The risk is not becoming less intelligent because we use AI. It arises when a tool provides an answer before we have formulated the problem, mobilized our knowledge and defined what a good answer should contain.
AI does not mechanically cause unlearning. The risk increases when it takes over scoping, production and evaluation of the same task. To preserve skills, it is necessary to maintain visible cognitive activity: formulate a hypothesis before the prompt, choose what remains human, verify against explicit criteria, and be able to explain the decision without the tool.
Delegating a task is not abandoning the reasoning process.
We have always delegated part of our effort to tools. A calculator handles calculations. A search engine makes information easier to access. A spellchecker spots errors. Generative AI goes further: it can suggest the problem, structure, arguments, wording and conclusion.
It is precisely this that makes it useful. It is also what makes the boundary less visible between what has been accelerated and what is no longer practiced.
A study presented at the CHI 2025 conference surveyed 319 professionals about 936 real-world situations involving generative AI. The authors do not claim that the tool eliminates critical thinking. They observe a shift: effort is focused more on verifying information, integrating the response and supervising the task. They also report that high confidence in AI is associated with lower self-reported critical effort.
The nuance matters. The issue is not delegation. It lies in delegating without knowing which competence should remain active.
A quick answer can save time. It proves neither that the problem was framed correctly nor that the reasoning was understood.
Four abilities can erode when they are no longer exercised.
Formulate the problem
A prompt creates the impression that the question is already clear. Yet a significant part of skilled work consists in deciding what we are actually looking for, for whom, under which constraints and according to which criteria. If AI systematically reframes a vague request, it may conceal that difficulty rather than solve it.
Draw on memory
We do not need to remember everything. We do, however, need a foundation of knowledge to recognise an inconsistency, connect two ideas and ask a better question. An external answer cannot replace this internal network. Without it, verification becomes superficial: the text seems plausible because we no longer have the reference points needed to challenge it.
Sustain attention
Reading a fluid response does not require the same level of engagement as building an argument. Attention can then shift to quickly selecting a pre-formed proposal. This comfort is sometimes appropriate. It becomes a problem if all complex tasks are reduced to validating outputs.
Exercise judgment
Judgment does not consist solely of spotting a factual error. One must assess relevance, consequences, interests at stake, and the expected level of proof. A technically correct answer may remain unsuitable for the context.
surveyed about 936 real-world uses of generative AI in the CHI 2025 study on critical effort at work.
Lee et al. · CHI 2025The smoothest response is not always the most reliable.
AI-generated language is often coherent, confident and well structured. These qualities of form can be mistaken for qualities of substance. The more an output resembles what we expect, the less likely we are to question its premises.
This phenomenon is reinforced by the pressure of work. When speed is required, verification becomes a final and compressible step. We reread the wording. We are less willing to check the source, date, scope or reasoning.
Useful trust is not blind trust in the tool. It is trust calibrated by task:
| Situation | Expected level of oversight |
|---|---|
| Rephrase a text that has already been approved | Check of meaning and tone |
| Summarise a supplied document | Control of omissions and citations |
| Explore hypotheses | Comparison of several options |
| Produce a business analysis | Validation by a competent person |
| Supporting a critical decision | Primary sources, traceability and explicit accountability |
Design practices that keep people in the thinking loop.
A usage policy is not enough if it does not change the work sequence. A few simple practices make cognitive effort visible.
These safeguards are not intended to slow down all tasks. They reserve effort for places where it creates value: framing, understanding, and decision-making.
Five questions before entrusting a task to AI.
Before the prompt, a team can ask itself five questions:
- Which skill does this task normally require?
- Do I need to learn, produce faster, or both?
- Which part of the reasoning must I retain so that I remain able to assess the answer?
- Which sources or data will make it possible to verify the result?
- Who will be accountable for the consequences if the proposal is incorrect?
AI can enhance our capabilities provided that its use does not render invisible the capacities on which it depends. The right question is not: can it do it? It is: what must we remain capable of doing when we ask it to do it for us?
Reference sources
- Lee et al., The Impact of Generative AI on Critical Thinking, CHI 2025
- Microsoft Research, The Metacognitive Demands and Opportunities of Generative AI
- Schwartz et al., Attending to Remember, 2025
FAQ
Does AI really cause skill loss?
Current research does not allow us to assert an automatic and uniform effect. They mainly show that the way effort is distributed between humans and tools changes reasoning, verification, and learning activities.
Should you avoid AI to learn?
No. It can explain, ask questions, act as a critical counterpart or provide feedback. It becomes less useful for learning when it systematically replaces the effort of recall, formulation and problem-solving.
Which task should remain human?
The choice depends on the business area and the risk. The problem framing, the definition of criteria, the decision-making process, and final responsibility should remain explicitly assigned.
How to verify that one still understands their own work?
A simple check is to close the tool and then explain the problem, assumptions, decision and limitations to another person.
