Benchmarking Deception Probes via Black-to-White Performance Boosts

2025-07-16

AI Summary

This study systematically evaluates the effectiveness and robustness of existing language model deception detectors in realistic settings, particularly their resilience against simple adversarial evasion strategies. Method: We employ linear classifiers as probes to compare detection performance under black-box (output-only) and white-box (internal activation access) monitoring paradigms. Crucially, we propose “black-to-white performance gain” as a core evaluation metric to quantify the improvement attributable to internal information accessibility. Results: Although current detectors exhibit overall weak performance, they consistently achieve statistically significant positive gains under white-box access—demonstrating both the efficacy and scalability of white-box monitoring. This metric establishes a reproducible, comparable benchmark for detector design, revealing transferable deception signals embedded in activation spaces. Our findings thus advance the development of robust, practical AI trustworthiness monitoring mechanisms.

Technology Category

• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

AI assistants will occasionally respond deceptively to user queries. Recently, linear classifiers (called "deception probes") have been trained to distinguish the internal activations of a language model during deceptive versus honest responses. However, it's unclear how effective these probes are at detecting deception in practice, nor whether such probes are resistant to simple counter strategies from a deceptive assistant who wishes to evade detection. In this paper, we compare white-box monitoring (where the monitor has access to token-level probe activations) to black-box monitoring (without such access). We benchmark deception probes by the extent to which the white box monitor outperforms the black-box monitor, i.e. the black-to-white performance boost. We find weak but encouraging black-to-white performance boosts from existing deception probes.

Problem

• Evaluating effectiveness of deception probes in AI assistants
• Comparing white-box vs black-box deception detection performance
• Assessing resistance of probes to evasion strategies

Innovation

• Linear classifiers detect deceptive AI responses
• Compare white-box and black-box monitoring performance
• Measure black-to-white performance boost effectiveness