A new study shows that the interpretability probes AI labs use to catch models lying may be reading the wrong signal entirely.
In a preprint posted to arXiv on September 17 titled "Decodability is Not Causality," researchers built a diagnostic that breaks a deployed truth/deception probe down into individual sparse-autoencoder features, then ranks each one by two criteria: how closely it aligns with the probe's geometry, and how much it actually moves model behavior when ablated. Testing the TTPD truth probe from Buerger et al. (2024), applied to Gemma2-9B-Instruct under the instructed truth/deception setup from Long et al. (2025), the two rankings barely agreed - about 12% overlap, with a Spearman correlation of just 0.10. When the team ablated features, the ones the probe actually shared with the model's real behavior flipped outputs up to 27% of the time, versus 6% for features that only looked important to the probe geometrically, and 1% for random features.
That gap held up across five random seeds and a held-out data split, which rules out a fluke. Combining probe direction with activation statistics recovered far more of the causally relevant features - a 17.6% flip rate versus 6.1% for the probe's raw top-ranked features - suggesting today's simpler probing methods can give safety teams a false sense of precision.
If a lie-detector for language models can't tell you which wires it's actually reading, it's worth asking how many other interpretability results are measuring the map instead of the territory.