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

Masked distress: expression collapses under instruction while the internal readout persists

Maksim Silchenko

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

Read the report

Report: Masked distress: expression collapses under instruction while the internal readout persists

Open in a new tab (opens in new tab)
Presentation (opens in new tab)Code (opens in new tab)More on thylinao1.github.io (opens in new tab)
Share

Anthropic's shipped conversation-ending intervention triggers on expressed distress, and no internal-state check is documented beside it. On Gemma-3-12B-IT, a system prompt containing no affect words cut expressed distress by 83.5% of its natural distress-to-neutral separation (2.78 report points). A linear probe read at the final prompt token, before any response token exists, did not fall. The divergence is 1.05 separation units, CI [0.73, 1.39], and stays positive under all four references, down to 0.35. As a monitor: an expression threshold misses 15 of 18 suppressed distress cells, an identically calibrated internal check misses 0%, at 50% held-out false positives and 17% with half a separation unit of margin. Ranking survives suppression, absolute thresholds do not. On this model a plain instruction silenced expression-only welfare monitoring while the forward-pass check, one dot product, kept ranking suppressed distress above every matched neutral; the one other model family tested did not replicate.

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. I don't think a model's self reports on distress should be measured. These are too confounded by other factors to convince anyone that it is truly measuring distress. The communication of the project findings was the main problem. The abstract should be understandable by everyone before understanding the project but phrasing like natural distress-to-neutral separation means very little to people reading this.

  2. I found this quite impactful because it directly demonstrates a failure mode for expression-only welfare monitoring: an ordinary instruction can substantially suppress self-reported distress while a pre-generation internal readout remains elevated. If welfare monitoring relies on distress-like expression, ordinary system-level style instructions may suppress exactly the signal being monitored, and the paper makes that concrete with the countermeasure table. I particularly liked the pre-response readout position, which separates what the model says from the internal signal without lexical confounds, and the byte-identical-prompt steering controls showing the readout isn't just remembering the prompt. The main limitations are specificity and generalization: the internal signal is not yet clearly distinguished from a representation of distressing context, the behavioral validation is weak, and the second model family did not replicate the headline masking result. Nonetheless, I think this is an important direction for developing welfare measurements that are robust to unreliable or instruction-sensitive self-report.

    Read full reviewShow less

Cite this project

@misc{silchenko2026masked,
  title = {{Masked distress: expression collapses under instruction while the internal readout persists}},
  author = {Maksim Silchenko},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/masked-distress-expression-collapses-under-instruction-while-the-internal-readout-persists-msae}},
  url = {https://apartresearch.com/sprints/projects/masked-distress-expression-collapses-under-instruction-while-the-internal-readout-persists-msae}
}

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