Paying AI to detect AI

We embed a ChatGPT-based AI detector in a laboratory experiment to test whether participants are willing to pay more to collaborate with AI than with human peers to accurately detect the proportion of AI-generated parts in deepfake news articles. Task difficulty varies with the model used to generate the articles (GPT-2 vs. GPT-4o). We find that participants' willingness to pay (WTP) for the AI detector exceeds that to collaborate with human peers, even though the AI detector does not provide better assistance and, in fact, humans do better than AI in for GPT-4o generated news. WTP for AI or peer collaboration does not rise with task difficulty. These patterns point to over-reliance on AI and raise concerns about the spread of deepfakes. The study improves understanding of human-AI interaction and informs safeguards for deepfake detection in the era of generative AI (GAI).