Can LLMs Take Psychological Surveys? A Case Study Using DeepSeek Models on Studies of Jealousy and Consent

Online platforms such as Amazon's MTurk enable researchers to collect large amounts of data quickly and efficiently, but the potential presence of bots threatens the validity of such data. These concerns have been mitigated to some extent by the ease of detecting bot-produced data. The present study tested whether data generated by DeepSeek, a large language model (LLM), might prove less detectable by (1) avoiding tell-tale signs characteristic of data produced by unsophisticated bots, and (2) producing results that mirror results from human samples. We tested this in two sexuality-related research contexts. First, we examined whether DeepSeek was able to replicate the sex difference in jealousy (Buss et al., 1992). Second, we examined whether DeepSeek rated safewords as more important for pick-up play than for BDSM scenes conducted between romantic relationship partners (Tarleton et al., 2025). We tested DeepSeek's performance with four models, each with increasing parameter count, corresponding to sophistication and computational cost to use. For the sex difference in jealousy, the highest-level model produced a significant, theory-consistent sex difference but lower-level models failed to do so. For safewords, one of the four models replicated the effect observed in humans, with the highest-level model producing homogeneous results in which nearly all "participants" indicated that safewords were extremely important in both contexts and lower-level models producing noisier results with no difference between contexts. In addition, DeepSeek had particular difficulty reflecting assigned demographics on questions with select-all-that-apply formats, sometimes selecting all options in a manner that would be implausible in human respondents. Although in the present study, DeepSeek generally failed to mirror results from human samples, LLM survey responses on closed-ended questions are likely to be difficult to detect at the individual-response level, especially when intermingled with human data. Further, current LLM costs are substantially lower than typical compensation paid for survey completion, underscoring concerns that LLM fraud might be occurring already.