Rating the Pitch, Not the Product: User Evaluations of LLMs Reflect Expectations More Than Performance

2026-07-06

AI Summary

This study investigates whether users’ evaluations of large language models (LLMs) reflect actual performance or are driven by prior expectations. Through a controlled experiment in a multitask human–AI collaboration setting, the authors manipulated participants’ pre-use beliefs about model capabilities and employed structured questionnaires with regression analysis to disentangle the effects of expectation, confidence, and objective performance on subjective ratings. Data from 162 participants reveal that task output quality depends solely on the model’s true capability, whereas user ratings are significantly influenced by induced expectations (β = 0.47–0.50, p < 0.001) and unrelated to objective performance. These findings uncover an “expectation-fulfillment bias” in LLM evaluation, challenging prevailing paradigms that prioritize user preference as a proxy for model effectiveness.

Abstract

Imagine two users interact with the same LLM. One has been told it is the cutting-edge flagship model; the other, an older, weaker model. They walk away with markedly different ratings of its usefulness and intelligence, yet they used the same model. In a controlled study, 162 participants each used one of six LLMs from two families across three collaborative tasks, after first viewing a landing page that matched, overstated, or understated their model's true capability. This pre-interaction framing shifted user opinions and interaction behavior while task performance did not. Oversold users rated the model more favorably and used more directive prompting, while Undersold users wrote longer, more collaborative prompts. The quality of what users and the model produced together depended only on the model's true capability, not on what users were told. Participants' change in model impressions after use, measured across two impression measures, was not predicted by task performance ($β= -0.01$ and $0.11$, both n.s.), but by whether the model met users' expectations ($β= 0.47$ and $0.50$, both $p < .001$) and how confident they felt working with it ($β= 0.47$ and $0.36$, both $p < .001$). After interaction, users are still rating the pitch, not the product: user-elicited LLM evaluations, including the preference data driving public leaderboards, measure expectation management at least as much as the model itself.

Problem

• large language models
• user evaluation
• expectation bias
• human-AI interaction
• model performance

Innovation

• expectation bias
• user evaluation
• large language models
• human-AI interaction
• performance vs perception