Automating knowledge work can save time while reducing opportunities to develop one’s own judgment. If employees accept ready-made answers without checking assumptions, they may gradually lose the ability to recognize unusual situations and errors. Researchers at INSEAD therefore recommend designing AI use so that it supports critical thinking rather than merely producing results faster.

Offloading a demanding cognitive task to an external tool is not new. Generative AI, however, makes it possible to delegate much more complex activities: analyzing a problem, constructing an argument, proposing a strategy or searching for a solution. Authors from INSEAD, ESSEC and Nova School of Business and Economics warn that regular “cognitive offloading” can weaken critical thinking, contextual expertise and the ability to make independent decisions in more complex tasks.

The risk is not only that the system gives a wrong answer. Even a correct output can reduce the amount of thinking the employee performs while solving the problem. If an expert repeatedly arrives only at a finished proposal, they lose part of the experience that comes from formulating hypotheses, comparing alternatives and looking for contradictions.

The authors use the experience of the design company With Company. Two teams worked on the same branding problem. One worked more traditionally, while the other relied heavily on AI. The AI team produced almost 1,700 visual concepts and greatly expanded the number of possibilities. The human team created fewer variants, but its proposals had a more coherent story and stronger context from the start. In the end, its work was selected. The lesson is not that the human team “beat” AI, but that the two approaches displayed different strengths.

A company therefore does not have to choose between banning the technology and fully automating the work. One option is a parallel process. The person first creates their own solution and AI then offers additional alternatives. The difference between the two outputs becomes a source of learning.

The second area is the user interface itself. A conventional conversational system often produces ever more blocks of text that the user then simplifies and reformulates. The authors point to an alternative design: the system can display source data, competing evidence and contradictions in a way that forces the person to actively assess the options rather than passively accept a summary.

Specialized mathematical tools provide an example. They do not merely give researchers a final answer but help them develop the working process, verify proofs and search for errors.

For management, this creates an important productivity question. If a company rewards only speed of completion, it creates an incentive to offload as much thinking as possible to AI. In expert roles, however, it is also necessary to monitor the quality of explanation, the ability to work without the tool and the diversity of approaches.

The point is not to protect an old way of working. It is to design the use of AI so that it increases people’s capabilities rather than gradually removing the activities through which expert judgment is formed.

KEY TERMS

• Cognitive offloading: Transferring demanding cognitive activity to an external tool.
• Expert judgment: The ability to interpret a situation on the basis of knowledge, experience and context.
• Cognitive homogeneity: A reduction in the diversity of ideas and approaches.
• Parallel problem-solving: Independent work on a problem by a person and by AI followed by comparison of the results.