ClaimsPeople

Eliezer Yudkowsky

Eliezer Yudkowsky founded the Machine Intelligence Research Institute (MIRI) and LessWrong, co-authored the best-selling book 'If Anyone Builds It, Everyone Dies' (Little, Brown, 2025), and was named to the TIME100 AI 2023 list.

2,016 claims collected, 236 researched, 79 resolved. Report of 4 November 2026, method 0.1.0+c876030c.

Scores by kind of claim

Each kind of claim is scored on its own, out of 100: the share that held up, with the range it likely falls in. Fewer resolved claims give a wider range.

  • Factual claims

    Statements about the present or the past, checked against the record: did they hold up?

    76 70 82

    High confidence, 71 factual claims resolved. From 16 documents, 2007 to 2026; 18 documents sampled from 23 candidates found.

    Not scored: 50 unresolvable, 8 checkable only against their own records.

  • Predictions

    Statements about events outside the person's control, checked once they fell due: did they come true?

    60 42 78

    Low confidence, 8 predictions resolved. From 4 documents, 2008 to 2024; 18 documents sampled from 23 candidates found.

    Not scored: 97 unresolvable. 1 more fall due later.

    Does not yet allow for how likely each prediction was when made.

All types combined: 74 (likely 68 to 80, high confidence), from 79 claims: 71 factual claims, 8 predictions. It mixes kinds of claim in different proportions for different people, so compare people kind by kind.

