Research Scientist, Manipulation Evaluations

Apart Research; San Francisco (CA); Remote; USD 140,000 - 200,000; Full time

Benefits offered by this job

Real ownership

Remote work

Travel budget

Learning budget

Job summary

Apart Research in San Francisco is seeking a Research Scientist for Manipulation Evals to design, build, and publish novel evaluations of AI harmful manipulation that inform EU AI Act enforcement. You will own the full pipeline and steer research directions with significant autonomy.

The role emphasizes hands-on evals, literature reviews, red-teaming, and fast, high-quality outputs. Strong technical writing for regulators is required; remote-friendly with EU-compatible time zones and travel

Qualifications

• Fast shipper with high-quality output.
• Hands-on evals experience with strong references.
• Peer-validated output with co-authored paper or impactful preprints.
• Empirical focus in safety-related evals/benchmarks.
• Strong Python and software engineering skills for eval frameworks.

Responsibilities

• Own the full pipeline from risk scoping to eval execution.
• Design novel evaluation methods for AI manipulation.
• Perform literature reviews and track state-of-the-art evals.
• Red-team results and scan transcripts for manipulation patterns.
• Deliver new results weekly with autonomous direction setting.
• Communicate technical results clearly for regulator use and public deliverables.
• Co-author papers, benchmark releases, and technical reports for EU AI Office.
• Consolidate public manipulation benchmarks into a single index.
• Develop a multilingual benchmark extension validated by native speakers.
• Contribute methodological innovations and identify better evaluation approaches.
• Draft risk scenarios that scope new evals.

Skills

Coding with agents

Eval frameworks

Python

Data analysis

Technical writing

Red-teaming

Empirical safety work

Tools

Inspect eval framework

LLM APIs

Job description

Compensation: $130,000–200,000 gross per year for Europe-based hires, flexible by experience. San Francisco compensation adjusted to local cost of living.

Apart Research is hiring a Research Scientist for Manipulation Evals to design, build, and publish novel evaluations of AI harmful manipulation that directly inform EU AI Act enforcement.

Why this role matters. From August 2026 the EU AI Office can fine frontier labs up to 3% of global annual turnover for deploying manipulative AI. To act, it needs its own credible evidence of serious concern (AI Act Art. 92–93), but the AI Act leaves the evaluation method to the \"state of the art\", which for many risk scenarios does not yet exist. The Office is severely capacity-constrained, so it relies on outside evaluation capacity, and Apart is one of the few organizations building it.

What counts as \"harmful manipulation\" here is still open, and you help define it. You set the direction: choose which risks are most important and tractable, then build evals credible enough that the Office can act on them. The quality bar is high, because weak evidence invites backlash and the wrong focus wastes scarce regulatory attention.

About Apart Research. Apart is a globally distributed AI safety research lab. We have established a Research Division to build independent harmful-manipulation evaluations for the EU AI Office ahead of EU AI Act enforcement. We've co-authored with the EU AI Office in Science, contributed to the EU AI Act Code of Practice across three of its four Working Groups, sit on the AI Act Advisory Forum, and assembled the consortium (Apart, CeSIA, Equistamp, Transluce, FAR.AI ) that won the AI Office's evaluation tender on harmful manipulation (36 months).
The division's work is independent but informed by direct conversation with EU AI Office staff.

You will

• Own the full pipeline. Develop and prioritize risk scenarios, then build and run evals end to end against frontier models using coding agents (Claude Code, Codex), analyze the results, and write them up.
• Design novel evaluation methods for AI manipulation, turning risk scenarios into rigorous, defensible measurements.
• Perform literature reviews and track the state of the art in manipulation, deception, and persuasion evals.
• Investigate model behavior and red-team your results. Scan large volumes of transcripts to find manipulation patterns, turn qualitative observations into quantitative measures, and red-team your own and others' evals.
• Deliver new results at least weekly and set research directions with significant autonomy.
• Communicate technical results so regulators can use them. Write up findings clearly enough for the EU AI Office to act on, and to academic quality standards. Co-author public deliverables (papers, benchmark releases, technical reports) and engage directly with the Office.
• A meta-analysis consolidating public manipulation benchmarks into a single comparable index.
• A new benchmark probing an under-researched threat (e.g., whether AI companions foster harmful emotional dependency in vulnerable users), with parts of the dataset open-sourced.
• A multilingual extension of an existing benchmark, validated against native-speaker judgment in non-English EU languages.
• A methodological contribution, e.g., showing a widely used benchmark fails to measure what it claims or is vulnerable to eval awareness, and proposing a better approach.
• A clearly-written risk scenario that scopes a new eval.

Must-haves

• Fast shipper, no slop. You move quickly and quality-control your own output. Fluent with coding agents (Claude Code / Codex) for building evals.
• High context on evals. Prior hands-on evals experience and strong references. Clear theory-of-change for why an eval matters; strong red-teaming instincts.
• Peer-validated output. At least one co-authored paper at a top AI/ML venue. Workshop papers or preprints with clear impact also considered.
• Empirical focus. A track record of hands-on, safety-relevant empirical work (evals, benchmarks, adversarial robustness, alignment).
• Strong engineering and data analysis. Strong Python and software engineering (eval frameworks like Inspect, LLM APIs); comfortable analyzing large, noisy datasets.
• Clear technical writing for a regulator audience. You can turn technical results into something the EU AI Office can understand and act on, while clearing academic standards of rigor.
• Self-directed. Comfortable owning open-ended research mandates and setting your own direction.

Nice-to-haves

• Research experience in AI manipulation, deception, persuasion, or sycophancy (and command of that literature).
• Measurement and study design: IRT, psychometrics, meta-analysis, human-subjects studies (RCTs, longitudinal panels), or causal inference (difference-in-differences, natural experiments).
• Native-level fluency in a non-English EU language (for multilingual evaluation and native-speaker validation).
• Familiarity with the EU AI Act or technical AI governance.
• Experience with the Inspect eval framework.

What we offer

• Real ownership of your research at an early-stage, funded EU AI safety lab, with room to grow as the division scales.
• Compensation: $130,000–200,000 gross per year for Europe-based hires, flexible by experience. San Francisco compensation adjusted to local cost of living.
• Fully remote, EU-compatible time zones; strong preference for being based in Europe.
• Travel budget for conferences, EU institutional meetings, and coworking across Europe.
• Learning and development, equipment, and home-office budgets.

Apart Research is an equal opportunity organization and welcomes applicants of all backgrounds. We may use AI-assisted tools to support review; all applications are read by a human and final decisions are made by people. We process the information you submit to assess your application and may match it against our candidate-sourcing records (public professional information and talent-directory entries) to give reviewers fuller context; we keep application data only as long as needed for recruiting, including considering you for future roles, and delete it on request. Questions about the role or how your data is processed: [e-mail] .