Mechanistic interpretability is a subfield of AI safety and research that aims to reverse-engineer trained neural networks at the level of individual weights, neurons, and circuits to explain precisely what computations they perform. Rather than treating a model as a black box and measuring its input-output behaviour, mechanistic interpretability dissects the internal representations — identifying which neurons detect specific features, how those features combine into higher-level circuits, and how information flows through layers to produce a given output. Key results in this area have identified induction heads in Transformers, curve detectors in vision models, and arithmetic circuits in small language models.

Think of it like performing a full autopsy on a calculator to figure out not just that it adds numbers correctly, but exactly which transistors implement each logical gate and why the design works.

A classic mechanistic finding concerns "indirect object identification" in GPT-2: researchers traced which Attention heads copy the indirect object token (e.g. "Mary" in "John gave Mary the book") to the output position, producing a circuit diagram of the behaviour:

Layer 3, Head 0  -->  copies subject token
Layer 9, Head 6  -->  inhibits incorrect completions
Layer 9, Head 9  -->  copies indirect object to output

Such diagrams let researchers predict and verify model behaviour by patching or ablating individual components. Mechanistic interpretability is closely connected to Embedding (the representations being decoded), Attention (the dominant circuit type in Transformers), and the study of Hallucination (tracing where factual errors originate in the network).

You will most often encounter mechanistic interpretability in AI safety research, in papers from Anthropic's interpretability team and Neel Nanda's work, and increasingly in practical debugging efforts when a model produces systematic errors that simple prompt-level fixes cannot resolve.