AI is making us faster. It may also be making us worse at thinking. A convergence of research from MIT, Harvard, Microsoft, and Wharton now shows that whilst generative AI tools boost short-term performance by 14-40%, they simultaneously erode the critical thinking, memory, and independent judgement they are supposed to augment. For UK organisations betting on AI-driven productivity, this creates an uncomfortable question: what happens when the tool you adopted to sharpen your workforce is quietly dulling it instead?
The cognitive debt nobody is measuring
The most striking recent evidence comes from MIT Media Lab. Kosmyna and colleagues monitored 54 participants with EEG during essay-writing sessions across multiple rounds. The ChatGPT group showed the lowest neural engagement across all 32 measured brain regions — systematically lower than both Google users and those writing unaided. By the third session, LLM users were mostly copy-pasting rather than engaging with the material.
Here is the part that should worry business leaders: when switched to writing without AI, those participants could not recover the neural connectivity patterns of the brain-only group. The researchers called this accumulated “cognitive debt.” 83% of ChatGPT users could not accurately recall key points from their own essays.
Strategic Reality: This is a preprint with a small sample, so treat it as early evidence rather than settled science. But it is the first neurological data showing AI use may create lingering cognitive effects — and it aligns with larger behavioural studies.
A rigorous randomised controlled trial published in PNAS reinforces the pattern. Bastani and colleagues at Wharton worked with roughly 1,000 Turkish high school maths students. Access to standard ChatGPT-4 during practice improved performance by 48%. But when AI was removed, students who had used unguarded ChatGPT performed 17% worse than those who never had access. They had used AI as a crutch and were, in the researchers’ words, “overly optimistic about their learning capabilities.”
The critical finding: a version of ChatGPT designed as a tutor — giving hints rather than answers — eliminated the negative effects entirely. Design matters enormously.
What the numbers actually say
| Study | Sample | Key finding | Limitation |
|---|---|---|---|
| MIT Media Lab (2025) | 54 participants, EEG | Lowest neural engagement with ChatGPT; cognitive debt effect | Preprint, small sample |
| Wharton/PNAS (2025) | ~1,000 students, RCT | 48% improvement with AI; 17% decline after removal | Single subject (maths), specific age group |
| Microsoft/CMU (CHI 2025) | 319 workers, 936 use cases | Higher AI confidence = less critical thinking | Survey-based, correlation |
| Gerlich (2025) | 666 participants | AI usage correlates with cognitive offloading (r = +0.72) | Self-report, correlational, MDPI journal |
| BCG follow-up (2026) | 244 consultants, ~5,000 interactions | 27% became Self-Automators | Single firm, elite population |
Critical Context: No long-term longitudinal studies yet track cognitive changes in regular AI users. The evidence base is growing fast but remains young. Every finding here should be read as a direction indicator, not a destination.
The real problem is not AI — it is how people use it
The 2026 BCG follow-up study tracked 244 consultants through approximately 5,000 AI interactions and identified three distinct collaboration modes. Cyborgs (60% of participants) engaged in continuous iterative dialogue with AI and developed new AI-related expertise. Centaurs (14%) used AI selectively whilst maintaining firm human control, achieving the highest accuracy and deepening their domain expertise.
Then there were the Self-Automators. 27% of these highly trained consultants delegated entire workflows and developed neither AI skills nor domain skills. They became passive conduits. As the researchers put it: “When employees default to Self-Automator behaviour — which over a quarter of our highly trained consultants did — organisations may be inadvertently hollowing out the very expertise that creates competitive advantage.”
That 27% figure is worth sitting with. These were BCG consultants — people selected for analytical rigour — and more than a quarter of them defaulted to wholesale delegation when given the opportunity.
Reality Check: The pull toward cognitive passivity is strong even among the highly skilled. If elite consultants default to Self-Automator behaviour, what happens in organisations with less analytical culture?
The Microsoft Research/Carnegie Mellon study captures the mechanism clearly. Surveying 319 knowledge workers across 936 real-world AI use examples, they found higher confidence in AI was associated with less critical thinking, whilst higher confidence in one’s own skills predicted greater critical engagement. The researchers identified a sharp irony: “By mechanising routine tasks and leaving exception-handling to the human user, you deprive the user of the routine opportunities to practice their judgement and strengthen their cognitive musculature, leaving them atrophied and unprepared when the exceptions do arise.”
The UK is walking into this problem with its eyes half-open
The UK context makes this especially pressing. According to the DSIT AI Adoption Survey (January 2026), only 16% of UK businesses currently use AI, with 80% having no plans to adopt. Among those who do adopt, 85% use natural language processing or text generation tools — precisely the category most associated with cognitive offloading.
Meanwhile, the UK AI skills shortage is acute and worsening. Harvey Nash’s Digital Leadership Report found 52% of UK tech leaders report an AI skills gap, up from 20% the prior year — the sharpest shift in 15 years. DSIT projects AI-related jobs could rise from 158,000 in 2024 to 3.9 million by 2035, yet only 17% of UK adults can explain AI in detail and just 21% of workers feel confident using it.
Hidden Cost: Cognitive offloading risks compounding the skills gap. If workers use AI to bypass skill development rather than enhance it, the underlying talent deficit deepens even as surface-level capability appears to grow. Organisations get faster employees who are becoming less capable.
The regulatory picture adds another layer. UK professional bodies hold individuals accountable for competence, but none have defined what maintaining competence means when core analytical tasks are increasingly delegated to machines. The GMC says doctors remain responsible for decisions made using AI. The ICAEW warns members they are liable to disciplinary processes if they fail to adhere to professional principles when using AI. But neither body has frameworks to assess or prevent cognitive dependency.
