A new wave of research into artificial intelligence-driven consumer modelling is raising concerns about the reliability of synthetic populations, with findings suggesting that “more data” may actually reduce accuracy rather than improve it. The study explores how large language models simulate consumer behaviour and why current approaches often break down when too many behavioural and psychological inputs are combined.
Researchers involved in the study examined thousands of AI-generated consumer interviews and model interactions across multiple countries and platforms, comparing synthetic responses with real-world purchasing behaviour. The results highlight a consistent pattern: while AI systems can replicate general attitudes and identity-based preferences with reasonable success, they struggle significantly when asked to mirror real-life habitual decisions and deeply embedded behavioural loyalty.
At the centre of the findings is a structural limitation described as a “domain-dependent” approach to human decision modelling. Instead of treating consumer behaviour as a single unified system, the research shows that human decisions operate across distinct cognitive layers—identity-based expression, deliberative trade-offs, and habitual actions. When these layers are merged indiscriminately within AI prompts, performance deteriorates sharply.
Brandspur Banking News Desk reports that the study identifies this issue as “layer interference,” a phenomenon where conflicting types of input data disrupt the model’s ability to prioritise real behavioural signals. In practical terms, adding personality traits, decision frameworks, and purchase history into the same prompt causes AI systems to prioritise narrative consistency over factual behavioural patterns, leading to distorted predictions.
The research outlines a framework known as the Domain-Dependent Information Hierarchy, which categorises consumer modelling into three core levels. The first level focuses on identity-driven behaviour such as lifestyle choices and brand perception. The second covers deliberative decision-making involving trade-offs, risk, and long-term evaluation. The third and most critical level relates to habitual actions such as routine shopping patterns and repeat purchases.
Findings suggest that each of these levels requires different data structures for accurate simulation. Personality traits and narrative profiles perform well when modelling identity-based behaviour, while structured decision parameters are more effective for trade-off analysis. However, raw behavioural data alone remains the most reliable input for predicting routine consumer actions.
A major conclusion of the study is that combining these layers reduces accuracy significantly. In some cases, prediction performance dropped by double-digit percentages when personality data was added to behavioural datasets. Researchers attribute this to what they describe as a “character simulation effect,” where AI models begin role-playing constructed personas instead of following empirical data.
Another key insight from the research is the concept of an “identity-operation gradient,” which shows that AI systems perform better when simulating behaviour tied to self-perception or narrative identity than when replicating purely operational habits. For example, models more accurately predict aspirational or expressive choices, but struggle with repetitive behaviours such as grocery purchasing frequency or brand loyalty.
The study also highlights a consistent failure in simulating extreme brand loyalty. Even when models are explicitly given loyalty constraints, they tend to default to rational cost-saving decisions, switching brands when presented with minor price differences. Researchers attribute this to a built-in bias toward rational economic behaviour, which conflicts with real-world consumer inertia and emotional attachment.
One of the most striking findings is the difficulty AI systems face in modelling “loss aversion,” a behavioural economics principle describing consumers’ tendency to avoid perceived losses more strongly than they pursue equivalent gains. Across all tested systems, this parameter showed extremely low predictive accuracy, contributing directly to the inability of AI models to replicate rigid consumer habits.
The report further explains that prompt design and data architecture matter far more than model selection. According to the findings, differences in how data is structured account for the vast majority of performance variation, while the choice between advanced AI models contributes only marginal improvement. This shifts focus away from model scaling toward system design and information architecture.
In response, researchers propose the development of a domain-aware cognitive routing system that dynamically selects which type of data should be used depending on the behavioural context being simulated. Such a system would separate identity-driven modelling from behavioural prediction, ensuring that only relevant inputs are used for each task.
The study ultimately argues for a hybrid approach to market research. Synthetic populations, it suggests, are highly effective for early-stage exploration, concept testing, and understanding consumer sentiment. However, they are not yet reliable enough to replace human data when predicting precise behavioural outcomes such as purchase frequency, brand switching, or loyalty intensity.
Instead, the future of consumer research is framed as a layered integration model, where AI is used to rapidly explore possibilities and narrow options, while traditional human-based research validates final behavioural assumptions before market execution.
