AI advice study: AI makes people three times less accurate before correction

A staggering 75% of participants in a recent AI advice study were found to be less accurate in their decision-making when they received AI advice before making a decision, with the study revealing that people are three times less accurate when relying on AI guidance initially. This phenomenon has significant implications for how AI systems should be integrated into workflows, particularly in high-stakes environments like medical diagnosis, financial planning, and legal analysis.

The Study Design and Its Findings

Researchers from multiple universities designed a controlled experiment to test the impact of AI advice on human decision-making accuracy. Participants were asked to make quantitative estimates on various tasks, including counting objects in images and predicting numerical outcomes. The study split participants into two groups: one group received AI advice before making their estimates, while the other made their estimates first and then received AI advice as a check. The results were striking. The group that received AI advice first was three times less accurate in their final estimates compared to the group that made their own estimates first. This finding held true across multiple task types and difficulty levels, suggesting a robust cognitive bias at play.

The Anchoring Effect in AI-Human Interaction

The study's lead researcher attributed the accuracy drop to the well-documented anchoring bias. When humans see an AI-generated estimate before forming their own judgment, they anchor on that number and struggle to adjust sufficiently away from it, even when the AI's advice is clearly wrong. This effect is particularly dangerous because AI systems often present their outputs with high confidence, potentially leading users to override their own better judgment. The researchers noted that participants who made their own estimates first were significantly more likely to identify and correct AI errors, suggesting that the order of information presentation matters as much as the information itself.

What This Means

For AI founders and product builders, these findings carry urgent implications. The default assumption that more AI assistance always leads to better outcomes is empirically false. The study suggests that the optimal design for AI-augmented decision-making may involve presenting AI advice as a secondary check rather than a primary input. This has direct applications in AI copilot tools for coding, AI-assisted medical diagnosis, and AI-powered financial analysis platforms. Products that surface AI suggestions before users form their own judgments may inadvertently degrade decision quality. Designers should consider implementing a "human-first" interaction pattern where users commit to a preliminary decision before receiving AI input. This approach not only preserves human agency but may also produce more accurate outcomes overall. Additionally, the study highlighted that AI systems deployed in high-stakes environments should include uncertainty estimates and confidence intervals to help users calibrate their trust appropriately.

Broader Implications for AI Adoption

The research adds to a growing body of evidence that the human-AI interaction design is as important as the underlying model quality. As AI tools become more integrated into professional workflows, understanding these cognitive dynamics will be critical for building systems that actually improve outcomes rather than simply adding friction. For regulators and industry standards bodies, the study suggests that guidelines for AI deployment should address interaction patterns, not just model performance metrics. The finding that AI can make people less accurate, even when the AI is correct, challenges the narrative that more AI automation is always better. It underscores the need for careful UX research and iterative testing before deploying AI assistants in production environments where decisions have real consequences.

EXCERPT: A new study reveals that receiving AI advice before making a decision makes people three times less accurate due to anchoring bias. The research suggests that AI tools should be designed as secondary checks rather than primary inputs to preserve human judgment and improve decision quality.