Cognitive offloading is a habit, not a knowledge gap. That's why training alone won't fix it

Every organisation rolling out AI has the same moment in it somewhere. It's 4pm, the deadline is close, the draft looks plausible, and someone accepts it without really reading it. Not because they don't know better. Because checking is effort, the tool has been right before, and the habit of handing thinking over forms one convenient shortcut at a time.

That habit now has a name in the research literature, cognitive offloading, and 2025 was the year the evidence stopped being speculative. If you're responsible for learning and development, the findings are worth ten minutes of your attention, because they change what a credible response looks like.

The research is in, and it's consistent

Four findings, from four very different methods, point the same way.

Frequent AI use correlates with weaker critical thinking. A study of 666 participants published in Societies found a significant negative correlation between frequent AI tool use and critical-thinking ability, and identified cognitive offloading as the mediating mechanism. The effect was strongest in younger participants, the people with the longest careers ahead of them.

The switch that turns thinking off is confidence in the tool. Researchers at Microsoft and Carnegie Mellon surveyed 319 knowledge workers across 936 real examples of AI use at work. Higher confidence in the AI predicted less critical thinking; higher confidence in one's own ability predicted more. The study also describes how knowledge work is shifting from doing tasks to verifying and supervising AI output. The uncomfortable implication: verification is precisely the work that lapses as trust in the tool grows.

The effect accumulates. An MIT Media Lab team measured brain activity while people wrote essays with an LLM, with a search engine, or with no tools at all. The LLM group showed the weakest neural connectivity of the three, remembered less of their own writing, and felt less ownership of it. The researchers call the pattern cognitive debt: capability quietly borrowed against, one assisted task at a time.

Deskilling is fast, measurable, and reaches experts. The starkest result comes from medicine. A multicentre study in The Lancet Gastroenterology and Hepatology tracked experienced clinicians, each with thousands of procedures behind them, after AI-assisted colonoscopy was introduced. Within months, their unassisted detection rates had fallen from 28.4% to 22.4%, a 20% relative drop. If a few months of AI assistance measurably erodes expert clinical skill, no profession should assume immunity.

The International AI Safety Report 2026, compiled by researchers from 30 countries, now lists routine cognitive delegation to AI among the emerging risks, citing early evidence of effects on critical thinking and memory. This has moved from conference-panel speculation to a documented workforce risk.

Offloading isn't the enemy. Unexamined offloading is

Worth saying clearly: cognitive offloading is not new, and not always bad. The foundational literature (Risko and Gilbert, Trends in Cognitive Sciences, 2016) defines it simply as using tools or actions to reduce the mental demands of a task, and people have done it since the first shopping list. Offloading arithmetic to a calculator freed mathematicians for harder problems.

The line the current research draws is specific. Offloading becomes a risk when the thinking you hand over is the thinking that would have built or maintained a skill worth keeping, and when nobody notices it happening. The clinicians in the Lancet study didn't decide to get worse at detection. It happened underneath their awareness, in the flow of work, which is exactly where no training course can follow.

Why the standard response falls short

The standard organisational response to a named workforce risk is a course. Commission a workshop on critical thinking with AI, get everyone through it, file the completion records. And a good workshop has real value: it names the risk, teaches the counter-moves, and creates shared vocabulary.

But look at what the research actually describes, and the gap is obvious. The behaviour forms in the daily flow of work, months after the workshop is forgotten. The driver is confidence in the tool, and confidence isn't recalibrated by being told once; it's recalibrated by repeated, examined experience. Awareness decays in weeks. And the completion certificate at the end records that someone attended, not that anything changed.

A course treats cognitive offloading as a knowledge gap. The evidence says it's a habit. Habits need a different instrument.

What changing a habit actually takes

The behaviour-change literature is unglamorous and consistent on this: habits change through practice that is situated in real work, repeated over months, and individually followed up. That's the loop Skilly Work runs.

Each quarter, everyone writes a short, structured reflection on a real, recent use of AI in their own work. Not a scenario, not a quiz: what they actually handed over, what they checked, what they'd do differently. The prompts rotate through the five SHARP habits, and two of them are, precisely, the counter-habits to offloading. Scrutinise is the habit of treating AI output as a first draft to check rather than an answer to accept. Practise is the habit of noticing where reliance is eroding a skill worth keeping, and keeping it live on purpose.

Each reflection is scored against a four-level rubric, from Unreflective (uses AI without noticing the choices being made) to Fluent (the habits are automatic and visible to colleagues), with warm, formative feedback and a human reviewer who keeps the final word. And each reflection ends with a committed next step that the following cycle opens by asking about. Did it stick? That commitment loop, spaced across quarters, is the part no one-off course can deliver.

There's a quieter design point underneath all of this. Writing a structured account of how you used AI, what you checked and what you'd change is retrieval, self-explanation and metacognition: cognitive work that cannot itself be offloaded. And if someone hands the reflection to an AI anyway, the scoring recognises shallow, generic writing for what it is. The medium is the counter-mechanism.

For L&D leaders: change the question you ask providers

The market is about to fill with courses on AI and critical thinking, and procurement will be hard to tell apart. One question cuts through: show me what changed, not who attended.

An attendance list shows activity. A capability record shows outcomes: where each person sits against observable habits, how that moved between cycles, which commitments stuck. Because Skilly Work's record accrues from real work over time, the same activity that builds the habits also produces the evidence that they're building, per person and per cohort. That's a record an L&D team can put in front of a board, and it's a materially stronger answer than a certificate when anyone asks what the organisation actually did about cognitive offloading.

The risk is real, the research is unusually consistent, and the fix is not another module in the LMS. It's a habit loop in the flow of work, with evidence coming out the other side. The interactive preview lets you feel the reflective loop in your browser in about two minutes, and the rollout playbook shows what a first cycle looks like.

Explore Skilly Work →