An AI Safety Researcher works on preventing harm from AI systems as they become more capable and more autonomous. The focus is on risks that come from the models themselves, such as dangerous capabilities, behaviour that departs from what developers intended, and failures that only appear at scale or in new settings. The work produces evaluations, training methods and safeguards that inform whether and how a model is deployed.
Organisations hire AI safety researchers because capability gains bring new failure modes that ordinary testing does not reveal. A model that can write code, browse the web and act through tools needs evaluation for misuse potential, for reliability when nobody is watching, and for whether its safety training holds. Regulators and frameworks such as the EU AI Act and the NIST AI RMF also expect developers of advanced systems to assess and manage these risks.
What does an AI Safety Researcher do?
Day-to-day work varies by team, but most AI safety researchers design experiments that reveal how models behave in situations that are hard to observe in normal use, then develop methods to improve that behaviour.
- Designing and running dangerous capability evaluations in areas such as cyber offence, biological and chemical risk, persuasion and autonomous replication
- Studying failure modes such as reward hacking, specification gaming, sycophancy and deceptive behaviour under evaluation
- Building model organisms of misalignment to test whether safety methods detect and correct specific failures
- Researching AI control, including monitoring, oversight and containment protocols for untrusted models
- Evaluating and improving safety training methods such as RLHF and constitution-based approaches
- Developing evaluation harnesses and benchmarks that can be rerun across model generations
- Writing up results for internal decision makers, policy teams and, where appropriate, the research community
Key skills and backgrounds
Most AI safety researchers have a PhD or equivalent research experience in machine learning, computer science, mathematics, physics or a related quantitative field. Some come from cognitive science, economics or philosophy and bring strong empirical or conceptual skills. Engineering ability is important in empirical safety work, where experiments on large models require solid infrastructure and careful code.
Look for candidates who can state a safety problem precisely, design an experiment that could prove them wrong and report the result even when it is disappointing. Ask how they would know if an evaluation was being gamed.
- Strong deep learning foundations, including training and fine-tuning of large language models
- Evaluation design, including threat modelling for capabilities and resistance to gaming
- Knowledge of RLHF, preference learning and their known limitations
- Familiarity with interpretability, robustness and scalable oversight research
- Python with PyTorch or JAX, and experience running experiments at scale
- Statistical care and honest reporting of uncertainty
- Clear writing for technical colleagues and policy audiences
Where AI Safety Researchers work
Frontier model labs employ AI safety researchers across alignment, evaluations, policy research and preparedness teams. Government AI safety and security institutes hire them to evaluate models independently and advise on regulation. Non-profit research organisations and academic groups run dedicated safety programmes, often funded by philanthropy. Third-party evaluation and assurance organisations need safety researchers to test models for developers and governments, and large enterprises deploying advanced AI are starting to hire safety specialists into responsible AI teams.
AI Safety Researcher vs related roles
AI safety is a broad field, and several research roles sit inside or next to it.
- An Alignment Researcher focuses on methods that make models reliably pursue intended goals, which is one major strand within the wider AI safety agenda.
- An Interpretability Researcher studies the internal computations of models to explain behaviour, providing tools that safety work can draw on.
- A Privacy-Preserving ML Researcher develops methods such as differential privacy and federated learning to protect data, a different risk from the model behaviour safety research addresses.
Salary and market data
Pay for AI safety researchers varies widely between frontier labs, government institutes and non-profit research organisations, and with seniority and research record. Our GenAI and agentic systems salary guide sets out current ranges, the AI and deep tech salary guides cover adjacent research roles, and the deep tech salary calculator lets you test an offer.
Hiring AI Safety Researchers with DeepRec
DeepRec recruits for AI research and safety teams across Europe and the US, from offices in London, Berlin, Zurich, Dublin and Boston. We screen candidates technically, discussing their research questions, experimental design and evaluation methods, and we understand the questions safety researchers ask about mission, research freedom and compute. We hire on a permanent, contract, retained and executive search basis, from research engineers to heads of safety. To start a search, go to Looking to Hire.