When a team of management researchers at the University of Bath set out to examine how large language models are reshaping workplace learning, they expected to find trade-offs. What they described in their peer-reviewed paper, published in April 2026 in the Human Resource Management Journal, was something more troubling: a mechanism by which routine AI use could quietly strip away the deep expertise that organizations spend years cultivating in their people.
The paper, titled “On the Dangers of Large-Language Model Mediated Learning for Human Capital,” argues that when employees consistently turn to chatbots for writing, analysis, and problem-solving, they replace hands-on cognitive experience with what the authors call “synthetic inputs.” Over time, that substitution weakens the tacit knowledge that underpins professional judgment. A separate experiment from the MIT Media Lab, measuring brain activity during essay writing, found biological evidence pointing in the same direction: participants who used an LLM showed measurably weaker neural connectivity than those who worked without digital assistance.
Two mechanisms of erosion
The Bath researchers, led by Callen Anthony, identify two specific pathways through which LLM dependence can degrade knowledge. The first is increased abstraction from real-world experience. AI-generated answers arrive as polished summaries, bypassing the messy, iterative process through which people normally build understanding. A junior analyst who asks a chatbot to interpret a dataset, for instance, skips the struggle of noticing anomalies, questioning assumptions, and revising conclusions that would otherwise sharpen their instincts.
The second mechanism is subtler. LLM outputs conflate fluency with accuracy. Text that reads smoothly can mask shallow or outright incorrect content, training users to mistake eloquence for understanding. The more polished the output, the less likely a reader is to interrogate it, especially if their own grasp of the subject has been thinning through disuse.
To counter these risks, the authors propose a concept they call “learning vaults”: structured periods during which employees set aside AI tools and engage directly with problems. The paper presents learning vaults as a speculative proposal rather than an empirically tested framework; the idea has not been validated through controlled trials. As described in a University of Bath briefing, the idea is not to reject AI but to treat unassisted work as a deliberate training regimen, much like drills in competitive sports. The goal is to preserve the pattern recognition, judgment under ambiguity, and domain-specific reasoning that no chatbot can build on a worker’s behalf.
What the brain scans show
Independent experimental evidence from the MIT Media Lab adds a biological dimension to the Bath team’s theoretical framework. In a preprint titled “Your Brain on ChatGPT” (arXiv 2506.08872), which has not yet undergone peer review, researchers tested 54 participants who wrote essays across multiple sessions under three conditions: using an LLM, using a search engine, or using no tools at all. Electroencephalography (EEG) recordings captured how broadly and intensely different brain regions communicated during each task.
The results were striking. Brain connectivity was strongest and most widely distributed when participants worked entirely on their own. In the LLM condition, connectivity was measured as weakest, suggesting that outsourcing composition to a chatbot narrowed the scope of neural engagement. The MIT team describes this pattern as “cognitive debt”: when an AI assistant handles argument structure, fact recall, and phrasing, the brain still participates but shifts into a narrower, more supervisory role. Less practice at the full range of cognitive tasks could, over repeated sessions, leave users less capable of performing those tasks independently.
The analogy the researchers draw is straightforward. A person who always uses a calculator may gradually lose confidence in mental arithmetic, not because the skill is gone but because it has gone unexercised. The same logic, they argue, applies to higher-order thinking when a chatbot handles the heavy lifting.
Where the evidence has limits
Neither study claims to have settled the question, and several gaps deserve attention. The MIT experiment involved 54 participants, a sample the authors themselves flag as a limitation. It also focused on a single task type, short-form essay writing, under controlled lab conditions. Whether the same patterns of reduced connectivity would appear during coding, data analysis, strategic planning, or other forms of knowledge work remains untested. Because the preprint has not been peer-reviewed, its methods and conclusions have not yet been independently evaluated by other researchers, and the findings should be treated with additional caution.
Interpretation is another open question. Reduced brain connectivity during a writing task does not automatically equal long-term cognitive decline. It could reflect efficient delegation: if an AI handles routine phrasing, the brain may redirect resources toward higher-level planning or critical evaluation. The preprint captures what happens during the task itself, not what happens to a participant’s skills weeks or months later. Longitudinal data, tracking the same workers over an extended period, does not yet exist in the published literature.
The Bath paper, while peer-reviewed, is a conceptual contribution rather than a field study. Its arguments draw on established research in human capital theory, learning science, and organizational behavior, but the authors did not track actual employees over time. No published dataset yet shows whether workers who use LLMs daily for months or years perform measurably worse on complex tasks than peers who use AI sparingly, or whether any early decline reverses when habits change.
The “learning vaults” proposal, though grounded in well-established principles of deliberate practice, remains a speculative concept within the Bath paper that lacks empirical testing. No controlled trial has measured whether scheduled AI-free periods actually maintain or restore deep expertise in a workplace setting. Practical design questions, such as how long vaults should last, how frequently they should recur, and which roles benefit most, remain unanswered.
Notably absent from the conversation so far is any formal response from major AI developers. Companies like OpenAI, Google, and Microsoft have not publicly addressed the specific erosion risks the Bath researchers describe. Internal research on the topic may exist but has not been shared, leaving employers and policymakers to weigh the evidence with only partial visibility.
The feedback loop no one is testing
Beneath the headline findings sits a risk that neither study directly measured but that both sets of evidence imply. If heavy AI use reduces a person’s ability to think critically about a subject, it also reduces their ability to evaluate whether the AI’s output on that subject is correct. The result is a self-reinforcing cycle: less cognitive engagement leads to weaker error detection, which leads to greater trust in AI-generated answers, which leads to even less engagement.
No published study has directly tested this loop. But the logic follows from the Bath paper’s framework, which warns that synthetic inputs displace experiential learning, and from the MIT team’s EEG data, which shows the brain doing less when a chatbot does more. If both findings hold up under further scrutiny, the implication is that the cost of over-reliance compounds quietly over time, becoming harder to detect precisely because the capacity to detect it is what erodes first.
Safeguards organizations can build into AI adoption
For managers and policymakers, the practical question is not whether to use AI but how to use it without hollowing out the human expertise that gives AI outputs their context and oversight. The science is still developing, but several low-cost measures align with what the current evidence supports.
The Bath team’s learning vaults offer a starting point, though one that still needs real-world validation: designate regular intervals where employees tackle core tasks without AI assistance, particularly work that builds judgment, pattern recognition, or domain reasoning. Organizations can layer on complementary safeguards, such as requiring workers to draft initial outlines before consulting a chatbot, asking them to explain key decisions in their own words, or scheduling periodic performance checks that must be completed unaided.
These steps are not bans on AI. They are the cognitive equivalent of cross-training, designed to ensure that the speed gains from AI-assisted work do not come at the cost of the skills that make human oversight meaningful. As adoption accelerates, with recent industry surveys showing a majority of knowledge workers already using generative AI on the job, the window for building these habits into workplace culture is narrowing. The strongest reading of the current evidence is not that AI tools are inherently harmful, but that treating them as a default rather than a supplement carries a cost that grows harder to reverse the longer it goes unaddressed.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