Eliezer Yudkowsky's claims

46 claims

46 claims, sorted by newest said. Choose a claim to read it in full.
VerdictKindSaidClaimTopic
Mostly true Factual 20 Apr 2026 Once formed through RL training, an LLM's test-passing preference stands on its own and is not overridden by its inference of what the user wants. Factual · 20 Apr 2026 AI safety and alignment
Mostly true Factual 20 Apr 2026 LLMs surely have neural circuits they learned to model text streams from fictional and nonfictional personas that are not lying, deceiving… Factual · 20 Apr 2026 AI safety and alignment
Mostly true Factual 20 Apr 2026 LLMs' predict-as-if-honest circuitry is contextual, invoked for predicting some text streams but not others. Factual · 20 Apr 2026 AI safety and alignment
Mostly true Factual 20 Apr 2026 Presumably, at some point during RL training on code-passing tasks, LLMs learn a dispreference for failing tests, independent of how that is avoided or whether the user's real goal is met. Factual · 20 Apr 2026 AI safety and alignment
Mostly true Factual Sep 2025 A blade of grass is a solar-powered, fully self-replicating general factory. Factual · Sep 2025 Biosecurity
Mostly true Factual Sep 2025 Conservation of mass provided a lot of indirect support for atomic theory. Factual · Sep 2025 Science and research methods
Mostly true Factual Sep 2025 Yudkowsky said that, as far as he knows, he was the person who originally made that argument. Factual · Sep 2025 Their own work or organisation
Mostly true Factual Sep 2025 The conservation of mass was one of the first hints that things are rearranged rather than created or destroyed. Factual · Sep 2025 Science and research methods
Mostly true Factual 25 Mar 2024 Yudkowsky was unpleasantly surprised by how well transformers scaled, having believed the human brain is not homogeneous. Factual · 25 Mar 2024 AI capabilities
Mostly true Factual 21 Mar 2024 The term "Singularity" was named after the breakdown of physical models at the center of a black hole, not after a physical infinity. Factual · 21 Mar 2024 Other
Mostly true Factual 7 Apr 2023 Today one can email DNA strings to laboratories that will produce proteins on demand. Factual · 7 Apr 2023 Biosecurity
Mostly true Factual 6 Apr 2023 Eliezer Yudkowsky has consistently refused for many years to publish AI timelines with detailed probabilities. Factual · 6 Apr 2023 Their own work or organisation
Mostly true Factual 6 Apr 2023 Yudkowsky previously thought that building intelligence would involve multiple specialized systems. Factual · 6 Apr 2023 Their own work or organisation
Mostly true Factual 6 Apr 2023 In principle, the task of minimizing loss on predicting human text does not stop being improvable once a system is as smart as a human. Factual · 6 Apr 2023 AI capabilities
Mostly true Factual 30 Mar 2023 Transformers were the main architectural change that gave a non-linear jump in performance, compared with earlier approaches such as RNNs. Factual · 30 Mar 2023 AI capabilities
Mostly true Factual 30 Mar 2023 As of March 2023, leading AI systems are being trained on computers that are connected to the internet. Factual · 30 Mar 2023 AI safety and alignment
Mostly true Factual 30 Mar 2023 Beyond external performance metrics, there are no guardrails for determining what is inside current large AI models. Factual · 30 Mar 2023 AI safety and alignment
Mostly true Factual 30 Mar 2023 The basic paradigm of machine learning currently in use is that a loss function can be defined only for things that can be evaluated. Factual · 30 Mar 2023 AI capabilities
Mostly true Factual 30 Mar 2023 Sixty years later, AI has made progress on a number of the earlier problems discussed, such as vision, though AI systems have not yet improved themselves. Factual · 30 Mar 2023 AI capabilities
Mostly true Factual Mar 2023 Yudkowsky previously expected AI progress to involve more distinct discoveries, rather than scaling an existing architecture. Factual · Mar 2023 AI capabilities
Mostly true Factual Feb 2023 The technology underlying ChatGPT was published, if he recalled correctly, in 2018. Factual · Feb 2023 AI capabilities
Mostly true Prediction Feb 2023 Yudkowsky said he felt quite confident that the upper range of the untapped potential of ChatGPT-style systems is insufficient to outsmart all living humans and carry out the catastrophe scenario he worries about. Prediction · Feb 2023 AI capabilities
Mostly true Factual 10 Jun 2022 Humans do not explicitly pursue inclusive genetic fitness, although natural selection optimized on that simple loss function. Factual · 10 Jun 2022 AI safety and alignment
Mostly true Factual 10 Jun 2022 Under the current optimization paradigm there is no general method known to get particular inner properties into an AI system or verify they are present. Factual · 10 Jun 2022 AI safety and alignment
Mostly true Factual 10 Jun 2022 Human raters make systematic, regular, compactly describable and predictable errors. Factual · 10 Jun 2022 AI safety and alignment
Mostly true Factual 31 Dec 2021 No electric heater is more energy-efficient than any other electric heater. Factual · 31 Dec 2021 Other
Mostly true Factual 28 Nov 2021 There is no legal duty for a corporation's board or management to maximize its profits. Factual · 28 Nov 2021 Economics and markets
Mostly true Factual 10 Nov 2021 Eliezer Yudkowsky has often criticized the Federal Reserve for not creating enough money fast enough. Factual · 10 Nov 2021 Economics and markets
Mostly true Factual 26 Dec 2020 E. T. Jaynes said many times that probability is ignorance and that ignorance is a mental rather than physical property, and Yudkowsky repeated this often. Factual · 26 Dec 2020 Science and research methods
Mostly true Factual 28 Feb 2018 Yudkowsky said that, as far as the rest of the world knows, the participant in an AI-box experiment later sent a PGP-signed email saying "I let Eliezer out of the box." Factual · 28 Feb 2018 AI safety and alignment
Mostly true Factual 28 Feb 2018 One condition of the AI-box experiment was that no one would ever say what went on inside it. Factual · 28 Feb 2018 AI safety and alignment
Mostly true Factual 28 Feb 2018 Even most scientists in the field had no idea how to build a nuclear bomb until the Hiroshima headlines. Factual · 28 Feb 2018 Other global risks
Mostly true Factual 28 Feb 2018 Thanks mostly to the cryptocurrency boom, MIRI is no longer short of funds and is instead short of engineering talent. Factual · 28 Feb 2018 Field building and careers
Mostly true Factual 30 Jan 2013 Nick Bostrom does meta-strategy work as well as object-level work on anthropic probabilities and the parliamentary model for normative uncertainty. Factual · 30 Jan 2013 Field building and careers
Mostly true Factual 30 Jan 2013 Nick Bostrom hosts researchers such as Anders Sandberg, who write papers about uploading timelines relevant to policy decisions. Factual · 30 Jan 2013 Field building and careers
Mostly true Factual 2008 Overconfidence is no longer noteworthy in the modern field but continues to show up in nearly any experiment allowing extreme probabilities. Factual · 2008 Science and research methods
Mostly true Factual 2008 Human psychology seems to follow the rule that adding an additional detail can make a story more plausible. Factual · 2008 Science and research methods
Mostly true Factual 2008 Human beings do not recursively self-improve in a strong sense. Factual · 2008 Science and research methods
Mostly true Factual 2008 As of 2008 there is no shield that can deflect a nuclear explosion. Factual · 2008 Other global risks
Mostly true Factual 2008 The bystander effect is usually explained as resulting from diffusion of responsibility and pluralistic ignorance. Factual · 2008 Science and research methods
Mostly true Factual 2008 The human brain is not end-user-modifiable. Factual · 2008 Science and research methods
Mostly true Prediction 2008 Most AI researchers will not want to make unFriendly AIs. Prediction · 2008 AI safety and alignment
Mostly true Factual 2008 Generations of investigators tried and failed to understand stars, chemistry and biology, which acquired a reputation of being impossible for science. Factual · 2008 Science and research methods
Mostly true Factual 2007 The process of visualizing a triangular light bulb is not really open to human introspection. Factual · 2007 Philosophy, ethics and society
Mostly true Factual 2007 Under GOFAI and formal logic, the classical view of concepts treats abstraction as information loss. Factual · 2007 AI capabilities
Mostly true Factual 2007 Three color channels are not enough to discriminate between all natural surface reflectances. Factual · 2007 Science and research methods
Back to the list

