Trapped by Expectations: Functional Fixedness in LLM-Enabled Chat Search

Prior experience shapes how people use new AI tools in limiting ways: The study of 450 participants found that users' expectations for ChatGPT were heavily influenced by their previous experience with other systems—search engines, virtual assistants, and ChatGPT itself—causing them to interact with the new tool using familiar but potentially suboptimal patterns rather than exploring its unique capabilities.

Different user backgrounds lead to different interaction styles that may reduce effectiveness: Users experienced with virtual assistants tended to use command-like prompts, search engine users maintained rigid, structured queries with minimal changes, while frequent ChatGPT users were more adaptive. These ingrained habits limited participants' ability to fully leverage the chat interface for complex, exploratory tasks like public safety decisions or AI ethics questions.

When AI systems fail to meet expectations, users become more creative and engaged: The research found that when ChatGPT didn't deliver what participants expected, they responded by generating more detailed prompts with greater linguistic variety, suggesting that "unmet expectations can motivate behavioral adaptation" and potentially unlock better use of the tool.

Functional fixedness represents a significant barrier to responsible AI adoption: The study demonstrates that cognitive biases preventing users from exploring an AI system's full capabilities could lead to suboptimal decision-making in important domains like health management and sustainability, highlighting a gap between what these tools can do and how people actually use them.