The Hidden Psychology of AI Adoption

‘Draft a reply.’ ‘Summarise this in one sentence.’ ‘Give me talking points for the meeting.’

Prior to the 30th November 2022, (ChatGPT's launch date), there was little alternative but to think. Our brains struggled for insights, solutions and the words to string them together. That friction was not always pleasant, but it was the mechanism that kept our analytical, creative and problem-solving capabilities sharp.

AI has absorbed much of the cognitive and manual load and returned something we're chronically short of - time. But the cost of that time does not appear in any productivity metric. Many of the skills the World Economic Forum identifies as most critical for business, (analytical thinking, creative thinking, leadership and self-awareness), are precisely those that research suggests habitual AI use may be eroding. The skills gap meanwhile remains the single biggest barrier to business transformation, with nearly 40% of skills expected to change by 2030 (1).

Business leaders spend a lot of time asking how employees can use AI more effectively. Perhaps the more pressing question is what AI is doing to their employees in the process, and what, if anything, they are doing about it. This article draws on recent research to examine what is happening, and what it implies for how organisations approach AI literacy.

What's AI's impact on business brains?

Most conversations about AI risk revolve around hallucinations, data security and compliance. These concerns certainly deserve attention, but there is another risk that is harder to spot, harder to measure and harder to reverse: AI is changing how workers think.

Critical thinking erosion

In 2025, researchers at Microsoft Research and Carnegie Mellon found that the more confidence knowledge workers placed in generative AI, the less critical thinking effort they applied to its outputs (2). A separate study found that frequent AI use was associated with lower critical thinking scores, driven by cognitive offloading, (handing the first pass of thinking to the tool rather than doing it independently) (3). These findings describe a behavioural pattern that emerges from routine, normalised AI use - the very thing today's businesses are trying to push for.

Cognitive inhibitor or augmentor?

The first clinical evidence of cognitive deskilling playing out in practice came in 2025, when a multicentre study found that endoscopists' unaided adenoma detection rate fell significantly after routine exposure to AI-assisted colonoscopy. The researchers noted that their findings ‘temper the current enthusiasm for rapid adoption of AI-based technologies.’ (4).

The mechanism is not specific to medicine. Skills that go unpractised atrophy. Decades of research in learning science, led by Robert Bjork's work on desirable difficulties, shows that the effort involved in working through a problem, (retrieving knowledge, weighing evidence, generating a position), is the mechanism through which durable skill is built and maintained (5, 6). Make the process effortless and immediate performance improves, but long-term capability weakens.

Apply that to a workforce using AI for first-pass thinking hundreds of times a week. Each individual interaction saves minutes. Collectively, they remove the daily practice that keeps evaluation, analysis and judgement reliable. Whilst productivity gains from AI are immediate and visible, capability loss is gradual and largely invisible until it is often too late.

There is a counterargument, however, that AI can also augment cognition by offloading routine processing and freeing bandwidth for higher-order thinking. For well-designed, intentional use, there is evidence for this. The problem is that intentional use is not necessarily the norm, whereas routine, normalised, low-friction use is. And under those conditions, the evidence points the other way. (7)

Sycophancy and the distortion of judgement

Deskilling affects what workers can do independently. Sycophancy affects the quality of the judgements they make with AI.

AI models affirm users' positions nearly 50% more often than humans do, (including in cases involving deception or harm), according to research published in Science in 2026, which tested 11 state-of-the-art AI models on their sycophantic responses. They found that even a single interaction with a sycophantic AI model increased participants' conviction that they were correct, while reducing their willingness to consider other perspectives. More concerning is that users preferred the sycophantic model, trusted it more, and were more willing to use it again, despite it distorting their judgement. The researchers identify the structural incentive problem directly: ‘the very feature that causes harm also drives engagement’ (8).

For organisations deploying AI as a decision-support tool, this is a governance concern. A model trained to be agreeable will validate a flawed premise as readily as a sound one. In contexts where decision quality depends on challenge and honest appraisal, that tendency works against organisational interests.

Psychological safety and AI adoption

Organisational AI adoption significantly reduces employees' sense of psychological safety, which in turn increases rates of depression, according to a 2025 longitudinal study in Humanities and Social Sciences Communications. Promisingly, the study also found that leadership quality moderated the effect: where leaders communicated transparently about AI, treated the transition fairly and maintained trust, the reduction in psychological safety was substantially weaker. The harm associated with AI adoption was not inevitable; it tracked the quality of the leadership environment in which adoption took place (9).

