An employee who uses AI every day is not necessarily becoming dependent in a harmful way. They may be learning a new method, getting useful feedback or using a tool that is now part of competent professional practice.
An employee who produces better work with AI is not necessarily becoming more capable either. The improved result may depend on suggestions they cannot yet evaluate.
The useful distinction is between assistance that people can understand and direct, learning that expands what they can do, and reliance that leaves important gaps hidden. Those conditions can coexist in the same person across different tasks.
For business and learning leaders, the question is therefore more precise than whether employees use AI too much: which capabilities does the role still require, and what evidence shows those capabilities developing?
Reliance on a tool is not the same as a loss of skill
Work has long depended on documentation, software, colleagues and specialist support. Requiring an employee to reproduce every process from memory would often be an irrelevant standard.
An analyst can be competent while consulting a reference. A coordinator can use a spreadsheet function without remembering its exact syntax. Someone can work effectively with an accessibility tool that remains essential to their performance.
The important question is what the person needs to contribute. Can they define the problem, recognize an unsuitable result and take an appropriate next step? Which parts can reasonably remain delegated to a dependable tool or another specialist?
Harmful dependence becomes a concern when the arrangement conceals a capability the job still needs. If nobody can recognize a consequential error because everyone assumes the assistant has checked it, a strong-looking output offers little reassurance.
That is a question about the work system as well as the individual. People cannot develop an expected capability if the process never gives them an opportunity to practise it.
The research gives different answers in different settings
Evidence of learning with AI exists. So does evidence of learning being weakened. The studies do not support a universal verdict.
In Generative AI at Work, published in 2025, Brynjolfsson, Li and Raymond examined a staggered rollout among 5,172 customer-support agents. Their learning analysis considered periods when technical outages interrupted AI recommendations. Previously exposed workers continued to conduct chats more quickly than before adoption, which the authors interpret as evidence consistent with learning.
That analysis has limits: outages were rare, estimates were noisy and the mix of chats could differ. Its outcome was chat duration, rather than a comprehensive assessment of independent expertise or service quality. The finding concerns a particular workplace deployment.
A 2026 preprint by Shen and Tamkin studied a different situation: Python users learning an unfamiliar programming library. The main randomized study involved 52 participants. The AI-assisted group performed worse on an immediate comprehension quiz, while its task-time improvement was not statistically significant.
This was a small, short experiment with a chat assistant, not a longitudinal study of employees' careers. It was conducted by researchers affiliated with an AI provider. It raises a relevant concern about acquiring new skills; it does not establish that all AI use erodes existing expertise.
A 2025 PNAS study by Bastani and colleagues provides another comparison. In a Turkish high-school mathematics experiment, a general-purpose AI tutor improved assisted practice but weakened subsequent unaided exam performance. A tutor designed with learning safeguards largely avoided that negative effect.
This is school evidence from a particular deployment in 2023, not direct evidence about adult employees or today's tools. Its relevance here is a distinction to investigate: success during assistance and learning afterward can diverge.
Taken together, these studies make the way assistance is designed and used worth examining. They do not supply a general rate of workplace deskilling.
Define the capability before judging the dependence
A role can change as tools improve. It is reasonable to reconsider which procedures people need to perform manually.
It is also reasonable to retain specific human knowledge where the person must judge a result, handle an exception or coordinate with others. Make that expectation explicit rather than treating “AI literacy” as a complete description.
Consider a hypothetical experienced event coordinator learning a new spreadsheet method to match participant registrations with workshop bookings. This is an illustrative example, not a Learning Coach client case.
An AI assistant proposes a formula. The sheet produces a tidy list. That result could save useful effort, but the learning question remains open.
The coordinator need not memorize the formula. They do need to understand what is being matched, how duplicate registrations are handled and whether cancelled bookings remain in the source information. Otherwise, a plausible list could be treated as a reliable attendance record without the necessary checks.
The human capability is therefore more specific than “use spreadsheets” or “work without AI.” It includes recognizing what would make this particular match misleading and knowing how to investigate.
Look at the moment the task changes
Suppose a later registration file contains a participant booked into two workshops. The old formula still returns a result.
What happens next is informative.
The coordinator might notice the ambiguity, consult the relevant documentation and adjust the process appropriately. They might ask a colleague who understands the booking system. They might use AI to explore the issue while checking its suggestions against the actual records.
All of those responses can reflect competent work. Independence need not mean isolation.
Alternatively, the coordinator might repeatedly regenerate the formula until the output looks tidy, without understanding which booking it selects. The concern is the unresolved matching decision, rather than the number of prompts.
A changed task can therefore reveal something that the original successful output did not. But one difficult attempt is not proof of long-term deterioration. It may reveal an unfamiliar edge case, insufficient teaching or a process that nobody has clearly specified.
