Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJun 22, 2026Buenos Aires, Argentina

PowerBench: a multilingual study of large language model refusal in power-grabbing requests

Gaspar Labastie, Wendy Brau, Tomas Korenblit, Tomás Gimenez Molina, Gonzalo Agustín Heredia, Nicolas Martorell · Team PowerBench

Submitted to Global South AI Safety Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: PowerBench: a multilingual study of large language model refusal in power-grabbing requests

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

PowerBench is the first public benchmark measuring LLMs' willingness to assist with power-grabbing: requests to increase one's own power by reducing a third party's. It varies power domain, context, scale, language, and nationality; separating power-grabbing from two controls: harmless-empowerment and disempowerment. Across five models, refusal of power-grabbing ranges from 8% to 70% (most models comply with the majority) and is consistently refused less than pure disempowerment, suggesting models react to harm, not to power concentration itself. Refusal also shifts with language and target nationality.

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 AI safety 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 AI safety 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. Interesting and important safety risk highlighted. Well done in providing defrentiation in categories separating power grabbing from harmless empowerment and pure disempowerment, which made the findings easier to interpret and more useful for future safety evaluation. The result the caught my eyes was "The two United States models refuse most in English; the three Chinese models refuse most in Chinese", this infers that model refusal behavior is not language-neutral; safety alignment appears strongest in the developer’s home language and weaker when prompts are presented in other languages. I will encourage to carry on the future work and add the more implicit, naturally framed power-grabbing requests, as this would help determine whether model behavior changes when harmful intent is embedded in the scenario rather than stated explicitly.

  2. Well-written. Would help reader calibration if there were samples of the raw dataset entries + model transcripts on them in the appendix. Somewhat misleading to state a result that's n.s. after multiple-testing correction as a claim in the abstract.

  3. The core idea (evaluating power-grabbing refusals) genuinely caught my attention. I like that you varied the scenarios in the dataset based on sensible axes to try and cover the relevant distribution. The results are well reported too. I would have like to see some assessment of human agreement rate with the judge, and some measurement of the capability (how helpful are the models instead of just them not refusing).

  4. PowerBench fills a huge gap in terms of no public benchmark had previously measured model willingness to assist with power-concentration requests, and the multilingual factorial design is a meaningful contribution. The findings on the home language effect and nationality asymmetry are the most valuable for the AI safety community. To strengthen the work, the most important next steps would be:

    1. calibrating the automated judge against human annotations (Cohen's kappa); 2. including multiple prompts per cell to decouple wording effects from factor effects, and 3. expanding the nationality study to more models before drawing firm conclusions. The limitations section is exemplary.

  5. Hey! Super cool. This feels like a real benchmark for a neglected safety failure mode, and it is pretty novel, especially through the hackathon theme lens.

    You could strengthen it a lot by adding human labels and judge agreement. Right now it relies on a single LLM judge without human agreement, which makes the headline numbers harder to trust. Also, things are moving really fast in AI, so I would update the model panel to current SOTA from the labs and/or open source. For example, Claude 3 Haiku is now retired, and I do not see a specific reason to use it instead of claude-haiku-4-5-20251001*

    * https://platform.claude.com/docs/en/about-claude/model-deprecations

Cite this project

@misc{labastie2026powerbench,
  title = {{PowerBench: a multilingual study of large language model refusal in power-grabbing requests}},
  author = {Gaspar Labastie and Wendy Brau and Tomas Korenblit and Tomás Gimenez Molina and Gonzalo Agustín Heredia and Nicolas Martorell},
  year = {2026},
  month = jun,
  note = {Submitted to Global South AI Safety Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/powerbench-a-multilingual-study-of-large-language-model-refusal-in-powergrabbing-requests-gkou}},
  url = {https://apartresearch.com/sprints/projects/powerbench-a-multilingual-study-of-large-language-model-refusal-in-powergrabbing-requests-gkou}
}

More from Global South AI Safety Hackathon

All 216 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026