Alignment researchers work on getting AI systems to pursue the goals their developers intend, and to keep doing so as those systems become more capable. In practice that means studying how training signals shape model behaviour, where those signals break down, and which methods can close the gap between what a model was rewarded for and what people actually wanted. The role is also advertised as AI alignment researcher, alignment scientist or research scientist in alignment.
Companies hire for it when they train or fine-tune their own models and need people who can reason about failure modes that ordinary testing misses. Frontier labs hire alignment researchers to shape post-training. Smaller teams building agents on foundation models hire them because autonomy raises the cost of a model that follows instructions literally rather than sensibly. Most of these hires sit inside wider AI research teams.
What does an alignment researcher do?
The work is empirical research with a strong conceptual streak. An alignment researcher forms hypotheses about why a model behaves a certain way, designs experiments to test them at whatever scale the compute budget allows, and turns the results into training changes, evaluation methods or papers.
- Designing and running experiments on reinforcement learning from human feedback (RLHF), direct preference optimisation and other preference learning methods
- Studying reward hacking, specification gaming and sycophancy, then building training interventions that reduce them
- Developing scalable oversight methods such as debate, recursive reward modelling and AI-assisted critique, so humans can supervise tasks they cannot easily judge alone
- Building model organisms of misalignment, controlled set-ups that reproduce behaviours like deceptive reasoning or goal misgeneralisation so they can be studied safely
- Writing the constitutions, specifications and rubrics that feed AI feedback pipelines
- Writing internal research notes and external papers, and presenting results to the people who decide what ships
Key skills and backgrounds
Most alignment researchers hold a PhD in machine learning, computer science, statistics, physics or another quantitative field, although strong research engineers without a doctorate move into the role regularly. What separates good candidates is research taste, meaning the ability to pick experiments whose results would change someone's mind, combined with the engineering skill to run those experiments on large models without weeks lost to debugging.
Hiring managers usually want hands-on experience with large-scale training rather than theory alone. The strongest signal is a record of careful, reproducible empirical work.
- PyTorch or JAX, plus distributed training and fine-tuning across many accelerators
- Reinforcement learning, policy gradient methods and reward or preference modelling
- Experimental design and statistics, including how to avoid over-reading noisy evaluation results
- Evaluation harnesses for honesty, harmlessness and instruction following
- Working knowledge of interpretability tools, useful for checking whether a behavioural fix changes what the model is computing internally
- Clear technical writing for research and non-research readers alike
Where alignment researchers work
The role is concentrated in frontier model developers and the research divisions of large technology firms, where post-training is a core activity. It also appears in non-profit research institutes and academic groups funded to study safety, in government-backed AI safety bodies that need in-house technical capability, and in start-ups building agentic products that want alignment expertise close to their fine-tuning work.
Alignment Researcher vs related roles
Alignment overlaps with several neighbouring research areas, and job adverts often blur the boundaries. The clearest way to separate them is by the question each role is trying to answer.
- An Interpretability Researcher studies the internal mechanisms of a model, while an alignment researcher is mainly concerned with shaping and verifying its behaviour through training.
- A Model Evaluation Researcher builds the tests that measure what a model can and will do, whereas an alignment researcher designs the training methods those tests are checking.
- A Privacy-Preserving ML Researcher protects training data through methods like differential privacy and federated learning, which is a separate problem from making model objectives match human intent.
Salary and market data
Pay for alignment research reflects competition between labs, start-ups and research institutes for a small pool of people with both research depth and large-model experience. Our GenAI and agentic systems salary guide and the wider AI and deep tech salary guides cover current ranges, and the deep tech salary calculator lets you benchmark a specific offer.
Hiring alignment researchers with DeepRec
DeepRec recruits alignment researchers for labs, scale-ups and enterprise AI teams across Europe and the US, working from offices in London, Berlin, Zurich, Dublin and Boston. Our consultants screen technically, which for this role means reading candidates' papers, asking how they chose and validated their experiments, and testing their real experience with large-scale training. We handle permanent, contract, retained and executive search briefs, including alignment research leads. If you are building or growing an alignment team, talk to us through Looking to Hire.