Human Decisions and Artificial Intelligence

GAICIL Insights

Here is the awkward fact at the center of medical AI: in this study, the machine alone beat the doctors, with or without its help.¹ Yet ethical and legislative guidelines do not permit AI to make diagnostic decisions directly. Human Decisions and Artificial Intelligence, by Zimeng Li, Edoardo Gallo, Jingyuan Wang, Ke Tang and Han Hong, takes that constraint as given and asks the practical question it leaves: how should AI talk to the humans who remain responsible? The authors run an experiment in which doctors see an AI system's diagnoses and get chances to revise their own decisions.

The answers are specific enough to build products on. Explanations accompanying the AI's diagnosis help doctors improve. Explanations also mitigate diagnostic expectation, the tendency to overweight the probability of a positive case. And doctors suffer choice overload: they are more likely to follow the AI's suggestion when shown 3 explanatory features from its model than when shown 10 or 30. Inexperienced and impatient doctors rely more on AI's assistance.

Sit with the choice-overload result, because it upends a favorite assumption of AI governance. More transparency, measured as more features displayed, made assistance less effective. Doctors drowning in explanatory detail were less likely to follow advice that, in the study, was more accurate than their own. The interface is the bottleneck. When humans must stay in the loop, the design of that loop decides how much of the machine's accuracy ever reaches a patient.

Call it the last-meter problem of AI. A model can be superb in the lab and still deliver little at the bedside if the handoff to a human is clumsy. A poorly designed interface leaves performance on the table. A well-designed one captures more of it. The value of medical AI is partly an exercise in communication design.

That has sharp implications for regulation. Mandatory explanations are common in AI governance proposals. This experiment shows the form of explanation matters enormously. A rule requiring disclosure of every feature could overwhelm users and reduce reliance on accurate advice. A small number of salient features may improve decisions. Regulation that specifies outcomes, such as improved decision accuracy, leaves designers room to discover what works. Prescriptive disclosure formats risk freezing the least effective version into law.

The reliance pattern adds a second angle. Inexperienced and impatient doctors leaned on the AI more. For junior practitioners, AI assistance may work like on-the-job support, narrowing performance gaps. That points to AI as a tool for spreading expertise, especially where experienced specialists are scarce. It also raises a calibration question worth studying: do users follow AI when it is right and override it when it is wrong? Reliance is valuable only when it tracks accuracy.

The mitigation of diagnostic expectation shows something hopeful too. Well-designed assistance can correct systematic human biases. An AI that nudges clinicians away from overweighting positive cases offers a second opinion that can also debias the first.

The authors call for research combining AI and behavioral science expertise to make explanatory information from algorithms more effective. They show such information can raise diagnostic accuracy and suggest future work on patient outcomes, for instance helping patients accept AI-assisted diagnoses. That agenda models responsible innovation: rigorous experiments that measure how humans and AI interact, followed by designs that improve results.

The lesson travels far beyond medicine. Human-AI collaboration is an engineering and behavioral problem with measurable solutions. Simple design choices, such as limiting the number of explanatory features, shift behavior meaningfully. As AI enters law, finance and public administration, experiments like this should guide both product design and the rules that govern it.

The provocative bottom line: in a world that keeps humans in charge, the most underrated AI research may concern a screen layout. Getting the interface right could be one of the largest and least appreciated sources of gains from AI.