Combining human judgment with AI does not automatically produce better decisions. Research described by INSEAD shows that when people receive an AI recommendation, they may pay less attention to other available information and move their judgment closer to the system. Leaders therefore need to design not only model quality but also the sequence in which the human and AI enter the decision-making process.
Companies often base AI adoption on a simple assumption: the algorithm processes a large amount of data and the person adds experience and context to its output. The result should combine the strengths of both. In reality, however, the mere presence of an AI recommendation can change how a person uses the other information available.
INSEAD summarizes research into radiologists’ decision-making when assessing X-rays with different access to AI predictions and to patients’ clinical histories. When they had an AI recommendation, they remained active but paid less attention to the clinical context, and their conclusions moved closer to the system’s prediction. The authors interpret this as a shift in attention rather than a simple reduction in work effort.
For business decision-making, this mechanism matters more than the specific medical setting. A manager may have high-quality data about a customer, market or operation, but if the first thing they see is a persuasively worded AI recommendation, they may begin to judge the other information subconsciously in relation to the already proposed solution.
The design of the process can therefore be just as important as the accuracy of the model. For significant decisions, for example, a person can first be asked to formulate their own conclusion and the reasons for it, and only then be shown the system’s recommendation. The difference between the two positions then becomes the subject of review. This creates a moment in which the person has to genuinely compare two views instead of automatically accepting the first one.
The same INSEAD article recalls additional research on group decision-making. Discussion among people helps them calibrate their own view of expertise more accurately and can improve the use of the group’s collective knowledge. This suggests that decision quality cannot be reduced to the question of “human versus AI.” The structure of collaboration among people before and after the system intervenes also matters.
In practice, therefore, monitor three things in decision-support tools. First, which information the user sees before the AI recommendation. Second, whether they must justify their own position in some way. Third, whether the system displays only one recommendation or also uncertainty, alternatives and information it did not incorporate into its conclusion.
A high-quality model can be dangerously persuasive precisely because it is usually right. A person can gradually get used to confirming its proposals until they stop using the knowledge that was the reason for keeping them in the decision process in the first place.
KEY TERMS
- Context displacement: A situation in which an AI recommendation reduces the attention a person gives to other relevant information.
- Judgment calibration: The ability to estimate realistically whether one’s own view is correct and how certain one should be.
- Decision sequence: The order in which information, the person’s own judgment and the AI recommendation are presented to the decision-maker.
- Human oversight: Active review of a system’s output rather than merely formal confirmation of its recommendation.