Cheap talk is among the most commonly applied strategies for reducing hypothetical bias in
discrete choice experiments, but evidence on its effectiveness is mixed. This study tests whether
a cheap talk script generated by artificial intelligence can outperform conventional cheap talk.
Using an online food choice experiment, we randomly assign respondents to one of three
treatments: conventional cheap talk, an AI-generated cheap talk script, and a real-payment
benchmark. Conventional cheap talk reduces hypothetical bias but leaves willingness-to-pay
estimates 46–52 percent above the benchmark. The AI-generated script, by contrast, produces
estimates close to the benchmark. These results suggest that AI-generated cheap talk may
provide an alternative and more effective approach to mitigating hypothetical bias with relatively
low cost on the researchers’ time.
