Truth probes cannot distinguish true and prescribed actions perfectly
The Truth Was Never Gone: Perfect Aliasing in Compliant-Context Truth Probes
Machine LearningArtificial IntelligenceComputation and Language
Summary
Some tools called truth probes try to detect when a system is telling the truth based on its behavior. This paper finds that if what the system does matches exactly what it’s supposed to do, the probe can’t tell whether the system is truly truthful or just following orders. The authors show that this confusion, called perfect aliasing, happens in many cases and prove ways to separate true beliefs from just prescribed actions. However, they caution that this doesn’t mean the probes measure actual beliefs or that they can detect deception reliably.
What this means in practice
- •For ai model auditors: Evaluate whether probes truly capture model truthfulness or simply reflect task compliance in behavior.
- •For machine learning engineers: Design better interpretation tools by separating true internal understanding from mere output correctness in trained models.
Authors
Dylan Jayabahu
Abstract
A truth probe fitted where truthful reporting and a task's prescribed action coincide cannot distinguish those targets from its fitting labels alone. We call this failure of semantic identification perfect aliasing. In a controlled binary reporting game, truth and prescribed-action probes fitted on compliant contexts solve the same optimization. On rival contexts their labels are complements, forcing their AUROCs to sum to one; this identity holds across 751 cell-layer pairs to floating-point precision. We separate prescribed output symbols from semantic action using randomized codebooks, then separate truth from prescribed action by fitting on mixed compliant and rival contexts. For a reward-trained Gemma-2-9B policy that answers falsely on all evaluated rival trials, the conventional probe scores $0.006 \pm 0.005$ AUROC across three training seeds, while mixed-fit probes score $1.000$ on the same held-out activations. Mixed fitting uses more training examples and access to labelled rival contexts, so this comparison establishes linear recoverability rather than isolating the benefit of decorrelation. We also show that two compliant-fit probes, both perfect in-distribution, score $0.080$ and $0.986$ on the same rival activations. The findings concern what a probe measures: they do not establish preserved functional belief, causal use of the recovered direction, or a deployable deception detector. Code and aggregate results accompany the paper.