Adversarial Attacks Reveal Divergent Deception Competence in Humans and Generative Language Models

arXiv:2610.13363v1 Announce Type: new 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.

Study reveals both strengths and weaknesses of humans and language models in deception detection.

The study demonstrates that generative language models are highly effective at making lies sound truthful, exceeding human capability. This creates a significant risk by making it easier to spread falsehoods undetected, thereby undermining trust and information integrity. The models' inability to mimic human deception cues is a 'two-sided' failure, meaning their lie detection capabilities may also be compromised.

The study shows that language models can make lies sound truthful, which could facilitate deception. This has concerning implications for trust in information.

SOURCE     arXiv cs.CY (Computers and Society) (tier 1)
   ↓
DOCUMENT   5b59a93d-2266-48c2-b434-9ba89e1ad332
           https://arxiv.org/abs/2610.13363
   ↓
EVIDENCE   2 extracted excerpts
   ↓
MODEL RUN  4 runs, methodology 1.0.1
   ↓
SCORE      -55  (Adverse)
   ↓
CONFIDENCE 80%