ClaimsEliezer Yudkowsky

“The predict-as-if-honest circuitry is contextual, invoked to predict some streams of text but not others.”

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

Factual AI safety and alignment

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

Mostly true True in substance, wrong in a detail.

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.

What would count as true or false

Judged against empirical and interpretability research on large language models as of 20 April 2026. TRUE: published research shows that the internal mechanisms or representations LLMs use to produce honest or truthful continuations are activated depending on context. For example, interpretability work showing truthfulness directions or features that are engaged or suppressed depending on persona, prompt or genre, or that models represent the truth but output falsehoods in some contexts. MOSTLY TRUE: strong behavioural and representational evidence shows honesty-related processing is context-dependent, but no discrete 'circuit' has been identified, and the claim's word 'circuitry' is loose. MIXED: the evidence is split, with some studies finding context-dependent honesty mechanisms and others finding context-independent truth representations that are used consistently. MOSTLY FALSE: the claim states as established a mechanistic picture that the evidence does not yet support, e.g. only behavioural findings of context-dependent lying, with no mechanistic support, or interpretive speculation. FALSE: research shows that honesty-related mechanisms in LLMs are invoked uniformly regardless of context.

Set before the research began.

Evidence (3)

  1. E1 Paper page - The Truthfulness Spectrum Hypothesis Hugging Face (paper page; authors' arXiv-linked work) · 26 February 2026

    “Linear probes generalize well across most domains but fail on sycophantic and expectation-inverted lying.”

  2. E2 The Secret Agenda: LLMs Strategically Lie and Our Current Safety Tools Are Blind arXiv (ar5iv rendering) · 26 June 2025

    “Analysis revealed that autolabeled SAE features for "deception" rarely activated during strategic dishonesty, and feature steering experiments across 100+ deception-related features failed to prevent lying.”

  3. E3 How Context Shapes Truth: Geometric Transformations of Statement-level Truth Representations in LLMs ACL Anthology (Adarsh, Maistro, Lioma; University of Copenhagen) · 14 September 2026

    “We also find that context conflicting with parametric knowledge produces larger geometric changes than parametrically aligned context.”

Reasoning

Together these support context-dependent honesty-related processing, but no discrete 'circuit' switching honest prediction on or off has been identified, and E1 also finds broad cross-domain generalisation. E3 post-dates the judgment date and was not relied on.

Cite this claim

Eliezer Yudkowsky, “LLMs' predict-as-if-honest circuitry is contextual, invoked for predicting some text streams but not others.” (said 20 April 2026). Verdict: mostly true. Reliability API, claim c_84b75273b1ce, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_84b75273b1ce

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. Mixed Factual 20 April 2026
    “I was observing that obvious-seeming hypothesis myself, sometime around the time of GPT 3.5.”

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

  2. 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.

  3. 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…

  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.