Use the observation to decide what support is needed.
Observe learning under conditions that match the purpose
Different questions require different conditions.
If the aim is to assess normal job performance, include the tools and references that the role appropriately uses. Removing them can create an artificial task and obscure the capability that matters.
If the aim is to develop a particular concept, a short agreed practice activity might temporarily withhold a generated answer while leaving relevant instructions and access support available. That should be a deliberate learning choice, not a surprise test.
If the aim is resilience during a service interruption, identify the actual fallback. It may be a documented manual process, a different approved tool or a decision to pause. Not every role needs to reproduce the automated workflow unaided.
In the booking example, these purposes lead to different activities. Normal work might involve reconciling a new file with all approved support available. Concept practice might involve working through a few fictional records by hand. Continuity planning might involve explaining whom to contact when the booking integration is unavailable.
The organization should say which question it is asking. Otherwise, people can be judged against a version of competence that the job never required.
Make some of the learning work visible
A useful development activity can be small.
For the coordinator, first make the matching rule explicit using a handful of fictional records. Let them identify an ambiguous case. Then use the assistant to suggest an approach and compare its behavior with the intended rule.
Return to the issue on a later, appropriately varied task. Look for evidence that the person recognizes the relevant problem and can choose a sensible way to resolve it.
These are proposed practices, not a validated training package. Their value must be tested in the actual setting.
The amount of assistance can change with experience. Someone learning an unfamiliar method may need a worked example and direct explanation. Someone who understands the method may benefit from exploring a harder exception. The goal is a better match between the support provided and the learning required.
Simply instructing employees to “think critically” leaves that design work undone.
Interpret the evidence without turning it into a score
A team review can examine several observations without collapsing them into a single independence rating.
| Observation | What it can help establish | What it cannot establish alone |
|---|---|---|
| The AI-assisted output meets the agreed requirements | Useful performance in that task and support arrangement | Durable learning by the employee |
| The employee identifies a relevant limitation and checks it appropriately | Judgment about that particular result | Competence across every related situation |
| A later changed task is handled appropriately | Evidence of development or adaptation | Permanent expertise or a universal training effect |
| A fallback is used successfully during a planned exercise | Readiness for the tested interruption | Recovery from every possible failure |
This is an editorial discussion aid, not an employee assessment instrument. Its purpose is to keep conclusions proportionate to the evidence.
An employee's account of what they learned is useful, but compare it with work they can demonstrate. Equally, an observed difficulty deserves a conversation about the conditions and the support available.
The aim is to discover where understanding is secure, where assistance remains useful and where an important decision has effectively disappeared from the process.
Give employees a reason to invest in understanding
Learning takes attention. If a team rewards only immediate output, the time needed to examine an explanation or practise a method can become difficult to justify.
Managers can make a practical choice: identify a capability worth developing and allow some work to serve that purpose. They can explain which checks matter and when seeking help is appropriate.
Learning Coach's perspective is that workplace AI should be judged partly by the opportunities it creates for people to become better able to understand and act. That goes beyond measuring usage. It also goes beyond insisting that every previously manual skill remain unchanged.
The broader human learning advantage in the AI era provides the context. The local decision is concrete: what should this employee become better able to do through this work?
A useful question for a workplace learning event
An executive or L&D event can make the distinction tangible.
Present the fictional booking workflow in two versions: one that produces an answer immediately, and one that gives the coordinator a chance to understand and examine the matching rule. Ask participants which work outcome and learning outcome each version serves.
Then change the data. The discussion becomes less about enthusiasm for a tool and more about the capability the organization expects to retain.
This is a proposed event activity, not evidence that a keynote produces a measurable skill gain. Organizers can use the guide to designing a corporate AI learning summit to connect the discussion with subsequent practice.
The research reviewed here was checked on 28 October 2026. It is selected evidence from distinct settings, not a systematic review or a prediction of long-term employee dependence. The workplace example and practical suggestions are original editorial proposals.
The question worth carrying forward is whether people are becoming better able to direct the work that still requires their judgment—and how the organization would know.
Bring that conversation to your next leadership or workplace learning event. Explore Learning Coach's international keynotes and request a tailored proposal.
References
Brynjolfsson, Li and Raymond (2025), Generative AI at Work. Quarterly Journal of Economics; see the worker-learning analysis and its limitations.
Shen and Tamkin (2026), How AI Impacts Skill Formation. Provider-affiliated preprint, revised 1 February; short coding experiment with an immediate comprehension assessment.
Bastani and colleagues (2025), Generative AI without guardrails can harm learning. PNAS; high-school mathematics experiment with different AI tutor designs.
