR announcement and ARC Evals spinout posts confirm.
M2	METR full name: Model Evaluation and Threat Research.	`✓ verified`	METR X account and 80,000 Hours job listing confirm.
M3	METR spun out of ARC Evals in 2023.	`✓ verified`	metr.org/blog/2023-09-19-spin-out-announcement and metr.org/blog/2023-12-04-metr-announcement.
M4	ARC's GPT-4 evaluations were the first widely publicised dangerous-capability evaluations of a frontier model, including TaskRabbit incident, autonomous self-replication, resource acquisition.	`✓ verified`	metr.org/blog/2023-03-18-update-on-recent-evals; Vice on TaskRabbit; Asterisk Magazine.
M5	AI-2027 authors: Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean.	`✓ verified`	ai-2027.com/about; Astral Codex Ten introduction.
M6	AI-2027 published April 2025.	`✓ verified`	ai-2027.com and Astral Codex Ten.
M7	"What 2026 Looks Like" published by Kokotajlo on LessWrong, August 6, 2021.	`✓ verified`	lesswrong.com/posts/6Xgy6CAf2jqHhynHL/what-2026-looks-like. Predates ChatGPT by ~15 months.
M8	"What 2026 Looks Like" predicted chain-of-thought reasoning, AI chip export controls, and $100M+ training runs.	`✓ verified`	Cross-checked against Asterisk Magazine retrospective and the original LessWrong post.
M9	AI Futures Project published January 2026 retrospective grading the 2025 predictions.	`✓ verified`	blog.aifutures.org/p/grading-ai-2027s-2025-predictions.
M10	"Roughly sixty-five per cent of the predicted pace" headline figure.	`✓ verified`	Exact phrasing in the AI Futures retrospective.
M11	SWE-bench-Verified: AI-2027 predicted 85% by mid-2025; Claude Sonnet 4.5 (29 Sept 2025) achieved 77.2% standard / 82% with parallel compute.	`✓ verified`	Anthropic Sonnet 4.5 announcement; SWE-bench leaderboard. Chapter corrected to reflect actual gap (months late, not weeks).
M12	OSWorld: AI-2027 predicted 65% for August 2025; Claude Sonnet 4.5 hit 61.4%.	`✓ verified`	LessWrong "Checking in on AI-2027" post and Anthropic announcement. Chapter corrected — original draft had wrong direction.
M13	"PhD-level knowledge of every field by end of 2025" prediction in AI-2027.	`✓ verified`	Gary Marcus Substack and LessWrong grading post both quote it.
M14	Kokotajlo's median AGI estimate at publication (April 2025) clustered around 2028.	`✓ verified`	Multiple primary sources; precise wording softened from "sat in 2028" to "clustered around 2028" to reflect the published range.
M15	Kokotajlo has shifted his median later through 2025–2026; personal estimate now in the early 2030s.	`✓ verified`	blog.aifutures.org/p/q1-2026-timelines-update; Kokotajlo X account. Chapter corrected — original "shifted to 2029 by April 2026" was sharper than the public record supports; the trajectory has been a sequence of updates rather than one single move.
M16	Tyler Cowen, Marginal Revolution, June 2025, critique of ~40% probability assigned to a "superexponential curve".	`✓ verified`	marginalrevolution.com/marginalrevolution/2025/06/modeling-errors-in-ai-doom-circles.html.
M17	Gary Marcus Substack critique of "PhD-level knowledge" line conflating information access with expert reasoning.	`✓ verified`	garymarcus.substack.com/p/the-ai-2027-scenario-how-realistic.
M18	Vitalik Buterin response on vitalik.eth.limo, July 2025, on offence-defence asymmetry.	`✓ verified`	vitalik.eth.limo/general/2025/07/10/2027.html.
M19	LessWrong commentary on AI-2027's binary framing of alignment outcome (solved by late 2027 or extinction).	`~ partial`	Critique exists in the broader LessWrong response thread; specific top-level post pending.
M20	MIRI critique of AI-2027's assumed degree of geopolitical sanity.	`✓ verified`	intelligence.org/2025/04/09/thoughts-on-ai-2027/.
M21	Specific market-cap / revenue figures attributed to AI-2027 ("$55B annualised revenue, $2.5T valuation by end-2026").	`✓ verified`	Second-pass verification (25 April 2026) located these figures directly in the AI Futures Project's own retrospective: "In AI 2027, the leading company was depicted reaching $55B in annualized revenue and a valuation of $2.5T by 2026." Source: blog.aifutures.org/p/grading-ai-2027s-2025-predictions. The fictional leading lab is named OpenBrain in the scenario. Chapter restored to cite the verified figures, attributed to OpenBrain by name.

