Sparse autoencoders gave the field real traction. They did not give it the ability to certify a model as safe, and the distinction matters.
Mechanistic interpretability has had a genuinely good few years. Work on sparse autoencoders and feature decomposition produced something the field had been missing: a method for pulling human-legible structure out of activation space that survives contact with models people actually deploy [Bricken et al., 2023; Templeton et al., 2024]. Attribution graphs then began to show how those features connect, tracing multi-step reasoning and planning inside a production model [Lindsey et al., 2025].
It is worth being precise about what that does and does not establish, because interpretability is increasingly cited as evidence in safety arguments, and evidence has to bear weight.
What the results support
Features recovered by sparse decomposition are frequently interpretable, sometimes strikingly so: specific people and places, programming constructs, and more abstract concepts such as deception or sycophancy. They can be intervened upon; steering behaviour by amplifying or suppressing a feature direction works often enough to be useful. And the features are not arbitrary: similar ones recur across training runs and, to a degree, across architectures.
There is also evidence that interpretability helps with real tasks. In a 2025 auditing exercise, teams were given a model deliberately trained with a hidden objective and asked to find it; teams with access to interpretability tools, alongside behavioural testing and training-data search, were able to uncover it [Marks et al., 2025].
That is a real result. It suggests representations are more structured, and more accessible, than the "inscrutable matrices" framing implied.
What they do not support
Three gaps are load-bearing.
Coverage
Decomposition recovers some features, not all of them. Reconstruction error is not zero, and the residual is not obviously benign: work on the "dark matter" of sparse autoencoders shows that a meaningful share of what a model computes is not captured, and that some of the error has structure of its own [Engels et al., 2024]. There is no argument that the uninterpreted remainder is safety-irrelevant. An explanation of part of a computation is not an explanation of the computation.
Faithfulness
A feature that correlates with a concept and a feature that causes behaviour related to that concept are different claims. Intervention experiments address this, but intervening on a direction and observing an output change does not establish that the direction played the role its label implies during normal operation. Labels are themselves generated by humans or other models reading examples, and can be confidently wrong.
The field has been admirably candid here. In 2025, Google DeepMind's interpretability team reported that sparse autoencoder features underperformed simple linear probes on a downstream safety-relevant task, and said it was deprioritising fundamental SAE research as a result [Smith et al., 2025]. That is not a failure of interpretability. It is interpretability being checked against the jobs people want it to do.
Adversarial pressure
Interpretability results are typically established on natural inputs. A model under adversarial pressure is operating in a regime where the mapping from features to behaviour may simply differ. Very little work tests whether interpretations hold there, and safety cases care about precisely that regime. If a model can recognise when it is being evaluated, the regime we inspect may not be the regime that matters.
Why this matters in practice
If interpretability is offered as evidence in a safety case ("we inspected the model and found no deceptive circuitry"), then coverage and faithfulness are not academic caveats. They are the difference between evidence and reassurance. Finding a mechanism is informative. Failing to find one, with tools of known partial coverage, is weak evidence of absence.
The useful framing
Interpretability is currently a strong tool for generating hypotheses about model behaviour and a weak tool for certifying its absence. Both are valuable; only one is what a regulator or a deployment decision actually needs, and public communication about the field does not always distinguish them.
The researchers doing this work are, in our reading, considerably more careful about these limits than the commentary around it. The overclaiming happens downstream.
Update, September 2026
Two newer methods are worth adding. Anthropic's natural language autoencoders translate activations directly into text and check each explanation by reconstructing the original activation from it; applied to Claude, they surfaced evaluation awareness that the model never stated. Circuit condensation, meanwhile, post-trains models so that specific behaviours run through much smaller circuits, making them easier to inspect. Both move the field forward on faithfulness and tractability. Neither closes the coverage gap, and so neither changes the conclusion above. We cover both in more depth in our piece on chain-of-thought faithfulness and our survey of circuit learning at scale.
References
Bricken, T. et al. (2023). "Towards Monosemanticity: Decomposing Language Models With Dictionary Learning." Transformer Circuits Thread. transformer-circuits.pub.
Engels, J. et al. (2024). "Decomposing The Dark Matter of Sparse Autoencoders." arXiv:2410.14670.
Lindsey, J. et al. (2025). "On the Biology of a Large Language Model." Transformer Circuits Thread. transformer-circuits.pub.
Marks, S. et al. (2025). "Auditing Language Models for Hidden Objectives." arXiv:2503.10965.
Smith, L. et al. (2025). "Negative Results for SAEs On Downstream Tasks and Deprioritising SAE Research." Google DeepMind Mechanistic Interpretability Team, AI Alignment Forum.
Templeton, A. et al. (2024). "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet." Transformer Circuits Thread. transformer-circuits.pub.