Difficulties with Evaluating a Deception Detector for AIs

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Smith, Lewis, Chughtai, Bilal, Nanda, Neel
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866914202411597824
author Smith, Lewis
Chughtai, Bilal
Nanda, Neel
author_facet Smith, Lewis
Chughtai, Bilal
Nanda, Neel
contents Building reliable deception detectors for AI systems -- methods that could predict when an AI system is being strategically deceptive without necessarily requiring behavioural evidence -- would be valuable in mitigating risks from advanced AI systems. But evaluating the reliability and efficacy of a proposed deception detector requires examples that we can confidently label as either deceptive or honest. We argue that we currently lack the necessary examples and further identify several concrete obstacles in collecting them. We provide evidence from conceptual arguments, analysis of existing empirical works, and analysis of novel illustrative case studies. We also discuss the potential of several proposed empirical workarounds to these problems and argue that while they seem valuable, they also seem insufficient alone. Progress on deception detection likely requires further consideration of these problems.
format Preprint
id arxiv_https___arxiv_org_abs_2511_22662
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Difficulties with Evaluating a Deception Detector for AIs
Smith, Lewis
Chughtai, Bilal
Nanda, Neel
Machine Learning
Building reliable deception detectors for AI systems -- methods that could predict when an AI system is being strategically deceptive without necessarily requiring behavioural evidence -- would be valuable in mitigating risks from advanced AI systems. But evaluating the reliability and efficacy of a proposed deception detector requires examples that we can confidently label as either deceptive or honest. We argue that we currently lack the necessary examples and further identify several concrete obstacles in collecting them. We provide evidence from conceptual arguments, analysis of existing empirical works, and analysis of novel illustrative case studies. We also discuss the potential of several proposed empirical workarounds to these problems and argue that while they seem valuable, they also seem insufficient alone. Progress on deception detection likely requires further consideration of these problems.
title Difficulties with Evaluating a Deception Detector for AIs
topic Machine Learning
url https://arxiv.org/abs/2511.22662