Humans and AI differ in fooling deception detectors
Adversarial Attacks Reveal Divergent Deception Competence in Humans and Generative Language Models
Computers and Society
Summary
Detecting lies is usually easier for AI than for humans, but AI can be tricked with clever changes to input text. The authors compared how humans and AI rewrite truthful or lying sentences to fool an automated lie detector. They found humans are better at making truthful sentences seem dishonest, while AI is better at making lies sound truthful. This difference happens because humans focus on specific signs of lying, while AI makes broader changes that generally appear honest. The work shows both strengths and weaknesses in human and AI deception methods.
What this means in practice
- •For security software teams: Design stronger deception detectors by accounting for how humans and AI differently manipulate statements to evade detection.
- •For marketing content creators: Create more believable messages by understanding how AI-generated changes make false statements sound truthful.
Authors
Lucca Pfründer, Riccardo Loconte, Bruno Verschuere, Bennett Kleinberg
Abstract
Machine learning models typically outperform humans in deception detection. A threat to machine learning models is their vulnerability to adversarial attacks: subtle yet targeted perturbations of input data that trick a model into misclassification. Yet, this also provides an opportunity: understanding how humans versus machines construct attacks offers new insights into second-order beliefs about deception. In a preregistered experiment (n = 501), we compare the effectiveness, efficiency, and nature of a paraphrasing attack by humans and generative language models on an automated deception classifier. We find that humans were more effective attackers when rewriting truthful statements to be deceptive, while the generative language model was more effective and efficient at making lies sound like truths. Computational text analysis suggests that this is because humans make targeted use of deception cues, while generative language models, instead, apply generic changes that signal truth. The failure to mimic deception is two-sided: while it exposes a blind spot in the machines increasingly used to detect lies, it also means that language models might facilitate successful lying.