AI and critical thinking: What the research actually shows

An approval button tells you that a person was present. It tells you very little about whether they understood the decision.

That is the question I care about when putting humans in the loop of an AI system. If the agent supplies the diagnosis, the evidence, and the recommended action, what work is left for the person approving it? Reading the explanation can feel like understanding the problem. The two are not interchangeable.

Recent research gives us reasons to examine that distinction. It does not establish that AI makes people less intelligent, damages their brains, or destroys an engineering team's diagnostic ability within a few months. Those are much bigger claims than the studies can carry.

The narrower findings are useful enough.

What the essay-writing study measured

In Your Brain on ChatGPT, Nataliya Kosmyna and colleagues studied essay writing with three conditions: an LLM, a search engine, and no external tool. Their arXiv preprint reports 54 participants across the first three sessions, with 18 completing a fourth session that switched some participants between the LLM and unaided conditions. [1]

The researchers used EEG to examine brain connectivity during writing. They reported the strongest and most distributed connectivity in the unaided group, intermediate engagement in the search group, and the weakest connectivity in the LLM group. LLM users also reported less ownership of their essays and had difficulty accurately quoting their own work.

These are findings about engagement, recall, and writing under particular experimental conditions. An EEG connectivity difference is not a diagnosis of brain damage or a general intelligence score. Nor does a study conducted over four months establish that four months of everyday AI use causes lasting cognitive decline.

The task matters, too. Having a model produce an essay differs from using it to challenge an argument you have already written. In the crossover session, the authors reported higher memory recall and engagement for participants who moved from unaided writing to LLM assistance. That result makes a blanket "AI use is bad for thinking" conclusion particularly unhelpful.

For an engineering team, this is a reason to ask how a tool changes participation in a task. It is not evidence that approving an agent's incident diagnosis causes the same effects as LLM-assisted essay writing.

Confidence changes the work people report doing

Hao-Ping Lee and colleagues approached the question differently. Their CHI 2025 paper surveyed 319 knowledge workers, who contributed 936 examples of using generative AI at work. [2]

Higher confidence in the AI was associated with less reported critical thinking. Higher task-specific confidence in one's own ability was associated with more. The researchers also described a shift in the work: toward checking information, integrating responses, and taking responsibility for the task.

That last finding deserves attention. AI can move thinking rather than simply remove it. Checking whether an answer fits a specific system can be demanding work, even when drafting the answer was easy.

But this was a survey of people's accounts, not a longitudinal test showing that their reasoning skills deteriorated. An association between confidence and reported effort does not establish that trust caused skill loss. Less effort on a routine task can also be a benefit. The question is whether the person still performs the checks that matter.

For human oversight, I would pay more attention to the quality of those checks than to whether an approval happened.

Friction is a proposal, not a proven cure

Kuangzhe Xu and colleagues argue for "scaffolded cognitive friction": interactions that prompt people to examine assumptions rather than accept a fluent answer. [3] Their perspective preprint includes an analysis of research literature and proposes ways to study cognitive engagement. It is not an intervention trial demonstrating that extra interface steps preserve engineers' skills.

The idea is worth trying carefully. Asking someone what evidence would disprove a diagnosis could be useful. Making them click through three confirmations of the same recommendation might add nothing.

More friction is not automatically better oversight. It has to earn its cost.

What I would change in an agent workflow

Consider a hypothetical service incident. An agent notices a rise in errors, proposes a restart, and presents a convincing explanation. This is an illustrative scenario, not a report from a particular company or production system.

The reviewer could approve it because the explanation sounds reasonable. Or they could inspect whether the errors began after a deployment, check whether a dependency is failing, and consider what a restart would disrupt. The second review has an independent basis for accepting or rejecting the recommendation.

These studies do not measure the difference between those two incident workflows. They motivate a design question that a team can test. My starting point would be:

1. Make evidence inspectable. Show the observations behind a recommendation, their source and age, and the important gaps. A generated explanation should not be the only way to examine the situation.
2. Ask for reasoning where the risk warrants it. For an ambiguous or consequential action, ask the reviewer to identify a competing explanation or a condition that would make the action unsafe. Avoid turning every routine action into an essay.
3. Practise outside live emergencies. Use synthetic exercises to let people form an initial diagnosis before seeing the agent's recommendation. Do not withhold useful assistance during a real incident just to create a learning opportunity.
4. Evaluate the review, not the click. In controlled exercises, examine whether reviewers catch intentionally flawed recommendations, use relevant evidence, and recognise when to escalate. Track the time cost and unnecessary rejections as well.

These are proposed practices, not outcomes demonstrated by the cited papers. A team could compare assisted-only exercises with exercises that include independent diagnosis, then see whether error detection or reasoning improves. If an added step creates work without improving the decision, change it.

And do not make the human reviewer the entire safety system. Access controls, action limits, independent validation, and recovery procedures still have to work. A thoughtful operator is not a substitute for a bounded agent.

Keep the claim proportional to the evidence

The phrase "cognitive debt" is useful if it prompts a concrete question: what practice or understanding are we postponing when a tool does this task for us? It becomes misleading when treated as a measured rate of deterioration that applies to every user.

The EEG study, the worker survey, and the perspective paper provide different kinds of evidence. They should not be combined into a single story that claims to prove inevitable decline. We still need evidence about long-term capability, different kinds of AI use, and which interventions help in which tasks.

Until then, I would make the oversight requirement explicit: the person approving a consequential action needs enough evidence, understanding, and authority to disagree with it.