Can natural language explanations redefine what interpretability means?

Interpretable ML grew around inherently-interpretable models (sparse linear, GAMs, trees) and post-hoc techniques (feature importance, visualization, distillation). This position paper argues LLMs change the game: their capacity to explain in natural language expands the scale and complexity of patterns that can be conveyed to a human, so interpretability can be redefined with a far more ambitious scope — including using LLMs to audit LLMs themselves. The cost is new failure modes: hallucinated explanations and immense compute. The two priorities it names — using LLMs to directly analyze new datasets (knowledge discovery) and to generate interactive explanations — reframe interpretability from "inspect the model" toward "converse about the model and the data."

The keeper is the reframing of the medium: explanation in natural language is not just a nicer output format but a capacity expansion — it lets humans receive more complex patterns than feature attributions or saliency maps can carry. The catch is that this is exactly where faithfulness risk concentrates.

This sits at the optimistic pole of the vault's explanation thread, and it must be read against the cautionary results: Can LLM explanations actually help humans predict model behavior? shows NL explanations can feel trustworthy without being predictive, and Can cognitive science methods unlock how LLMs actually work? supplies the structure this ambition needs to avoid collapsing into plausible-but-unfaithful narration.