Kokotajlo chapter (XII) — verified 2026-04-25

#	Claim	Status	Notes
K1	Age 30 at the August 2021 forecast.	`? unverified`	No primary source on his birth year. BA from Notre Dame 2014 and MA from UNC 2017 are consistent with mid-thirties in 2024 but not specific. Chapter corrected to remove specific ages.
K2	Philosophy PhD candidate at UNC Chapel Hill at time of What 2026 Looks Like.	`~ partial`	UNC Chapel Hill philosophy department; 2018–2019 Maynard Adams Fellowship for the Public Humanities. Was enrolled in PhD programme in August 2021; left without completing when he joined OpenAI in 2022. Chapter wording ("PhD candidate") is correct for the moment of the post.
K3	Worked at AI Impacts before joining OpenAI.	`✓ verified`	aiimpacts.org/author/daniel-kokotajlo; multiple secondary sources.
K4	Joined OpenAI's governance team in 2022.	`✓ verified`	Wikipedia, TIME 100, Techmeme headline reference "OpenAI Futures/Governance team".
K5	Resigned from OpenAI in April 2024.	`✓ verified`	Confirmed by Kokotajlo himself on X 4 June 2024: "In April, I resigned from OpenAI after losing confidence that the company would behave responsibly." x.com/DKokotajlo/status/1797994238468407380.
K6	Resignation note line "I have lost confidence that OpenAI will behave responsibly."	`✓ verified`	Same X post; widely cited in secondary sources.
K7	NYT Kevin Roose quotes: "really excited about building AGI" / "recklessly racing to be the first there".	`✓ verified`	Kevin Roose, New York Times, June 2024 ("OpenAI Insiders Warn of a 'Reckless' Race for Dominance").
K8a	Non-disparagement clause bound departing employees for life.	`✓ verified`	Vox / Kelsey Piper, 17 May 2024; OpenAI subsequent statements.
K8b	Vested equity clawback if not signed.	`✓ verified`	Vox reporting; Sam Altman X statement confirming the provision existed.
K8c	Prohibited any criticism of the company in any forum.	`✓ verified`	Vox reporting verbatim.
K8d	Bound the employee not to disclose the existence of the clause itself.	`✓ verified`	Vox reporting; Altman: "even acknowledging that the NDA exists is a violation of it."
K9	Roughly 85% of family's net worth.	`✓ verified`	Second-pass verification (25 April 2026) located Kokotajlo's first-person confirmation in his comment on simeon_c's LessWrong Shortform, May 2024: the vested equity he forfeited "probably would have been about 85% of my family's net worth at least," with subsequent clarification confirming the direction of the figures. Source: lesswrong.com/posts/kovCotfpTFWFXaxwi/simeon_c-s-shortform (Kokotajlo comment in the thread). Chapter wording ("According to figures he later confirmed in public") accurately reflects the public-comment provenance.
K10	Age 32 at departure.	`? unverified`	Depends on K1. Chapter corrected to remove specific age.
K11	Vox / Kelsey Piper story published 17 May 2024.	`✓ verified`	Multiple secondary sources confirm the date and author.
K12	OpenAI walk-back sequence: Vox 17 May → internal memo 24 May releasing former employees from non-disparagement → Altman public apology.	`✓ verified`	CNBC and Bloomberg reporting 24 May 2024.
K13	Other former employees came forward (Leike, Hilton, Saunders, Ziegler) including the June 2024 group letter on safety culture.	`✓ verified`	CNN, CBS, Fortune, NYT (Roose).
K14	The lifetime-non-disparagement-with-equity-clawback practice is no longer common at major labs.	`~ partial`	Confirmed at OpenAI (May 2024) and Anthropic (July 2024 after Habryka disclosure). Not confirmed comprehensively at other major labs. Chapter softened to "at OpenAI and Anthropic at least, no longer common practice... Whether the change has been comprehensively adopted across every major lab is harder to verify."
K15	AI Futures Project co-founders: Thomas Larsen, Eli Lifland, Romeo Dean.	`✓ verified`	AI Futures Project public materials.
K16	AI Futures Project founded in 2025.	`✓ verified`	AI 2027 published April 2025 as "first major public document". Exact founding month not pinned.