In education, the shift has been explosive. The HEPI/Kortext 2025 survey found 92% of UK students now use AI (up from 66% the previous year), with 88% using it for assessments. Only 36% had received any AI skills training from their institution. A generation is learning to work with AI before learning to work without it.
What organisations should actually do about this
The evidence points toward a clear design principle: AI tools should scaffold thinking, not replace it. The Bastani study’s finding that tutoring-style AI (hints rather than answers) eliminated negative learning effects is not just an educational finding. It is an organisational design principle.
| Maturity level | Priority actions |
|---|---|
| Early adopters | Define which tasks are “AI-assisted” vs “AI-independent.” Require manual competency demonstrations for critical skills quarterly. |
| Scaling organisations | Monitor for Self-Automator patterns. Build AI workflows that require iterative engagement (Cyborg mode) rather than one-shot delegation. |
| Advanced users | Invest in metacognitive training alongside tool training. Harvard Business Review research (January 2026) shows AI boosts creativity primarily for employees with strong metacognition. |
Implementation Note: The centaur model — strategic division of labour where humans retain control over tasks that build and exercise judgement — is the most evidence-supported approach. But it requires deliberate organisational design, not just good intentions.
Cognitive Load Theory offers a useful diagnostic. AI that reduces extraneous cognitive load — unnecessary friction, formatting, information retrieval — is beneficial. AI that reduces germane load — the effortful processing that drives learning and skill development — is harmful. The distinction is not about what task is being performed but about which cognitive demands are being removed.
The Jones Walker four-phase dependency model provides a warning trajectory: Enhancement (AI assists, humans understand), Integration (AI handles more, comfort grows), Dependency (professionals struggle without AI), Atrophy (skills necessary for independent practice deteriorate). Most organisations have no way of knowing where their teams sit on this spectrum.
Four risks most organisations are not thinking about
1. The competence verification gap. Professional liability insurers have begun adding AI use disclosure requirements and training mandates to policies. Courts and insurers are starting to revisit what constitutes “competent” practice when automation is standard. Organisations with no AI competency framework are exposed.
2. Collective creativity narrowing. Doshi and Hauser (2024) found AI-assisted writers produced stories rated more creative individually, but the stories were also more similar to each other. Meincke and colleagues (2025) found only 6% of AI-generated ideas were unique compared to 100% of human ideas. Individual output improves; collective innovation narrows. Organisations relying on AI for strategic ideation may converge on the same ideas as their competitors.
Competitive Reality: If every firm uses the same AI tools for strategy, differentiation comes from the human thinking applied on top. Organisations that let that muscle atrophy lose their only source of competitive advantage that cannot be replicated by a rival with the same subscription.
3. The training pipeline problem. Medical settings illustrate this clearly. A study referenced in The Lancet Gastroenterology & Hepatology (2025) found endoscopists using AI for colonoscopy detection performed worse when the AI was removed after just three months. A junior professional who never practises independently may never develop the pattern recognition that distinguishes experienced practitioners.
4. The false confidence trap. The Microsoft/CMU study found that higher confidence in AI was associated with less critical thinking. Workers who trust AI the most check it the least. This creates systematic blind spots precisely where organisations feel most secure.
The strategic question UK leaders should be asking
The question is not whether to adopt AI. The productivity evidence makes that inevitable. The OECD projects AI could add £55-140 billion annually to UK output by 2030. Not adopting means falling behind.
The question is whether your organisation is building cognitive resilience alongside AI capability. Three factors distinguish organisations that get this right:
They treat AI deployment as a cognitive architecture decision, not a procurement one. This means designing workflows that preserve germane cognitive load — the effortful processing that builds expertise — while offloading extraneous work.
They monitor for dependency patterns. If nobody in your organisation has written a strategy document, diagnosed a complex problem, or evaluated a nuanced risk without AI assistance in six months, you have a problem. And you probably do not know about it yet.
They invest in metacognition, not just tool training. Teaching people to use AI is the easy part. Teaching them to remain cognitively active whilst using it — to plan, monitor, and refine their own thinking rather than outsourcing it — is harder and more valuable.
Take Action: Start with a simple audit. Ask your team leads: which critical skills have not been exercised without AI assistance in the last quarter? The answers will tell you where your cognitive risk lies.
Historical parallels offer both reassurance and warning. Socrates feared writing would produce “the appearance of wisdom, not its reality.” Calculators did not destroy mathematical ability when used thoughtfully. But GPS did measurably reduce spatial navigation skills — UCL research showed London taxi drivers had significantly larger hippocampi than bus drivers following fixed routes, and longitudinal GPS studies demonstrated steeper memory decline with greater GPS dependence.
The critical difference: calculators offloaded computation, GPS offloaded navigation, writing offloaded memory storage. AI offloads reasoning itself. And reasoning is what distinguishes expertise from mechanical execution.
Sources and further reading
This analysis draws on research from MIT Media Lab (Kosmyna et al., 2025), Wharton/PNAS (Bastani et al., 2025), Microsoft Research/Carnegie Mellon (CHI 2025), Harvard/BCG (Dell’Acqua et al., 2023; Candelon, Kellogg, and Lifshitz, 2026), Gerlich (2025, Societies), Doshi and Hauser (2024, Science Advances), Meincke et al. (2025, Nature Human Behaviour), Brynjolfsson, Li, and Raymond (2025, Quarterly Journal of Economics), and UK data from DSIT AI Adoption Survey (January 2026), EY AI Sentiment Index (2025), Harvey Nash Digital Leadership Report, and HEPI/Kortext Survey (2025).
This article is original analysis by Resultsense, synthesising publicly available research to provide strategic perspective for UK business leaders. For guidance on developing AI deployment strategies that preserve cognitive resilience, explore our AI Strategy Blueprint or AI Risk Management services.