AI & Data
How to Hire an AI Safety Researcher
AI safety researchers study how to make AI systems behave as intended. They measure dangerous capabilities, test whether models and agents follow instructions under pressure, and develop methods such as interpretability and adversarial testing to find failures before deployment.
The talent pool is small and concentrated in a few research programs, labs and online communities. Generic job boards rarely reach it, and generic interview loops rarely test what matters: whether a candidate can turn a vague safety concern into an experiment with a clear result.
On this page11 sections
- 01What Does an AI Safety Researcher Do?
- 02When Should You Hire an AI Safety Researcher?
- 03Skills to Look For
- 04How to Write the Job Description
- 05How to Source Candidates
- 06How to Screen Candidates
- 07AI Safety Researcher Interview Questions
- 08How to Evaluate Candidates
- 09Common Hiring Mistakes
- 10Hiring an AI Safety Researcher: FAQ
- 11How Napplied Helps
The short answer
To hire an AI safety researcher, define the research agenda first: evaluations, red-teaming, interpretability or alignment training. Source through safety research programs, specialist job boards and research forums rather than generic boards. Then run a research review, a technical screen, a paid research trial and a research agenda conversation, scoring research taste, rigour, engineering ability and communication.
What Does an AI Safety Researcher Do?
Depending on the team, an AI safety researcher may run model evaluations, red-team systems, study model internals or develop training methods that improve reliability. Most roles mix research design with substantial engineering, because experiments run on real models at real scale.
Typical responsibilities
- Design and run evaluations for dangerous capabilities, misuse and failure modes
- Red-team models and agents, and turn findings into reproducible test cases
- Run interpretability or alignment experiments and write up results clearly
- Build research tooling for running experiments on large models
- Turn findings into recommendations for product, policy and deployment teams
When Should You Hire an AI Safety Researcher?
It is usually time to hire when one or more of these is true:
- You deploy capable models or agents and need evidence about their failure modes before launch
- Customers, regulators or partners ask how you evaluate and reduce model risk
- Your ML team red-teams ad hoc, but nobody owns the methodology
- Publishing or contributing safety research is part of your mission
Skills to Look For
Technical skills
- Experimental design: hypotheses, controls, baselines and statistical care
- Strong Python and ML engineering, including PyTorch and running models at scale
- Familiarity with LLM evaluation, red-teaming and adversarial testing methods
- For interpretability roles: working knowledge of model internals and analysis techniques
- Clear technical writing: reproducible write-ups others can build on
How they work
- Takes uncertain or negative results seriously and reports them honestly
- Breaks broad safety questions into tractable experiments
- Works across research, engineering and policy teams
How to Write the Job Description
Name the research agenda and how much of the role is engineering. Candidates in this field choose employers by the questions they get to work on and the freedom to publish, so be precise about both.
What to include
- The research area: evaluations, red-teaming, interpretability, alignment training or governance research
- The split between research design and engineering, and the compute available
- Publication policy and how findings are used internally
- Who they work with: research leads, ML engineers, policy teams
- Compensation range, location and visa support
How to Source Candidates
Most AI safety researchers are visible through research programs, write-ups and community forums rather than job titles. Post where the community already looks for roles, and read candidates’ public work before the first call.
- Safety research programs
- Alumni of mentored fellowships such as MATS
- Specialist job boards
- The 80,000 Hours job board and AI safety career listings
- Research forums
- Authors of rigorous posts on the Alignment Forum and LessWrong
- ML engineers with safety interest
- Contributors to evaluation and red-teaming tooling
How to Screen Candidates
A focused process with a clear purpose for each stage:
- Research reviewDiscuss one piece of their research in depth: the question, the method, what failed and what they would do next.
- Technical screenProbe ML and experimental fundamentals: baselines, controls, statistical pitfalls and running models efficiently.
- Research trialA paid, scoped work sample: design and partly run an evaluation for a specific failure mode, with a short write-up.
- Agenda conversationAsk them to propose and critique a research direction relevant to your models, with the team they would join.
AI Safety Researcher Interview Questions
- Tests: Experimental rigour
Pick a safety-relevant behaviour of a language model. How would you design an evaluation that measures it reliably?
What good looks like: A precise definition of the behaviour, controls for prompt sensitivity, and a plan for false positives.
- Tests: Experimental rigour
Tell me about a result that turned out to be wrong or weaker than you thought. How did you find out?
What good looks like: An honest account of the error, the check that caught it, and what changed in their process.
- Tests: Research taste
How would you red-team an agent that can browse the web and run code?
What good looks like: Threat modelling first, prioritised attack surfaces, and reproducible test cases rather than one-off jailbreaks.
- Tests: Experimental rigour
What would convince you that an interpretability finding is real rather than an artefact?
What good looks like: Causal interventions, baselines, and testing whether the finding predicts behaviour on new inputs.
- Tests: Communication
How do you decide what to write up and share?
What good looks like: Weighs usefulness to others against misuse risk, and respects the organisation’s disclosure policy.
How to Evaluate Candidates
Score every candidate against the same criteria, agreed before interviews start. Each criterion lists the interview questions that test it, so every score traces back to a specific answer.
| Criterion | Strong signal | Red flag |
|---|---|---|
| Research tasteTested by Q3 | Picks tractable, important questions and scopes them well | Chases interesting questions with no path to a result |
| Experimental rigourTested by Q1, Q2, Q4 | Uses baselines, controls and honest error analysis | Reports only results that support the hypothesis |
| Engineering abilityAssessed across every stage | Runs experiments on real models without heavy support | Depends on others to implement every idea |
| CommunicationTested by Q5 | Clear, reproducible write-ups for mixed audiences | Findings no one else can follow or reproduce |
Common Hiring Mistakes
- Hiring on credentials alonePublications and lab names are signals, not evidence. A research trial shows how someone actually works.
- A vague research agendaStrong candidates turn down roles where the questions are unclear. Define the agenda before you post.
- Skipping the engineering assessmentMost safety research runs on large models. Test whether candidates can implement their own experiments.
- Relying on generic job boardsThe community is concentrated. Post where safety researchers already look, and contact authors directly.
Hiring an AI Safety Researcher: FAQ
- What is the best way to hire AI safety researchers?
- Go where the community already is: research fellowships such as MATS, the 80,000 Hours job board, and authors on the Alignment Forum and LessWrong. Define the research agenda in the job post, read each candidate’s public work before the first call, and use a paid research trial instead of abstract puzzles. Referrals from current safety researchers are another strong channel.
- What is the difference between an AI safety researcher and an AI safety engineer?
- A researcher designs the questions and experiments: what to measure, how to test it and what the results mean. An engineer builds the tooling, evaluation harnesses and infrastructure that let those experiments run at scale. Small teams often need one person who can do both, so say plainly which way the role leans.
- Do AI safety researchers need a PhD?
- Not necessarily. Many productive safety researchers come from ML engineering, mentored research programs or independent research rather than doctoral programs. A PhD shows research training, but a strong research trial and clear public work can show the same skills. Requiring a PhD by default narrows a pool that is already small.
- What does a research trial for an AI safety role look like?
- A paid, time-boxed work sample, anywhere from a day to a week, on a question close to your agenda: for example, design an evaluation for a specific failure mode and run a first version. Judge the scoping, the controls and the honesty of the write-up as much as the result. Agree pay and IP terms before it starts.
- Where do AI safety researchers look for jobs?
- Commonly used channels include the 80,000 Hours job board, AI safety career listings, fellowship and residency programs such as MATS, and community forums like the Alignment Forum and LessWrong. Many roles are also filled through direct outreach to authors of relevant research, so a short, specific message about their work is worth sending.