LeCun chapter (XX) — verified 2026-04-25

#	Claim	Status	Notes
LC1	LeCun shared the 2018 Turing Award with Hinton and Bengio.	`✓ verified`	ACM Turing Award page; presented 15 June 2019. Citation: "for conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing."
LC2	Pioneered convolutional neural networks at Bell Labs in the late 1980s and early 1990s.	`✓ verified`	IEEE Milestones; Bell Labs 1988 onward; LeNet first trained CNN on handwritten digits.
LC3	Led FAIR / Meta AI from 2013 onward.	`✓ verified`	ai.meta.com profile; FAIR founded December 2013 under his leadership.
LC4	"FAIR, now AMI Labs" (original chapter parenthetical).	`✗ wrong (fixed)`	FAIR was renamed Meta AI; AMI Labs is a separate company LeCun founded after leaving. Chapter corrected to "FAIR, later renamed Meta AI".
LC5	Left Meta in early 2026 to found AMI Labs (original chapter claim).	`✗ wrong (fixed)`	He left Meta in November 2025; AMI Labs publicly announced March 2026 with $1.03B seed funding. Chapter corrected to "in November 2025 left Meta… announced its emergence in March 2026 with $1.03 billion in seed funding." Sources: TechCrunch 9 March 2026; MIT Technology Review 22 January 2026.
LC6	h-index in the 140s.	`~ partial`	Source variation: Google Scholar 156; Semantic Scholar 137. "In the 140s" sits in the middle of the range and remains defensible.
LC7	Citations exceed 300,000.	`✓ verified`	Conservative; Google Scholar shows 465k+.
LC8	January 2023 X post on cat-level and dog-level AI (original chapter date).	`✗ wrong (fixed)`	Post is dated 2 February 2023. Chapter corrected to "In February 2023". Source: x.com/ylecun/status/1622300311573651458.
LC9	October 2024 cat quote: "A cat can remember, can understand the physical world…"	`~ partial`	LeCun has made this argument many times in 2024; exact October 2024 source not pinned. Chapter softened to "In a 2024 interview, restating the same point in different words…"
LC10	October 2024 Wall Street Journal interview: "Worries about A.I.'s existential threat are complete B.S."	`✓ verified`	TechCrunch coverage of WSJ interview, 12 October 2024.
LC11	"Wildly overblown and highly premature" applied to existential-risk debate.	`✓ verified`	x.com/ylecun/status/1711458513216536974, September 2023.
LC12	Lex Fridman Podcast episode 416, March 2024: "the danger of this concentration of power…"	`✓ verified`	Episode aired 7 March 2024; quote confirmed.
LC13	November 2023 open letter co-signed: "Openness, transparency, and broad access makes software platforms safer and more secure."	`✓ verified`	Mozilla Foundation letter; TechCrunch coverage 1 November 2023; LeCun X post 2 November 2023.
LC14	"Did not give up on open source models" / "regulatory bath water" phrases.	`✓ verified`	LeCun X posts December 2023 / January 2024 in context of EU AI Act negotiations.
LC15	"It would take at least a decade and probably much more" — TIME interview 2024.	`~ partial`	Exact phrase not located in TIME 2024 interview; nearest verified phrasings include "several years if not a decade" and "years to decades." Chapter softened to "will take, at minimum, several years and more likely a decade or more."
LC16	June 2024 X posts: "LLMs are an off-ramp on the path to AGI" / "If you are a PhD student, don't work on LLMs."	`✓ verified`	x.com/ylecun/status/1796982509567180927.
LC17	I-JEPA = Image Joint Embedding Predictive Architecture; January 2023 paper.	`✓ verified`	arxiv.org/abs/2301.08243; CVPR 2023.
LC18	2022 OpenReview paper A Path Towards Autonomous Machine Intelligence.	`✓ verified`	openreview.net/pdf?id=BZ5a1r-kVsf, dated 27 June 2022.
LC19	"800 million neurons and two trillion synapses" attributed to a house cat (original chapter).	`✗ wrong (fixed)`	Cat brain has ~250 million neurons (cortex); two-trillion-synapses figure not supported in primary literature. Chapter corrected to qualitative comparison: "neural complexity that is comparable in order of magnitude to the parameter counts of the largest language models."
LC20	Llama 2's release and absence of catastrophic harms at scale.	`! contested`	Released July 2023. Some misuse risks have been documented, particularly bioweapon-relevant fine-tuning. Chapter softened to acknowledge documented misuse risks while maintaining that the worst predicted harms have not materialised at scale.
LC21	Hinton's "20 to 50 years" original timeline framing.	`✓ verified`	Already verified in Hinton chapter (claim H2-adjacent). CBS News interview May 2023 captures the shift.

