top of page

GenAI Consumer Research Risks an Average Trap and Model Collapse

11 minutes ago
2 min read

Image via Wikimedia Commons (public domain).

What happens to consumer research when generative AI becomes a standard tool? Huang and Rust argue in the Journal of Consumer Research that GenAI follows a trajectory from democratization to an "average trap" of generic answers and finally to model collapse, and they set out concrete ways researchers can push back.

From democratization to model collapse

Consumer researchers now use generative AI to draft surveys, code open-ended answers, simulate respondents and analyze text at scale. The tools lower the cost of doing research, which is good news for many scholars. The harder question is what happens to the evidence base when more and more of that research runs through the same type of model. Ming-Hui Huang (National Taiwan University) and Roland T. Rust (University of Maryland) take up that question in a conceptual article rather than an empirical study.

According to Huang and Rust (Journal of Consumer Research, 2025), GenAI moves through three linked phenomena. Democratization comes first: more consumers take part in the data that feed research, which can improve representation but also carries real-world biases into the data. The average trap follows: because these systems are built on next-token prediction, they tend to produce generic outputs that lose the individuality of real consumers. Model collapse is the endpoint: as GenAI learns more and more from its own outputs, machine bias is amplified and the models drift away from how people actually behave.

The authors pair each stage with a remedy. To make good use of democratization, researchers can study marginalized consumers and keep human-centered research ahead of purely data-driven methods. To escape the average trap, they suggest fine-tuning models with task-specific data and data from marginalized consumption, and engineering responses for uniqueness rather than typicality. To prevent model collapse, they call for human and machine hybrid data and for applying theories of mind to realign AI with human-centric consumption.

The article closes with a research agenda in three parts: preserving the tails of data distributions so that democratization does not erase minority consumers, countering the average trap built into next-token prediction, and finding ways to reverse the slide from democratization toward model collapse.

Readers should keep the genre in mind. This is a theoretical framework, not a test of it. The paper does not measure how often synthetic respondents are generic or how fast collapse happens in practice. Its value lies in giving researchers a vocabulary and a set of testable directions, which later empirical work will need to confirm, refine or overturn.

Why it matters: If you use LLMs to simulate participants, generate stimuli or code consumer text, the "average trap" is a concrete validity threat: models pull answers toward the typical case and flatten the variation that consumer behavior research exists to explain. Huang and Rust's framework gives reviewers and authors a shared checklist, namely whether a study keeps real human data in the loop, whether it protects the tails of the distribution, and whether it relies on models trained on their own outputs. That is worth asking of any GenAI-assisted study before it reaches print.

Huang, M.-H., & Rust, R. T. (2025). The GenAI Future of Consumer Research. *Journal of Consumer Research*, 52(1), 4-17. Read the original. Independent editorial summary of third-party research.

Comments


This initiative is supported by the following organizations:

  • Twitter
  • LinkedIn
  • YouTube
logo_edited.png
bottom of page