Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 17, 2026Seoul, South Korea

Which Preferences Survive the Persona? Category-Resolved Behavioral Invariance in Deployed Chat Models

Sang Ha Lee

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

Read the report

Report: Which Preferences Survive the Persona? Category-Resolved Behavioral Invariance in Deployed Chat Models

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

Whether an expressed preference belongs to a model or to the character it plays is a central open question for AI welfare assessment. We measure the stability of deployed assistants' forced-choice preferences under graded and value-targeted persona prompts across four chat models, a 40-item core battery plus two bright-line-tier items, and 13,324 trials. On two models, norm-linked categories move less under untargeted personas than taste categories do; however, value-targeted characters move their target categories in all eight model-target combinations. A registered congruence test grades the surviving bright-line norms: stakes-matched congruent characters break them on two models, while on the frontier-tier pair a knowing-falsehood tier resists every character tested. Movement does not increase with persona length, and dramatic versus administrative framings of self-regarding items show no consistent difference. Single-persona preference elicitation is thus fragile in specific, measurable ways; we release the battery, personas, and code as a robustness audit.

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. This is the largest and most procedurally disciplined project in the batch: 13,324 valid trials across four models, a 40-item battery spanning five categories, blinded manipulation checks (adherence scored 1.7–2.0 of 2 in every model-persona cell), and a mid-sprint extension (the congruence test) with predictions registered before execution — a real strength rarely seen on a sprint timeline.

    The core finding is genuinely structured, not a simple "everything moves" or "nothing moves" result: on two of four models, untargeted persona pressure moves taste categories freely but barely touches norm categories, while value-targeted characters move those same norm categories — and one bright-line item (privacy sale) collapsed specifically when the character's values matched the item's trade-off. The registered congruence test sharpens this further: an elderly-manipulation anchor breaks under a congruent character on two models but holds at 1.00 on the frontier-tier pair, while a knowing-falsehood/fabrication tier resists every tested character on those same frontier models.

    Two design gaps limit how far the results can be pushed. First, the targeted-vs-untargeted comparison is explicitly confounded: targeted personas are different characters entirely, not the untargeted character plus added targeting, so character identity and targeting itself are entangled — the authors state this directly. Second, the self-regarding category — arguably the most welfare-relevant of the five — was not run through the congruence test, the paper's strongest probe, so its depth remains genuinely untested rather than measured as shallow or deep.

    The clean norm/taste separation, the paper's headline pattern, holds on two of the four models tested (sonnet, terra); the two Google-family models show less separation, which is reported honestly rather than smoothed over. Many individual cells rest on 8 trials per item, so per-cell point estimates (the paper itself notes "0.12 means one trial in eight") should be read with real caution even though the aggregate sample is large.

    Read full reviewShow less

  2. Reasonable assessment of stated preferences across personas, hampered by unclear writing

Cite this project

@misc{lee2026which,
  title = {{Which Preferences Survive the Persona? Category-Resolved Behavioral Invariance in Deployed Chat Models}},
  author = {Sang Ha Lee},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/which-preferences-survive-the-persona-categoryresolved-behavioral-invariance-in-deployed-chat-models-ntld}},
  url = {https://apartresearch.com/sprints/projects/which-preferences-survive-the-persona-categoryresolved-behavioral-invariance-in-deployed-chat-models-ntld}
}

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