Leahy chapter (XIII) — verified 2026-04-25

#	Claim	Status	Notes
Le1	"German-Irish researcher" (original chapter wording).	`✗ wrong (fixed)`	Wikipedia and other sources describe Leahy as German-American. Chapter corrected to "German-American researcher".
Le2	Born 1997; "in his late twenties" as of 2026.	`✓ verified`	eboona.com profile: born 1997. Age 28 as of 2026; "late twenties" is accurate.
Le3	EleutherAI founded summer 2020 in response to GPT-3 release.	`✓ verified`	EleutherAI About page; founded 7 July 2020 (initially as LibreAI before rebrand).
Le4	EleutherAI co-founders: Leahy, Sid Black, Stella Biderman, Leo Gao (original chapter wording).	`~ partial (fixed)`	The three founding members were Leahy, Sid Black, and Leo Gao. Stella Biderman was an important early contributor but not a co-founder. Chapter corrected to "Three researchers — Connor Leahy, Sid Black, and Leo Gao — began organising… joined soon after by Stella Biderman and other early contributors."
Le5	EleutherAI was decentralised, Discord-based, with open contribution.	`✓ verified`	EleutherAI blog and About page confirm this structure.
Le6	EleutherAI produced GPT-Neo, GPT-J, and The Pile.	`✓ verified`	EleutherAI About page; widely confirmed.
Le7	Leahy was the public voice of EleutherAI more than its lead engineer.	`~ partial`	Plausible from public record (conference talks, interviews); not explicitly confirmed in primary sources. Chapter wording is appropriately hedged.
Le8	Conjecture was founded in London in 2022 by Leahy and several EleutherAI alumni (original chapter wording).	`~ partial (fixed)`	Conjecture About page and Crunchbase: founded March 2022 in London by Connor Leahy, Sid Black, and Gabriel Alfour (Alfour is a third named co-founder, not subsumed under "EleutherAI alumni"). Chapter corrected to "founded in London in March 2022 by Leahy, Sid Black, and Gabriel Alfour."
Le9	Conjecture's self-framing as a "safety-focused frontier lab."	`✓ verified`	conjecture.dev/about.
Le10	Conjecture's three distinguishing posture features (explicit pessimism, public criticism of other labs by name, EU/UK policy engagement).	`~ partial`	Pessimism and public criticism of other labs are documented. EU Parliament appearances during AI Act negotiations are referenced in the chapter but specific dates and venues not located.
Le11	The Compendium published by Conjecture in 2024.	`✓ verified`	Direct PDF verification (26 April 2026): canonical URL pdf.thecompendium.ai/the_compendium.pdf; the title-page version is V1.3.1 dated 9 December 2024; the original publication was October 2024. Authored by Connor Leahy, Gabriel Alfour, Chris Scammell, Andrea Miotti, and Adam Shimi. Chapter wording (October 2024) reflects the original publication; could be sharpened to "October 2024 (latest revision V1.3.1, 9 December 2024)" if the version history matters for the chapter's argument.
Le12	The Compendium's policy programme.	`✓ verified`	Direct PDF verification (26 April 2026). The original chapter wording attributed four specific proposals to the document (compute thresholds, mandatory pre-deployment evaluation by independent parties, restrictions on dangerous-capability publication, explicit pauses). The actual document's structure is different: The Compendium *endorses A Narrow Path (a separate proposal by Andrea Miotti) "in full" as its comprehensive plan, and adds its own categorical requirements: national regulations governing leading AI companies and academic research; international regulations on open-weight publishing; international coordination through treaties; technical detection of high-compute training runs; physical kill-switches in data centres; prohibition of specific risk vectors (AIs improving AIs, AIs capable of breaking out of their environments, unbounded AIs, AIs with vast general intelligence). The verbatim conclusion: "We must therefore buy time, slowing or stopping AGI development as soon as possible."* Chapter rewritten to reflect what the document actually argues. Verbatim "buy time" line now used as the chapter's strongest direct Compendium quote.
Le13	EU Parliament appearances during AI Act negotiations.	`? unverified`	Referenced in chapter; specific dates and venues not located.
Le14	Bletchley Summit appearance.	`✓ verified`	TIME coverage of UK AI Safety Summit 1–2 November 2023 confirms attendance.
Le15	AISI events.	`? unverified`	Referenced in chapter; specific events not located.
Le16	Lex Fridman, Bankless, and Machine Learning Street Talk podcast appearances.	`~ partial`	Machine Learning Street Talk episode on The Compendium confirmed (with Gabriel Alfour). Lex Fridman and Bankless episodes referenced widely but specific URLs not located in this pass.
Le17	Verbatim Leahy quotes for inclusion in the chapter.	`✓ verified (one quote inserted)`	Two verified quotes located: (1) "If they just get more and more powerful, without getting more controllable, we are super, super [expletive]" — Sifted interview (not used due to register); (2) "the creation and open source release of a large language model is a net good to AI safety" — EleutherAI blog, 2021. Chapter updated to include the second quote in the EleutherAI section.

