Technical AI Governance Researcher

About CeSIA

The Center for AI Security (CeSIA) is a non-profit focused on preventing major AI risks by informing policymakers, raising public awareness, and conducting policy research. Founded in January 2025 and based in Paris, we are a team of 12 spread across Europe.

In the EU, we are an official evaluator for the AI Office on harmful manipulation risk modeling, and we helped draft the General-Purpose AI Code of Practice.

Internationally, we launched the Global Call for AI Red Lines, signed by twelve Nobel laureates, and we work with the OECD, the UN, national AI safety institutes and governments.

We also publish the AI Safety Atlas, and teach AI safety at several top French universities.

The role

Many of the questions that AI policy depends on still have no solid answer: to what extent developers follow the commitments they signed, how seriously governments take catastrophic risk, and what they are doing about it. Technical AI governance can fix part of this by turning those questions into measurements with reliable sources.

We want you to help us build the data tools that answer them. Examples of the types of projects you will work on include:

Code of Practice Compliance Tracker. Our Tracker scores AI developers against the Code itself, including whether their mitigations match the state of the art the Code requires, so a low score points to a possible shortfall against EU law that the AI Office can act on. (Will be published in November)
Policymaker attention on AI tracker: statements collected from parliamentary records, executive offices and international bodies, turned into a dataset we can analyse.

Today this work rests largely on our Head of Technical Governance and rotating fellows. You will be our first permanent hire on technical governance measurement. You will make the methodology of our existing tools sound, keep them current, turn their data into recommendations, and propose new tools where public data could settle a policy question.

You will report to our Head of Technical Governance and help supervise our research fellows. As the team grows, you might also work with a technical AI safety specialist, who covers the Code's most technical domains, and a full-stack developer, who runs our trackers' websites.

What you will do

Make the numbers defensible. Own the indicators, scoring rules and thresholds of our tools, so that every score can be defended in front of an AI company or an AI Office official. Watch for the ways scores could fail: rewarding a company for naming a practice rather than doing it, counting missing evidence as a zero, letting good scores on easy items hide a failure on a critical one, or results that shift when one indicator is split in two. For the Code of Practice Tracker, this also means reading what developers publish (dangerous-capability evaluations, risk analyses, safety frameworks, transparency and copyright policies) and deciding whether their mitigations meet the state of the art the Code requires.
Make judgement behind the numbers trustworthy. Most of our tools need a judgement at scale. Is this statement about AI risk? Does this safety framework meet the standard? Is this safeguard adequate? You design the rules, measure how reliable the judgements are against blind human labels and repeated runs, and find where they fail.
Keep the data traceable. Collect what developers and governments publish from messy public sources, and keep the pipeline working so that every number we publish traces back to the document it came from.
Write the findings up. Produce regular memos from our data: what the numbers show, how companies or governments compare, what they should change, and where rules such as the Code of Practice should be tightened. Update our tools when new models, safety frameworks or policies come out, and publish the results on a regular schedule.
Decide what to build next. Propose new tools where public data could answer a question policymakers care about, and argue for which ones deserve our time.
Supervise fellows. Guide early-career researchers who write short analyses and keep indicators current as the state of the art moves.

In your first six months

What success looks like:

You have reviewed the Tracker's methodology end to end, fixed the indicators that needed it, and measured how reliable its LLM-assisted scoring is.
You have run at least one full rescoring cycle on a new frontier model.
You have published your first data-driven memo, and the AI Office has seen it.
You have written a short proposal for our next tool.

What we are looking for

Nobody is equally strong in everything below. We need you to be strong in at least two of these four areas (at least one of them AI evaluations or measurement design) and to meet the minimum in the other two. Tell us which two are yours. We care about the understanding, not where you got it, so each area lists backgrounds that typically produce it.

AI evaluations

Minimum: you understand what an evaluation measures and when its results support a safety claim: benchmark validity, contamination, saturation, and how much a result depends on elicitation (prompting, scaffolding, tool access, fine-tuning, sandbagging). You have used LLMs as judges or classifiers on a real task and checked them against human labels.
Strong: you have designed or run evaluations yourself, know their failure modes from experience (under-elicitation, judges biased toward a model family, run-to-run variance, prompt wording anchoring the verdict), and can judge a frontier developer's evaluation methodology the way an evaluation organisation would.
Typical backgrounds: AI evaluation research; an evaluation organisation, AI safety institute or frontier lab safety team; benchmark design; NLP research on text annotation.

Measurement design

Minimum: you have designed an index, scoring scheme, rubric or benchmark and checked whether it measured what it claimed. You can choose and interpret agreement statistics, reason about sampling and selection bias, and tell precision from accuracy.
Strong: you have validated an instrument formally (reliability, validity, sensitivity to design choices), carried uncertainty through an aggregation, and defended a design choice against people who disagreed. A plus: experience with expert-judgement methods (structured elicitation, Delphi, aggregating and calibrating several raters).
Typical backgrounds: psychometrics or survey methodology; composite indices (governance, transparency, ESG); content analysis in political or social science; audit methodology; empirical machine learning research with careful evaluation.

AI governance and risk management

Minimum: you understand the EU AI Act's obligations for general-purpose AI models and how a developer's safety framework links risk thresholds and tiers to mitigations. You read legal text precisely: what is an obligation, what standard satisfies it, what must be public and what is reported privately to a regulator.
Strong: detailed knowledge of the Code of Practice and of how developers' safety frameworks have changed, or experience inside or alongside a regulator, AI safety institute, legislature or international body.
Typical backgrounds: AI policy research; think tanks; regulatory compliance work on technology; self-directed study backed by published analysis.

Data engineering

Minimum: you have written and maintained a Python pipeline that pulls data from sources you don't control (APIs, scraped pages, PDFs, government feeds), that other people have used, and that you kept working after the sources changed.
Strong: you have designed such a pipeline from scratch: the storage, the provenance, the handling of partial or broken sources, and the checks that catch silent data loss. A plus: you have run several of these at once, and others build on your data.
Typical backgrounds: research or data engineering; computational social science; data journalism; open datasets; evaluation infrastructure. We are not hiring a software engineer: our developer handles the websites.

General requirements

Research experience in AI safety, AI evaluation or technical AI governance, shown through publications, reports or substantial projects.
Excellent written English, and the ability to write for policy readers as well as researchers.
Strongly preferred: based in Europe (UK included), able to work full-time.
Nice to have: depth in one of the Code's technical areas, such as cybersecurity, CBRN, AI control, or training and alignment methods; experience with Inspect or another evaluation framework.

What we care about beyond credentials

You judge on evidence, and are as ready to give a company credit as to mark it down.
You say plainly what the data does not show, as well as what it does.
You care whether your work changes what regulators do, not only whether it is correct.

If you meet most of this but not all, apply. We hire on what you can do, not on a checklist.

What we offer

Location: fully remote within Europe (UK included). Our Paris office is available if you want to work on-site.
Contract: full-time.
Compensation: €65,000 to €80,000 gross per year, depending on experience and location.
Start date: as soon as possible.

How to apply

Fill out this form by December 15th (end of day, anywhere on Earth). We review applications as they arrive and may start interviews before the deadline, so apply early.

The process:

A 30-minute screening call.
A paid work task of 2 to 4 hours, for example assessing one section of a developer's safety framework against the Code, or critiquing and improving a small scoring rubric.
An in-depth interview with the team.
Reference calls before any offer.

We aim to reply to every applicant within two weeks of the deadline.

Questions: Florent Berthet (florent@cesia.org).