Large language models (LLMs) such as ChatGPT, Gemini and Grok can generate synthetic consumer responses that closely resemble those of real people, potentially making marketing research faster and less expensive, according to a new study. However, researchers caution that AI-generated responses still fall short of fully replacing human participants.
The study, Synthetic Consumer Insight Generation with Large Language Models, examined whether AI could produce realistic consumer insights for marketing research, where collecting data from people is often “costly, time-consuming, and difficult to scale.”
The researchers found that AI-generated responses and human responses shared many common themes.
“The results show substantial overlap between human and LLM responses in broad topics and associations, but also important differences in style, linguistic structure, and the way diversity is generated.”
The study focused on whether AI could perform “projective techniques”—marketing research methods designed to uncover consumers’ hidden emotions, associations and motivations through open-ended questions rather than direct surveys.
The researchers said projective techniques present a particularly difficult challenge because AI must generate “associations, emotions, symbolic meanings, and creative interpretations” instead of simply answering structured questions.
To test this, the team compared responses from 173 human participants with answers generated by several AI models, including ChatGPT 3.5, ChatGPT 4.1, ChatGPT 5.1, Gemini 2.5 Flash, Grok 4.1 and Mistral Medium 3.1. Participants and AI models completed the same tourism-related exercises involving five U.S. cities, including word association and sentence-completion tasks.
One of the study’s biggest findings was that the way AI is prompted makes a significant difference.
The researchers found that richer prompts and higher temperature settings produced more varied and creative responses.
“Higher temperatures give longer, more diverse responses, with a smaller proportion of ‘standard’ stop words.”
Similarly, the study found that more detailed prompts substantially improved the richness of AI-generated responses.
“The extended prompt has significant positive relationships with the number of words, the vocabulary size, and the word length.”
However, the researchers also discovered that AI tends to be much more verbose than humans.
“Across all three projective tasks, the human responses had the lowest, or close to the lowest, average number of words, indicating that the LLMs often generated more verbose responses than human participants.”
While longer responses may be useful for brainstorming, the study noted that they are “less suitable when the goal is to reproduce the brevity and succinctness of human projective responses.”
The researchers also found important differences among AI models.
Compared with Gemini 2.5, ChatGPT 5.1 and Mistral Medium 3.1 produced much longer responses and larger vocabularies, while ChatGPT 3.5 generated significantly shorter responses with a smaller vocabulary.
Beyond language style, AI and humans often discussed similar themes but differed in the details.
“There is a high concordance between the topics generated from the human responses and from the LLM responses.”
However, the study found that AI often created diversity by inventing elaborate compound adjectives rather than introducing genuinely new ideas.
“The word diversity in the LLM solution seems to be generated by complex compound adjectives… rather than true diversity of topics.”
The researchers conclude that AI can become a valuable tool for marketers, particularly during the early stages of consumer research, helping generate ideas, identify themes and reduce research costs. But they stress that synthetic consumer data should complement—not replace—real consumer research because important differences in creativity, language and interpretation remain.