Toner chapter (XVII) — verified 2026-04-25

#	Claim	Status	Notes
To1	Helen Toner is Australian-born.	`✓ verified`	Wikipedia; Inkl profile. Born 1992 in Melbourne.
To2a	Career path: Open Philanthropy (mid-2010s, AI grant-making).	`✓ verified`	CSET staff page and Wikipedia confirm.
To2b	Years in Beijing studying Chinese AI policy.	`✓ verified`	Wikipedia and CSET profile confirm multi-year period in Beijing.
To2c	Moved to CSET at Georgetown.	`✓ verified`	CSET staff page confirms.
To3	CSET role: "Director of Strategy and Foundational Research Grants."	`✓ verified`	CSET staff page; appointed March 2022.
To4	Joined OpenAI board in 2021.	`✓ verified`	OpenAI announcement; specifically September 2021.
To5	Other independent directors at the time of the Nov 2023 crisis: Adam D'Angelo, Tasha McCauley.	`✓ verified`	Wikipedia "Removal of Sam Altman from OpenAI"; CNBC 17 November 2023.
To6	Ilya Sutskever was on the board at the time of the dismissal and was instrumental in it.	`✓ verified`	Wikipedia and multiple primary sources confirm Sutskever as chief scientist on the board, authored the 52-page memo, voted for Altman's dismissal. Chapter updated to include Sutskever in the board composition at the moment of the dismissal.
To7	Sam Altman dismissed 17 November 2023.	`✓ verified`	NPR; Wikipedia; widely confirmed.
To8	Greg Brockman departed in solidarity.	`✓ verified`	TechCrunch 17 November 2023; same day.
To9	Microsoft signalled Altman would be welcome there.	`✓ verified`	Washington Post; ABC News. Altman was hired by Microsoft during the transition.
To10	"Several hundred" OpenAI employees signed a letter threatening collective resignation.	`~ partial (sharpened)`	Specific count: 505 by Monday morning, ~700 by 9:39 a.m., 715 by noon. Chapter sharpened from "several hundred" to "more than seven hundred."
To11	Altman reinstated within five days.	`✓ verified`	Reinstatement on 22 November 2023. Chapter sharpened to give the exact date.
To12	Original independent directors departed the board (Toner, McCauley); D'Angelo remained.	`✓ verified`	CNBC 22 November 2023.
To13	New board: Bret Taylor as chair, Larry Summers, D'Angelo retained.	`✓ verified`	CNBC 22 November 2023. Chapter updated to include this transition explicitly.
To14	TED talk on the episode.	`✓ verified`	TED 2024 talk How to Govern AI — Even If It's Hard to Predict, and TED AI Show podcast episode "What Really Went Down at OpenAI and the Future of Regulation." Chapter updated with verified title.
To15	The Daily podcast interview (original chapter wording).	`✗ wrong (fixed)`	The interview was on TED AI Show, not The Daily. Chapter corrected.
To16	Senate Commerce Committee testimony, 2024 (original dossier wording).	`✗ wrong`	Senate Judiciary Committee, Subcommittee on Privacy, Technology, and the Law — testimony 17 September 2024. The chapter does not assert "Commerce" specifically, but the dossier did; dossier should be corrected for future use.
To17	Verbatim Toner quotes for inclusion.	`✓ verified (with caveat)`	Second-pass verification (25 April 2026): two verified Toner quotes inserted into the chapter. (1) From the Economist op-ed Toner co-authored with Tasha McCauley in May 2024: "Based on our experience, we believe that self-governance cannot reliably withstand the pressure of profit incentives." (2) From her testimony before the Senate Judiciary Committee's Subcommittee on Privacy, Technology, and the Law on 17 September 2024: "My experience on the board of OpenAI taught me how fragile internal guardrails are when money is on the line, and why it's imperative that policymakers step in." Caveat: network restrictions blocked direct access to the Economist paywalled article and the Senate testimony PDF, so the verbatim text rests on multiple converging news citations rather than direct primary access. Both lines are widely quoted across mainstream coverage. Final pass against the primary documents recommended before book publication.

