A junior researcher uses an AI tool to produce a polished debrief draft in 10 minutes. It is clear, well-structured and probably better than the average first draft produced manually. Most research leaders would call that progress.

What did the researcher practice while producing it? What did they learn?

Many of us know Daniel Kahneman’s distinction between fast, intuitive thinking and slower, deliberate reasoning. Academics have now proposed additional frameworks for AI as an external cognitive layer. Some call it “System 3”, others “System 0”. Whatever the terminology, large language models are becoming part of the daily workflow for many insights professionals.

Alongside the excitement of suddenly being able to deliver things quicker, colleagues and clients have raised a concern I recognise myself. Under pressure, accepting AI-generated answers can become the default, even when they need to be challenged.

Did we catch a cognitive virus?

This reminded me of Nicholas Carr’s 2008 essay Is Google Making Us Stupid? Carr described finding prolonged reading increasingly difficult due to exposure to the internet and questioned whether changing reading habits might also change how we think. Today’s version appears in a satirical Substack article about a scientist struggling to explain her work without ChatGPT: “My own words? I haven’t used my own words since 2022.”

With flu, you generally know something is wrong. A cognitive virus is trickier when the symptoms look like a very productive week.

What interests me is how quickly new cognitive habits can spread through organisations.

Research agencies are very good at spreading useful ways of working. When someone finds a better way to analyse data, write a report or generate ideas, we turn it into a workflow, train the team, encourage adoption, showcase at conferences; that is exactly what we should do with AI.

The catch is that workflows spread thinking habits as well as efficiency. Last month’s preprint Large-language models as a cognitive virus explores how AI-assisted habits might spread and reinforce themselves, like a virus. Some of this research reads like Klingon, but the central idea is clear: dependence can become self-reinforcing. As more cognitive work is delegated to AI, people get fewer opportunities to practice independent reasoning. That, in turn, can make further delegation easier and more accepted as normal.

In the model, once that pattern becomes established, simply dialling AI use back may not be enough to restore the previous level of cognitive autonomy. This is a theoretical possibility, not evidence that AI is making people less intelligent, but it argues that prevention could be easier than rebuilding lost habits later.

Where expertise comes from

For leaders, this turns a hypothetical concern into a very practical question: how should we introduce AI in ways that preserve the parts of work through which people learn and develop? If leaders reward speed and accept surface-level polish without challenging how the conclusion was reached, the definition of good work begins to shift. In environments with fewer opportunities or incentives for independent cognition, further delegation to AI becomes easier and increasingly culturally acceptable.

Many of the activities that build research expertise are not especially efficient. You read the verbatims that do not fit the neat story. You notice when two consumers use the same word to mean different things. You defend an interpretation, have it challenged, and sometimes discover that the clever recommendation you liked is not actually supported by the data.

Those moments are slow. They are also part of the learning that becomes expertise.

If AI removes the friction, we need to be deliberate about which friction was pointless and which friction was practice.

Build the thinking into the workflow

In that preprint one possible response is cognitive immunisation. Not rejecting AI completely, but designing its use so that it strengthens rather than substitutes for the cognitive abilities we want to preserve. The authors distinguish between AI as scaffolding and AI as substitution: does the tool help someone do the thinking, or completely remove the need to do it?

Imagine a team develops a brilliant new workflow for finding the story in a dataset. It saves hours. But as part of designing that workflow, they also define what the researcher should still work through themselves. The useful question is not simply “Did they use AI?” but “Was AI scaffolding the thinking or substituting for it?”

The AI session could then become an on-the-job coaching session. Instead of immediately producing the interpretation, it might ask: What do you think is happening? What evidence does not fit your story? What else could explain this pattern? What would make you change your recommendation? It can challenge, probe and nudge without immediately doing the thinking for them.

The point is not to make everyone perform an AI-free ritual before every task. It is to preserve deliberate practice where that practice builds judgement.

And that will look different depending on experience. An experienced director might ask AI for a first questionnaire draft and interrogate it using years of experience. A graduate might initially receive more questions, challenge and guidance. As their judgement develops, the scaffolding can gradually disappear. There is a simple test of whether it worked: can they apply the same reasoning to a new problem without the workflow holding their hand?

That is cognitive virus immunisation in practice. As System 3 becomes embedded in our work, the challenge is not to keep it out of the building. It is to design the work so that Systems 1 and 2 still get exercised.

Alexandra Kuzmina is innovation director: discovery at MMR Research Worldwide