Activation-based deception detectors for large language models can catch lies that read as honest, but their usefulness hinges on how much trust their alarms deserve, according to a new arXiv white paper. The authors compare such systems to polygraphs: a detector that flags deceptive responses is only valuable if its warnings are reliable.

The abstract argues that accuracy is not enough. It frames uncertainty as a key issue, suggesting that researchers should evaluate not only whether a detector catches deception but also how confident users can be in its alarms. The paper's full details on what current detectors depend on are not included in the available abstract.

Since this is a single source, there are no independent findings to compare or contrast. The white paper's contribution, based on the abstract, is to shift attention from raw detection performance to the trustworthiness of the detector's signals.