Researchers developed an adaptive AI system that adjusts guidance to individual users, improving decision accuracy while reducing overreliance and encouraging stronger human judgement skills.
Researchers at Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS) have developed an adaptive artificial intelligence model that helps people make better decisions without encouraging excessive reliance on AI. The system, based on reinforcement learning, dynamically adjusts the amount of assistance it provides according to a user’s needs, aiming to improve both decision accuracy and long-term learning.
The researchers argue that many current AI assistants offer uniform recommendations regardless of the user’s expertise or confidence. This can lead to overreliance, where individuals accept incorrect AI suggestions instead of applying their own judgement, gradually weakening their decision-making skills over time.
To address this issue, the new model continuously learns from user interactions and decides how much guidance to provide in each situation. Rather than always supplying complete answers, it can offer partial explanations, full recommendations or, when appropriate, withhold advice altogether. The approach considers factors such as a person’s skill level, confidence and need for analytical thinking before determining the most effective form of support.
The research team evaluated the system through online experiments involving more than 1,000 participants using healthcare-related decision scenarios. Participants interacting with the adaptive reinforcement learning model consistently achieved higher decision accuracy than those using conventional AI systems that always delivered fixed recommendations. In many cases, the combined human-AI team outperformed both unaided participants and the AI system operating independently.
Researchers believe the findings have important implications for the design and regulation of high-stakes AI applications, including healthcare, law and public services. Instead of assuming human oversight alone is sufficient, they suggest AI systems should be intentionally designed to strengthen human capability while reducing dependence. The study concludes that adaptive decision support can improve performance without diminishing critical thinking, creating AI tools that enhance rather than replace human expertise.