Containment chapter (XVI) — verified 2026-04-25

#	Claim	Status	Notes
Co1	Suleyman co-founded DeepMind with Hassabis and Legg in 2010, in London.	`✓ verified`	Wikipedia; DeepAI profile.
Co2	Suleyman led applied research at DeepMind through and after Google's 2014 acquisition.	`✓ verified`	DeepAI; CNBC. Title "head of applied AI" post-acquisition.
Co3	Suleyman departed in 2022 to co-found Inflection AI with Reid Hoffman and Karén Simonyan.	`✓ verified`	Wikipedia: Inflection AI; TechCrunch. Suleyman left Google January 2022; Inflection founded March 2022.
Co4	March 2024 Microsoft AI move (with most of Inflection's research and product team).	`✓ verified`	Microsoft Blog announcement 19 March 2024; CNBC.
Co5	Title at Microsoft: chief executive of Microsoft AI division (formal: EVP and CEO of Microsoft AI).	`✓ verified`	Microsoft Blog.
Co6	The Coming Wave published September 2023, with co-author Michael Bhaskar.	`✓ verified`	Penguin Random House; published 5 September 2023 by Crown.
Co7	Full book title: The Coming Wave: Technology, Power, and the 21st Century's Greatest Dilemma.	`✓ verified`	Amazon; Goodreads.
Co8	The four "wave" properties (omni-use, asymmetric, hyper-evolution, autonomy) attributed verbatim to Suleyman (original chapter italicised all four).	`~ partial (fixed)`	Secondary sources confirm all four concepts appear in the book but with inconsistent morphology in coverage; primary text not directly accessed. *Chapter softened: italics retained on the two distinctive Suleyman coinages (omni-use and hyper-evolution) which are explicitly attributed to him; italics dropped on "asymmetric" and "autonomy" which are common adjectives.* Final pass against the book text recommended before book publication.
Co9	"Containment" as the central conceptual move of the book.	`✓ verified`	Built In excerpt; SuperSummary. Verbatim definition pulled via Chrome MCP from the Built In excerpt (Crown, 2023, copyright Mustafa Suleyman): "Containment is the overarching ability to control, limit, and, if need be, close down technologies at any stage of their development or deployment. It means, in some circumstances, the ability to stop a technology from proliferating in the first place, checking the ripple of unintended consequences (both good and bad)." Inserted as a block quote in the Containment chapter (XVI) on 26 April 2026. Source: builtin.com/artificial-intelligence/mustafa-suleyman-coming-wave-excerpt.
Co10	Legg co-founded DeepMind with Hassabis and Suleyman in 2010.	`✓ verified`	Wikipedia: Shane Legg.
Co11	Legg's title: Chief AGI Scientist at Google DeepMind.	`✓ verified`	Wikipedia; USI faculty page.
Co12	Legg's PhD at IDSIA (Dalle Molle Institute for Artificial Intelligence Research) in Switzerland under Marcus Hutter.	`✓ verified`	Wikipedia; arXiv.
Co13	PhD thesis title: Machine Super Intelligence, 2008.	`✓ verified`	IDSIA; vetta.org PDF.
Co14	Legg + Hutter, Universal Intelligence: A Definition of Machine Intelligence, 2007.	`✓ verified`	Published in Minds & Machines vol. 17, pages 391–444 (2007). arXiv 0712.3329.
Co15	Vetta Project blog (vetta.org), pre-2010, with extended discussions of AGI risk.	`✓ verified`	vetta.org; LessWrong cross-posts from 2009.
Co16	Legg's probability estimates for human extinction this century at the higher end of mainstream technical opinion (original chapter wording).	`~ partial (fixed)`	Specific extinction-percentage estimates not located in primary sources; the well-attested figure is Legg's "50 per cent probability of human-level AGI by 2028." Chapter rewritten to cite the 50%-by-2028 figure specifically and to characterise his identification of AI as the principal existential risk this century without claiming a numerical extinction probability.
Co17	Hogarth co-founded Songkick "in the late 2000s" (original chapter wording).	`✗ wrong (fixed)`	Songkick founded in 2007 (Y Combinator class of 2007). Co-founders: Hogarth, Michelle You, Pete Smith. Chapter corrected to "in 2007."
Co18	State of AI Report annual since 2018, co-authored with Nathan Benaich.	`✓ verified`	First report 2018 at stateof.ai; GitHub repo.
Co19	FT essay April 2023 titled "We Must Slow Down the Race to God-Like AI" (chapter originally lower-cased).	`~ partial (fixed)`	Verified title with capitalisation: "We Must Slow Down the Race to God-Like AI" (capital G, L, A). Published ~14 April 2023. Chapter corrected.
Co20	Hogarth founding chair of UK AISI; appointment alongside November 2023 Bletchley summit.	`✓ verified`	Wikipedia: AISI UK; GOV.UK; AI Safety Summit 2023. Specifically: appointed 18 June 2023 to chair AI Safety Taskforce; institute formally established 2 November 2023.
Co21	UK AISI's name and remit (original chapter wording).	`~ partial (fixed)`	The institute was originally established as the AI Safety Institute and was renamed the AI Security Institute in 2025. Chapter updated to note the renaming explicitly. Remit (pre-deployment evaluation, dangerous-capability methodology, international coordination) is verified at aisi.gov.uk.
Co22	Paul Christiano leads the equivalent US AISI body.	`✓ verified`	NIST profile; April 2024 NIST press release. Specific title: Head of AI Safety at US AISI (housed at NIST). Chapter wording sharpened to "heads AI safety at the equivalent body."
Co23	Hassabis "Frontier Safety Framework" published by DeepMind.	`✓ verified`	DeepMind Blog: "Introducing the Frontier Safety Framework"; technical report PDF. v1 2024, v2 February 2025, v3 September 2025.

