Interpretability researchers try to work out what a neural network is actually computing. Rather than treating a model as a black box and judging it only by its outputs, they study its weights and activations to identify the features it represents and the circuits that combine them into behaviour. The role is also called mechanistic interpretability researcher, interpretability scientist or, in more applied settings, model understanding researcher.
Organisations hire for it because behavioural testing alone cannot rule out a model doing the right thing for the wrong reasons. If a team can see the internal mechanism behind a refusal, a factual recall or a deceptive answer, it can debug models more precisely, audit safety training and catch problems that would never show up in an evaluation set. That is why interpretability now sits alongside alignment and evaluation in many AI research organisations.
What does an interpretability researcher do?
The job combines careful experimental science with a lot of custom tooling. An interpretability researcher picks a behaviour, forms a hypothesis about how the model implements it, then uses causal interventions to test that hypothesis rather than relying on correlation. Good work ends with an explanation that predicts what will happen when you change the model.
- Training sparse autoencoders, transcoders and other dictionary learning methods to decompose activations into more interpretable features
- Tracing circuits through attention heads and MLP layers using activation patching, path patching and attribution methods
- Investigating superposition and polysemanticity, and how models pack many features into fewer dimensions
- Building linear probes and steering vectors to locate concepts such as truthfulness or refusal, and to test whether editing them changes behaviour
- Applying interpretability to safety questions, for example detecting hidden objectives in deliberately misaligned model organisms
- Building internal tooling and visualisation interfaces that let other researchers explore features at scale
Key skills and backgrounds
Strong candidates usually have a PhD or equivalent research experience in machine learning, neuroscience, physics, mathematics or computer science. Computational neuroscientists often transfer well, because the habit of probing a complex system with targeted interventions carries over directly.
Beyond research ability, the role is engineering heavy. Candidates need to extract, cache and analyse activations from large models efficiently, and to know when an apparent finding is an artefact of their method.
- PyTorch, plus interpretability libraries for hooking into transformer internals
- Linear algebra and an intuition for high-dimensional geometry
- Causal intervention methods, including ablations, patching and counterfactual inputs
- Dictionary learning and sparse coding techniques
- Experience with transformer architectures at the level of individual components
- Data visualisation and interactive tooling for exploring model internals
- Scepticism about their own results, shown through sanity checks and baselines
Where interpretability researchers work
Most dedicated interpretability teams sit in frontier model developers and the research labs of large technology firms, where access to model weights and substantial compute makes the work possible. Non-profit research organisations and university groups run interpretability programmes too, often publishing openly. Increasingly, start-ups are forming around interpretability tooling and model auditing, and some regulated enterprises hire interpretability expertise to support explainability obligations for high-stakes models.
Interpretability Researcher vs related roles
Interpretability supports several other research functions, and candidates frequently move between them. The distinction lies in whether the work looks inside the model or measures it from outside.
- A Model Evaluation Researcher measures model capabilities and propensities through external tests, while an interpretability researcher examines the internal mechanisms that produce those results.
- A Frontier AI Risk Researcher assesses how advanced systems could cause serious harm at a societal level, whereas an interpretability researcher produces the technical evidence about model internals that such assessments can draw on.
- A Privacy-Preserving ML Researcher works on stopping models from leaking training data, a goal that sometimes uses interpretability methods but is focused on data protection.
Salary and market data
Interpretability is a specialised research area with a limited number of experienced practitioners, so compensation tends to follow research scientist bands at well-funded labs. See our GenAI and agentic systems salary guide and the full set of AI and deep tech salary guides for current ranges, or use the deep tech salary calculator to check a package.
Hiring interpretability researchers with DeepRec
DeepRec helps labs, start-ups and enterprise AI teams hire interpretability researchers and research engineers across Europe and the US, through our offices in London, Berlin, Zurich, Dublin and Boston. Our technical screening looks at the substance of a candidate's work, including the methods they used, how they validated causal claims and what they built to make their research possible. We run permanent, contract, retained and executive search assignments, from individual researchers to the person who will lead a new interpretability team. To start a search, visit Looking to Hire.