Xuesen Fan

Ohio State University · Agricultural, Environmental and Development Economics · 2026–27 job market

Fields

Agricultural, Applied Micro, Behavioral & Experimental, Public, Econometrics, Environmental & Energy, Innovation

Job market paper

Can AI-Generated Cheap Talk Correct Hypothetical Bias? Evidence from a Food Choice Experiment

To help researchers and policymakers obtain valid estimates of consumer willingness to pay (WTP) for food attributes that markets do not yet price, I study whether a cheap talk script generated by a large language model can mitigate hypothetical bias in discrete choice experiments more effectively than the conventional script that has dominated practice for two decades. Cheap talk is the most widely used device for reducing hypothetical bias, yet its effectiveness is inconsistent, in part because every researcher writes their own script, making the script itself an uncontrolled source of variation across studies. I field an online food choice experiment on certification-labeled snakehead fish fillets with Chinese consumers, randomly assigning respondents to three treatments: a classic cheap talk script, an AI-generated (ChatGPT) script used without human curation, and a real-incentive benchmark in which one choice set is binding. The results suggest that classic cheap talk leaves WTP overstated by roughly 46 to 52 percent relative to the benchmark, whereas AI cheap talk brings WTP close to the real-incentive level. The adjustment operates primarily through restored price sensitivity rather than through changes in attribute valuation, and a probe of an algorithm-appreciation mechanism finds no support. Because a fixed prompt produces a reproducible script at near-zero cost, AI cheap talk offers a standardized, low-cost alternative to classic cheap talk; I release the generating prompt and estimation code. These findings will inform researchers designing stated-preference studies and policymakers who rely on WTP estimates to evaluate food labeling and certification programs.