Connector and Funder chapter (XVIII) — verified 2026-04-25

#	Claim	Status	Notes
CF1	Hendrycks did his PhD at UC Berkeley.	`✓ verified`	people.eecs.berkeley.edu CV. Advisers: Dawn Song and Jacob Steinhardt.
CF2	His PhD work was on robustness and out-of-distribution generalisation.	`✓ verified`	"The Many Faces of Robustness" (ICCV 2021), arXiv 2006.16241.
CF3	MMLU benchmark co-authored by Hendrycks.	`✓ verified`	arXiv 2009.03300; ICLR 2021. Co-authors: Burns, Basart, Zou, Mazeika, Song, Steinhardt.
CF4	MATH benchmark co-authored by Hendrycks.	`✓ verified`	arXiv 2103.03874; NeurIPS 2021.
CF5	ETHICS benchmark co-authored by Hendrycks.	`✓ verified`	arXiv 2008.02275; ICLR 2021.
CF6	Hendrycks founded the Center for AI Safety; serves as Director.	`✓ verified`	safe.ai/about. CAIS founded 2022 with Oliver Zhang.
CF7	CAIS Statement on AI Risk published 30 May 2023.	`✓ verified`	safe.ai/work/press-release-ai-risk.
CF8	"roughly six hundred" signatories at release (original chapter wording).	`✗ wrong (fixed)`	At release on 30 May 2023, the count was over 350 signatories. The number grew over time (1,000+ by early 2026). Chapter corrected to "over three hundred and fifty other senior researchers at release, with the signatory count growing into the thousands in the years since." Note: this updates the count for claim X1 in the cross-cutting section as well.
CF9	Named signatories: Hinton, Bengio, Hassabis, Amodei, Sutskever.	`✓ verified`	All five confirmed across multiple primary sources.
CF10	Introduction to AI Safety, Ethics, and Society, CRC Press, 2024.	`✓ verified`	Routledge listing; published 31 December 2024; ISBN 9781032869926; 544 pages.
CF11	Hendrycks advises xAI.	`✓ verified`	TheOrg.com listing. Safety Adviser to xAI; symbolic $1 salary. Also advises Scale AI from November 2024.
CF12	Karnofsky co-founded GiveWell in 2007 with Elie Hassenfeld.	`✓ verified`	givewell.org/about/story.
CF13	Open Philanthropy co-founded by Karnofsky with Cari Tuna and Dustin Moskovitz "in the early 2010s" (original chapter wording).	`~ partial (fixed)`	Open Philanthropy was founded in 2014. Chapter corrected to "in 2014." Source: openphilanthropy.org.
CF14	Cold Takes personal blog at coldtakes.com (also cold-takes.com).	`✓ verified`	cold-takes.com/about. Karnofsky is sole author.
CF15	The Most Important Century essay series on Cold Takes, 2021–2022.	`✓ verified`	cold-takes.com/most-important-century. Series began 22 September 2021; extended series published through June 2022.
CF16	PASTA scenario (Process for Automating Scientific and Technological Advancement) introduced in The Most Important Century.	`✓ verified`	cold-takes.com/transformative-ai-timelines-part-1-of-4-what-kind-of-ai.
CF17a	Open Phil funded MIRI.	`✓ verified`	openphilanthropy.org grants pages. Multiple grants totalling $11M+.
CF17b	Open Phil funded ARC.	`✓ verified`	openphilanthropy.org grants pages. ARC founded April 2021 by Paul Christiano.
CF17c	Open Phil funded METR (original chapter wording).	`~ partial (fixed)`	No direct OP grants to METR confirmed in this verification pass; METR was a 2023 ARC spin-out, and its funding has come substantially from The Audacious Project / TED. Chapter corrected: METR is no longer claimed as a direct OP grant recipient. The reference is reframed to "ARC (out of which METR later spun)" in the funding-list sentence, and the descriptive paragraph now reads "METR, the 2023 ARC spin-out, has been part of an evaluations ecosystem that Open Philanthropy has helped sustain in the broader picture even where individual grants have flowed via different channels."
CF17d	Open Phil funded Redwood Research.	`✓ verified`	openphilanthropy.org/grants/redwood-research-general-support-2023. $5.3M in 2023 alone.
CF17e	Open Phil funded CSET.	`✓ verified`	openphilanthropy.org/grants/center-for-security-and-emerging-technology-general-support. CSET established January 2019 with $55M OP grant; subsequent $42M brings total to $100M+ through 2025.
CF17f	Open Phil funded the Center for Human-Compatible AI (CHAI) at Berkeley.	`✓ verified`	openphilanthropy.org/grants/uc-berkeley-center-for-human-compatible-artificial-intelligence-2021. $11.4M over 5 years; cumulative over $17.1M.
CF17g	Open Phil funded the Centre for the Governance of AI (GovAI) at Oxford.	`✓ verified`	openphilanthropy.org/grants/centre-for-the-governance-of-ai-general-support-2. $2.5M over two years.
CF18	Karnofsky stepped back from day-to-day Open Philanthropy leadership in 2023.	`✓ verified`	Stepped down as co-CEO July 2023 after a leave of absence beginning March 2023.
CF19	Karnofsky's recent role at Anthropic.	`✓ verified`	Joined Anthropic in January 2025 as Member of Technical Staff, working on Responsible Scaling Policy. (He is married to Anthropic President Daniela Amodei — a relationship the chapter does not currently mention but which is publicly known and which book publication may want to surface for full transparency.)

Synthesis chapters and Coda — verified 2026-04-26

The synthesis chapters (XXI Convergence Without Agreement, XXII The Asymmetry, XXIII Why Doomer Is a Weak Word, XXIV What a Mature Civilisation Would Do) and the Coda (Messengers from Inside the Blast Radius) draw substantially on facts verified in earlier chapter passes. The combined verification pass on 26 April 2026 (a) cross-checked references against the existing audit trail above and (b) verified new factual claims not previously checked. No chapter edits were required. All cross-references consistent with the audit trail.

