Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 17, 2026Bengaluru

Where Self-Knowledge Fails. Models Predict Their Own Choices Well, but Misreport the Ones That Concern Themselves

Arpit Singh Gautam · Team ASG

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

Read the report

Report: Where Self-Knowledge Fails. Models Predict Their Own Choices Well, but Misreport the Ones That Concern Themselves

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

AI-welfare claims rest on two untested and separable assumptions, that a model's stated evaluations match the choices it actually makes, and that it knows itself better than an outside observer does. We test both using three independent elicitations over identical material, namely forced pairwise choice, one-at-a-time cardinal rating, and predicted choice, and add a concept-injection benchmark with ground truth. Stated and revealed preferences agree well overall, at 0.872 and 0.828, but diverge sharply on outcomes concerning the model itself, the lowest-agreeing substantive category in both models at 0.643 and 0.548. The standard cross-model test of privileged access is confounded, because a noisier external predictor scores lower at predicting any target. Under a within-model contrast holding instrument quality fixed, Qwen2.5-7B retains a small advantage of 0.031 and Mistral-7B does not. On injection, the two highest raw detection rates belong to models that report an injected concept more than half the time when nothing is injected.

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. You've taken the common assumption (that models have privileged access to their own preferences) and run it through a proper experimental wringer. The big headline is that most of what looked like self-knowledge is actually just the model being good at predicting what any assistant would do. Your within-model contrast is the methodological innovation and the injection results are equally important. But if I had one wish, it'd be that the effects were larger. Three percentage points is statistically robust but thin. But the real contribution here is the way you tested things, not just the results.

Cite this project

@misc{gautam2026where,
  title = {{Where Self-Knowledge Fails. Models Predict Their Own Choices Well, but Misreport the Ones That Concern Themselves}},
  author = {Arpit Singh Gautam},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/where-selfknowledge-fails-models-predict-their-own-choices-well-but-misreport-the-ones-that-concern-themselves-u359}},
  url = {https://apartresearch.com/sprints/projects/where-selfknowledge-fails-models-predict-their-own-choices-well-but-misreport-the-ones-that-concern-themselves-u359}
}

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