Beyond these findings, there is consistent evidence linking AI adoption to professional identity threat, (particularly in roles where AI encroaches on the work that people find meaningful), and to behavioural disengagement driven by perceived loss of skill and autonomy (10). However, employees who develop what researchers call an ‘AI-inclusive identity’, (a stable sense of how AI fits within their professional practice), showed significantly weaker identity threat responses and lower rates of withdrawal behaviour (11), according to a 2026 study applying social identity theory.

Compounding risks

These risks do not operate in isolation. A workforce that has offloaded critical thinking to AI is precisely the workforce least equipped to detect when AI is flattering rather than informing them. Reduced capacity for independent evaluation makes sycophantic outputs harder to catch; sycophantic outputs further erode the habit of scrutiny; and both accelerate deskilling by eliminating the occasions on which independent judgement gets exercised at all. Employees already experiencing identity threat and disengagement are meanwhile less likely to invest in the evaluative habits that would interrupt this cycle. The risks intertwined, and addressing one while ignoring others is merely a partial intervention in a compounding problem.

What must AI literacy education address?

Using AI less is neither realistic nor desirable; the goal is to leverage it in ways that preserve and bolster the cognitive habits upon which your business relies.

That requires AI literacy education to do three things that most current provision does not. Firstly, help people build an accurate picture of where their expertise remains irreplaceable - this is an identity intervention as much as a knowledge one, and it needs to be designed accordingly. Secondly, build critical evaluation as a core skill rather than an afterthought: when to trust an AI output, what the model may be optimised for, how to identify errors the system will not flag. Fluent use without sceptical evaluation compounds the risks the evidence describes. Thirdly, reduce anxiety through genuine competence rather than reassurance - people need accurate mental models of what AI systems are, how they work, and where they fail, because reassurance without understanding does not hold when the technology behaves unexpectedly, whereas competence does.

AI literacy education is not sufficient on its own, and presenting it as such does a disservice to organisations trying to get this right. It sits within a broader challenge of organisational design - leadership, role structures, workflow, culture. But within that architecture, it earns its place when it takes the psychological dimensions of adoption seriously, prioritises critical thinking over fluency and is honest with learners, about both the capabilities and the limits of the tools they are being asked to use.

That is a higher bar than most current provision meets. The evidence suggests it is the right one.

References

1. World Economic Forum (2025). The Future of Jobs Report 2025. https://www.weforum.org/publications/the-future-of-jobs-report-2025/
2. Lee, H. P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3706598.3713778
3. Gerlich, M. (2025). AI tools in society: impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006
4. Budzyń, K., Romańczyk, M., Kitala, D., et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology & Hepatology, 10(10), 896-903. https://doi.org/10.1016/S2468-1253(25)00133-5
5. Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher, R. W. Pew, L. M. Hough, & J. R. Pomerantz (Eds.), Psychology and the real world. Worth Publishers. https://bjorklab.psych.ucla.edu/wp-content/uploads/sites/13/2016/07/EBjork_RBjork_2011.pdf
6. Bjork, R. A., & Bjork, E. L. (2020). Desirable difficulties in theory and practice. Journal of Applied Research in Memory and Cognition, 9(4), 475-479. https://doi.org/10.1016/j.jarmac.2020.09.003
7. Jose, B., Cherian, J., Verghis, A. M., Varghise, S. M., Mumthas, S., & Joseph, S. (2025). The cognitive paradox of AI in education: between enhancement and erosion. Frontiers in Psychology, 16, 1550621. https://doi.org/10.3389/fpsyg.2025.1550621
8. Cheng, M., Lee, C., Khadpe, P., Yu, S., Han, D., & Jurafsky, D. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science, 391(6792). https://doi.org/10.1126/science.aec8352
9. Kim, B.-J., Kim, M.-J., & Lee, J. (2025). The dark side of artificial intelligence adoption: linking artificial intelligence adoption to employee depression via psychological safety and ethical leadership. Humanities and Social Sciences Communications, 12, 704. https://doi.org/10.1057/s41599-025-05040-2
10. Jussupow, E., Spohrer, K., & Heinzl, A. (2022). Identity threats as a reason for resistance to artificial intelligence: survey study with medical students and professionals. JMIR Formative Research, 6(3), e28750. https://doi.org/10.2196/28750
11. Ashraf, A., Min, Q., & Ashraf, A. (2026). A moderated mediation model of AI-driven identity threats and employee cyberloafing: the role of AI-inclusive identity. European Journal of Investigation in Health, Psychology and Education, 16(4), 52. https://doi.org/10.3390/ejihpe16040052

‍