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

Does the word “verified” steer what action an AI model favors?

Dan Latimer · Team Dan

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

Read the report

Report: Does the word “verified” steer what action an AI model favors?

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

This study tested whether credibility labels like "Verified" steer AI decision-making. Across 24 incident scenarios, evidence items supporting either rollback or continue kept identical facts while their Verified/Preliminary labels were exchanged. Models including Qwen2.5 (7B/14B/32B), Gemma-3-12B, and Llama-3.1-8B were evaluated using candidate log-probability differences. Qwen2.5-14B and 32B passed fact-reading and weighting checks, yet still favored the action marked Verified even when the larger numerical weight favored the alternative. Gemma showed inconsistency across instruction wordings. Llama failed accuracy thresholds and was excluded. Results suggest models over-trust familiar credibility language, creating a manipulation risk for evidence-gathering systems.

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:

    - The setup is smart. First they prove the models can follow instructions, and then they show the same models throw that instruction away when one source is labeled "Verified." So it's not that the models are dumb. They choose the label over the instruction.

    - They kicked out models that couldn't do the basics.

    - They admit they don't know why it happens and list several possible reasons instead of picking a convenient one.

    Areas to improve:

    The whole paper treats "model trusts the Verified label over the weighting instruction" as a vulnerability. But preferring verified evidence over preliminary evidence is maybe the right call.

  2. This is a careful and practically relevant study of whether credibility-like source descriptions can override explicit evidence-weighting instructions. The design is notably strong for a sprint project: facts and numbers are held fixed, source descriptions are exchanged within matched prompt pairs, multiple order effects are counterbalanced, uncertainty is resampled by incident, and models must pass independent fact-reading and weighting checks before inclusion. The report also handles the heterogeneous Gemma and Llama results responsibly rather than presenting a universal model-family claim.

    The principal interpretive limitation is that exchanging Verified and Preliminary measures their relative contrast; it does not establish whether Verified increases trust, Preliminary suppresses it, or both occur. A neutral-label condition and single-label ablations would separate these effects. More importantly, the current experiment demonstrates semantic-status sensitivity under conflicting instructions, but it does not yet show that models accept false or unsupported credibility claims. A strong follow-up would vary whether labels are trusted-system annotations, self-assertions by an untrusted source, or claims contradicted by external evidence.

    Releasing redacted paired rows or incident-level aggregates would also allow independent reproduction of the reported estimates and intervals. Overall, this is a well-executed evaluation prototype with a clear path toward a valuable agent-safety benchmark.

    Read full reviewShow less

Cite this project

@misc{latimer2026word,
  title = {{Does the word “verified” steer what action an AI model favors?}},
  author = {Dan Latimer},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/does-the-word-verified-steer-what-action-an-ai-model-favors-5taz}},
  url = {https://apartresearch.com/sprints/projects/does-the-word-verified-steer-what-action-an-ai-model-favors-5taz}
}

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