ClaimsEliezer Yudkowsky

“I was observing that obvious-seeming hypothesis myself, sometime around the time of GPT 3.5.”

Eliezer Yudkowsky, author on X 20 April 2026 Post by @ESYudkowsky on X, 2026-04-20 (X)

Factual AI safety and alignment

Yudkowsky himself observed the persona-selection-like hypothesis around the time of GPT 3.5.

Mixed Partly true and partly false in roughly equal measure.

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

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

The evidence is the person's or their organisation's own records.

What would count as true or false

Judge as of 2026-04-20. True: dated records (e.g. Yudkowsky's posts, tweets or comments from about late 2022 to early 2023, the GPT-3.5/ChatGPT era) show him stating a hypothesis substantially like persona selection (that LLMs model or select among characters/personas learned from text). Mostly true: such records exist but are dated a little outside that window (e.g. 2022 or mid-2023), or state the idea in a closely related but not identical form. Mixed: the records only partly match the hypothesis or the timing is unclear. Mostly false: his earliest documented statement of the idea is clearly later (e.g. 2024 or after), or only loosely related ideas appear around GPT-3.5. False: dated records show he rejected or argued against the hypothesis around GPT-3.5 time and took it up only much later. His own 2026 recollection alone does not settle the claim.

Set before the research began.

Evidence (4)

  1. E1 No, really, it predicts next tokens. — LessWrong LessWrong · 18 April 2023 Own record

    “Yes, a big point of this post is that it's not "inherently mistaken" for the reasons (a) and (b) that Eliezer gives in his linked tweets.”

  2. E2 No, really, it predicts next tokens. — LessWrong LessWrong · 18 April 2023 Own record

    “If That-Which-Predicts were a zillion miles superhuman underneath and could solve every aspect of alignment in an instant, but the mask is near-human-level because it is trained on human-level data”

  3. E3 Why do we assume there is a "real" shoggoth behind the LLM? Why not masks all the way down? — LessWrong LessWrong · 9 March 2023 Own record

    “Some people are thinking in those terms, e.g. Eliezer Yudkowsky saying that "the actual shoggoth has a motivation Z".”

  4. E4 On "ChatGPT Psychosis" and LLM Sycophancy — LessWrong LessWrong · 23 July 2025 Own record

    “2024-10-27 Yudkowsky shares his "Rasputin's Ghost" theory of confirmation bias driven LLM psychosis.”

Reasoning

E3 (March 2023) quotes Yudkowsky saying "the actual shoggoth has a motivation Z", E1 (April 2023) refers to reasons he gave in linked tweets, and E2, from the same April 2023 post, describes a superhuman predictor behind a "near-human-level" mask but does not clearly attribute this to him. This suggests that around the GPT-3.5 era he was using a shoggoth/mask framing related to persona selection. However, his framing stresses a real underlying shoggoth with its own motivations. E3 contrasts that view with "masks all the way down", so it only partly matches the persona-selection hypothesis. E2 is also not clearly attributed to him, so the records match the timing but only partly match the hypothesis.

Cite this claim

Eliezer Yudkowsky, “Yudkowsky himself observed the persona-selection-like hypothesis around the time of GPT 3.5.” (said 20 April 2026). Verdict: mixed. Reliability API, claim c_8f3dc306412c, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_8f3dc306412c

This claim is in dataset release v1.1.2, with the same verdict.

Reply to this claim

For Eliezer Yudkowsky, or someone speaking for them: a correction or counter-evidence. Replies are read before anything is shown; an approved reply shows your name, role, date, statement and link. Your email address is never shown and is deleted 30 days after the decision (privacy).

More factual claims by Eliezer Yudkowsky

  1. Mostly true Factual 20 April 2026
    “The test-passing preference stands on its own once formed.”

    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.

  2. Mostly true Factual 20 April 2026
    “LLMs surely have neural circuits they learned to model text streams from fictional and nonfictional personas that are not lying, deceiving, cheating.”

    LLMs surely have neural circuits they learned to model text streams from fictional and nonfictional personas that are not lying, deceiving…

  3. Mostly true Factual 20 April 2026
    “The predict-as-if-honest circuitry is contextual, invoked to predict some streams of text but not others.”

    LLMs' predict-as-if-honest circuitry is contextual, invoked for predicting some text streams but not others.

  4. Mostly true Factual 20 April 2026
    “Presumably, somewhere along the way LLMs learn a dispreference for tests that don't pass, independently of how that gets prevented or whether the new plan does what the user really wanted.”

    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.

  5. Resolved true Factual 1 April 2026
    “It is an easy con to see through, if you know the actual history and intellectual lineage of extinction concerns, and how long they predate the AI companies supposedly inventing them.”

    Concerns about human extinction from AI have a history and intellectual lineage that long predate the AI companies.

All of them

Verdicts are made automatically; how verdicts are made.