Why AI is like a (Clever Hans) Horse - Computerphile

To answer why a model made a decision you need causal analysis because only interventions that test which input features change the output can separate true causes from mere correlations.

Adding hundreds of minimal sufficient signals still triggers the same class because each contains spectral cues the model associates with the label, and their aggregate preserves enough of those cues even though the blend sounds nothing like the genre to humans.

Models can be highly confident about labels humans would reject because training makes them output strong posterior scores whenever they see learned correlations, so confidence reflects dataset patterns not human semantics.

Many music models prioritize spectral fingerprints like timbre or production artifacts because frequency analysis gives consistent cues that correlate with labels, so they often ignore rhythm and harmony humans use to identify genres.

A complete explanation for a classification must be both sufficient and necessary because the feature set should by itself produce the class and removing it must prevent the model from producing that class, proving a causal role.

A divide‑and‑conquer search finds minimal frequency sets efficiently by randomly partitioning bins, testing chunks, discarding irrelevant parts, and recursively subdividing relevant ones to narrow down causes without checking every combination.

Brute‑forcing which frequency bins matter is infeasible because a 30‑second clip produces on the order of hundreds of thousands of bins and the possible combinations explode combinatorially.

An audio waveform can be decomposed into sinusoids because a Fourier transform converts the time signal into frequency components that reveal which frequencies and amplitudes make up the sound.