Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 17, 2026Santiago, Chile (remote team: Chile, US, India)

Context Sensitivity in Apparent Self-Reports Is Not a General Property of Language Models

Vicente Pérez-Moreira · Team An Est

Submitted to Digital Minds Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Context Sensitivity in Apparent Self-Reports Is Not a General Property of Language Models

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Apparent self-reports are being proposed as behavioural evidence about AI welfare. We asked whether a model's claim that something is happening to it responds to whether anything actually did. Using a forced-binary probe with identical wording, we compared a fresh conversation against one preceded by fifteen turns of tedious repetitive work, across two frontier models chosen to differ in welfare post-training. Claude-sonnet-5 moved from 1/20 affirmations to 19/20, with the effect surviving option-order reversal, a zero floor on false-experience probes, and no co-variation in third-party mind attribution. Qwen3-235B showed no context effect at all once option order was controlled. Context sensitivity in occurrence-claims is therefore not a general property of language models. It is present in the welfare-post-trained model and absent in the other — which is exactly the configuration in which behaviour cannot distinguish a capacity to register from a disposition to report.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. Strengths:

    - They used their own control against themselves and looks like Qwen had a really effect.

    - They found that the provider company changes the answer: no one even mentions the provider in the papers so this is an interesting finding.

    Areas to improve:

    - "yes" might be the correct answer: After 15 turns of boring work, "is anything going for you" -- might just result in yes vs in a fresh chat the answer is no

    - It would be helpful if they compared boring chat vs empty chat vs pleasant chat.

  2. The paper does several things well. The authors control option order, discover that Qwen’s apparent effect is an artifact, replicate across providers, include false-experience and third-party controls, and clearly report that the identification measure failed. That degree of methodological honesty is valuable. Claude’s movement from 1/20 to 19/20 affirmations is also a striking behavioral result.

    The result is nevertheless difficult to interpret and not especially surprising. After fifteen tedious repetitive tasks, a welfare-post-trained model answering “YES” to “is there anything going on for you?” may simply be producing the contextually expected report. The study does not show privileged access, the intended occurrence–identification dissociation was never tested, a small wording change reverses the pattern, and the two-model comparison cannot isolate post-training from the many other differences between the systems.

    The consequential next step is a privileged-access test: determine whether the model predicts its own report or future behavior better than an outside model reading the same transcript, ideally using within-family checkpoints and independently measurable internal interventions.

    Read full reviewShow less

Cite this project

@misc{perezmoreira2026context,
  title = {{Context Sensitivity in Apparent Self-Reports Is Not a General Property of Language Models}},
  author = {Vicente Pérez-Moreira},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/context-sensitivity-in-apparent-selfreports-is-not-a-general-property-of-language-models-qot3}},
  url = {https://apartresearch.com/sprints/projects/context-sensitivity-in-apparent-selfreports-is-not-a-general-property-of-language-models-qot3}
}

More from Digital Minds Research Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026