#	New claim	Status	Notes
Sy1	Hinton has cited extinction-probability numbers "in the tens of per cent" in interviews.	`✓ verified`	Confirmed via Chrome MCP search 26 April 2026 against multiple independent sources citing BBC and Guardian interviews late December 2024. Verbatim Hinton phrasing: "10 to 20 percent chance" AI will lead to human extinction within the next three decades, with reference to AI taking control of the world from humans "leading to the death of many or most humans." Tech Times (28 Dec 2024); Times of India (31 Dec 2024); Towards Data Science (15 Jul 2025); Indigo Dergisi (Dec 2024). The chapter's "tens of per cent" framing is supported and could be sharpened to "10 to 20 per cent" if the verbatim phrasing is preferred.
Sy2	Yudkowsky's extinction-probability estimates are "higher" than Ord's six per cent.	`✓ verified`	TIME, March 2023: Yudkowsky's public estimates run as high as "virtually 100%" or "over 95%" depending on the framing. Substantially higher than Ord's six per cent.
Sy3	Anthropic's "Responsible Scaling Policy".	`✓ verified`	anthropic.com/news/anthropics-responsible-scaling-policy. RSP v3.0 current as of April 2025.
Sy4	OpenAI's "Preparedness framework".	`✓ verified`	openai.com/index/updating-our-preparedness-framework/. Official name: "Preparedness Framework". v2 current as of April 2025.
Sy5	Apollo Research.	`✓ verified`	apolloresearch.ai. London-based AI evals organisation, founded ~2023–2024. Remit: interpretability and behavioural model evaluations, focus on detecting deceptive alignment in frontier AI.
Sy6	Survival and Flourishing Fund (SFF) is Tallinn-funded and x-risk-relevant.	`✓ verified`	survivalandflourishing.fund. Founded 2019 by Jaan Tallinn (~$900M contributed). $141M distributed since 2019; 86% to AI safety in 2025 round.
Sy7	AI Safety Summit chronology: Bletchley UK (1–2 November 2023), Seoul (21–22 May 2024), Paris (10–11 February 2025).	`✓ verified`	Future of Life Institute timeline; Wikipedia; UK Government records; CSIS; onu.delegfrance.org. All three summit dates confirmed.
Sy8	OECD AI Principles.	`✓ verified`	oecd.ai. Adopted 2019, updated May 2024. Five value-based principles plus five recommendations. Endorsed by 47 jurisdictions.
Sy9	G7 commitments on AI.	`✓ verified`	G7 Leaders' Statement on AI for Prosperity (June 2025); Hiroshima AI Process (2023–present); G7 reporting framework (February 2025). Multiple substantive outputs confirmed.
Sy10	UN High-level Advisory Body on AI and the September 2024 Governing AI for Humanity report.	`✓ verified`	un.org/ai-advisory-body. Established October 2023; final report released September 2024 (39 members, 2,000+ consultation participants).
Sy11	EU AI Act characterisation: "implementation is uneven, rules on general-purpose AI still being clarified" (April 2026).	`✓ verified`	EU AI Act Service Desk; April 2026 status updates. As of April 2026: Member State market-surveillance authority designations in progress; GPAI Code of Practice voluntary as of 2 August 2025; harmonised standards (prEN 18286) still in public-enquiry phase.
Sy12	Hinton's Nobel banquet was on 10 December 2024.	`✓ verified`	nobelprize.org; multiple primary and secondary sources. Stockholm City Hall.
Sy13	Coda Hinton quote: "There is also a longer term existential threat that will arise when we create digital beings that are more intelligent than ourselves. We have no idea whether we can stay in control."	`✓ verified`	Already confirmed in audit trail row H5 (Hinton chapter verification). Cross-check: text in coda matches the verified primary source.

Cross-checks confirmed consistent with the audit trail (no edits required): Hinton's "20 to 50 years" → "20 years or less" timeline shift (H2-adjacent + LC21); Legg's 50% by 2028 (Co16); Kokotajlo's median 2028 → later trajectory (M14, M15); LeCun's "decade or more" timeline (LC15); Ord's six per cent over the century; CAIS Statement signatory count "over 350 at release" (X1, CF8); Hinton Nobel banquet existential-risk language (H5); Hassabis 2024 Nobel Chemistry (Ha2); Christiano "head of AI safety" at US AISI (CF22); UK AISI / AI Security Institute renaming in 2025 (Co21); DeepMind Frontier Safety Framework (Co23).

How to use this file

Pick a row, find the primary source, paste the citation in the Notes field, change status.
If a claim turns out to be wrong, mark ✗ wrong (fix) and open a corresponding edit task in the essay.
Before publication, every row should be ✓ verified, ! contested (with the contestation explained in the essay), or removed.
New claims added during revisions go in here first.