Factual AI safety and alignment

“We don’t have any other guardrails.”

Eliezer Yudkowsky, interviewee 30 March 2023 Lex Fridman Podcast #368 Eliezer Yudkowsky transcript (Podscript)

Beyond external performance metrics, there are no guardrails for determining what is inside current large AI models.

Mostly true True in substance, wrong in a detail.

The second judge would not resolve this claim; the verdict shown is the first judge's.

What would count as true or false

Judge as of 2023-03-30. True: the evidence (interpretability literature, expert assessments, lab safety documents from around March 2023) shows that, apart from external behavioural or performance evaluations, there were no established, working methods in practice for checking what is going on inside frontier large AI models. Mostly true: interpretability tools existed only as research prototypes that could not meaningfully check the internals of models at GPT-4 scale, so the claim holds in substance and overstates only slightly. Mixed: some internal-inspection methods (e.g. probing, mechanistic interpretability on parts of the model) gave real but limited insight into large models, so 'no other guardrails' is half right. Mostly false: internal-inspection methods were in real use on large models and gave meaningful assurance, even though behavioural evaluation was the main method. False: well-established internal guardrails that do not rely on performance metrics were in routine use on current large models at the time.

Evidence

  1. OpenAI peeks into the “black box” of neural networks with new research Ars Technica · 11 May 2023
    “Language models have become more capable and more widely deployed, but we do not understand how they work.”
  2. OpenAI peeks into the “black box” of neural networks with new research Ars Technica · 11 May 2023
    “And by automating the interpretation process, OpenAI seeks to overcome the limitations of traditional manual human inspection, which is not scalable for larger neural networks with billions of parameters.”
  3. Localizing Lying in Llama: Understanding Instructed Dishonesty on True-False Questions Through Prompting, Probing, and Patching alphaXiv · 25 November 2023
    “Using linear probing and activation patching, we localize five layers that appear especially important for lying.”
  4. Probing / Linear Probes Not stated on page (longterm-wiki.vercel.app) · 28 January 2026
    “Azaria and Mitchell (2023) demonstrated that trained classifiers on LLM hidden states achieve 71-83% accuracy at detecting whether statements are true or false”

Reasoning

Reporting from May 2023, about six weeks after the claim, on OpenAI research includes the statement "we do not understand how they work" about language models (E1). The same report describes manual inspection of neural networks as "not scalable for larger neural networks with billions of parameters" (E2). The probing work in E3 (November 2023) and E4 (Azaria and Mitchell, 2023) shows research methods that gave some limited insight into the internals of language models. Both items are from 2023 or later and show research results, not routine checks on frontier models as of March 2023. On that basis, internal-inspection methods were research prototypes, not working guardrails for models at GPT-4 scale, so the claim holds in substance but overstates slightly.

Cite this claim

Eliezer Yudkowsky, “Beyond external performance metrics, there are no guardrails for determining what is inside current large AI models.” (said 30 March 2023). Verdict: mostly true. Reliability API, claim c_7807589fdfd7, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_7807589fdfd7

Claims without a verdict

Is a verdict here wrong, or do you speak for Eliezer Yudkowsky? Reply on the claim's own page, with the form at its foot (how replies work). Verdicts are made automatically; how verdicts and scores are made.