An employee receives a report at 9 a.m. By 9:15, an artificial intelligence-powered assistant has summarized the document, identified the major risks, proposed three recommendations and drafted an executive response.
The work is faster; the output may also be better. From a performance perspective, this looks like success. But from a learning and development perspective, a pivotal question emerges: What did the employee actually have to think through in order to produce that result?
That distinction is becoming increasingly important for L&D leaders. Generative AI can remove repetitive work, increase access to expertise, provide feedback and help employees solve problems that previously required considerably more time. None of that should be dismissed.
A randomized field experiment involving thousands of knowledge workers found measurable productivity benefits from generative AI, including faster document completion. But productivity and learning are not the same outcome.
As organizations integrate AI more deeply into everyday work, L&D may need to confront an uncomfortable possibility: Employees can become better at producing results without becoming proportionally better at producing those results independently. Performance is visible. Learning is harder to see.
L&D researchers have long distinguished between performance during practice and learning that persists after practice. As Nicholas Soderstrom and Robert Bjork observed, what people do successfully while training is not always a reliable indicator of what they will retain or transfer later.
That distinction becomes particularly consequential in AI-enabled workplaces.
Imagine two analysts asked to interpret a complex dataset. The first examines the information, develops hypotheses, compares explanations, tests assumptions and eventually produces a recommendation. The second asks an AI system to analyze the same data and then evaluates its answer. Both may produce excellent recommendations. But they have performed different cognitive tasks.
The second analyst may have worked more efficiently—and efficiency may be exactly what the organization needs. Yet if developing analytical capability is also an objective, L&D must ask whether the mental processes required to build that capability were exercised.
Not all cognitive effort is waste
Modern workplace technology is often designed to reduce friction. That makes sense when the goal is execution. But learning follows a different logic.
Research on the generation effect has repeatedly shown that people tend to remember information better when they actively generate it rather than simply receive it. A meta-analysis covering 86 studies found a meaningful memory advantage for generated versus read information, and further meta-analysis further confirmed that the conditions under which people generate information affect memory outcomes.
Learning leaders should distinguish between unproductive cognitive load and productive cognitive load. Productive cognitive effort requires learners to retrieve, explain, compare, decide, generate, test, reflect or detect errors. AI can be extraordinarily valuable when it removes the first. Problems can emerge when it systematically removes the second.
For L&D, the objective should not be to make learning easier. It should be to remove the wrong difficulty while preserving the useful difficulty.
Answer engine or learning partner?
It would be equally misleading to conclude that AI weakens learning. The evidence is more nuanced.
A 2025 randomized study published in Scientific Reports found that a carefully designed AI tutor produced stronger learning gains than an in-class active learning condition in the educational context studied. The AI system did not simply deliver answers; it was explicitly designed around established learning principles.
Research in workplace contexts also suggests that employees perceive generative AI as capable of supporting individual learning, skill development and collaborative knowledge sharing. A 2025 study involving employees in a multinational corporation examined how workers experienced AI productivity tools as part of workplace learning.
The distinction is crucial. AI can function either as an answer engine that removes a cognitive task or as a learning partner that helps perform it more effectively
An answer engine removes the need to perform a cognitive task. A learning partner can ask the employee to perform it better. Instead of immediately generating a recommendation, for example, AI might ask:
- What evidence supports your conclusion?
- What alternative explanation have you considered?
- What assumption would make this recommendation fail?
- How confident are you in this judgment?
- What information would change your decision?
Critical thinking is shifting
Recent research among knowledge workers provides another important warning against simplistic conclusions.
A 2025 study presented at the CHI Conference examined 319 knowledge workers and 936 real-world examples of generative AI use. Workers reported that greater confidence in AI was associated with less critical-thinking effort, while critical-thinking activities increasingly shifted toward verifying information, integrating responses and task supervision
This suggests that AI may not be reducing cognition. It may redistribute it.
Employees may spend less time generating information but more time evaluating, judging and verifying it.
If AI assumes more production, humans may need stronger skills in verification, metacognition, judgment and error detection. And those capabilities cannot be developed merely by giving employees access to AI. They require deliberate practice.
5 questions every CLO should ask
Before integrating AI into a learning activity, L&D leaders can ask five questions.
Is the objective performance, learning or both? If immediate execution is the only objective, extensive cognitive offloading may be perfectly appropriate. If capability development matters, the design may need to preserve more learner participation.
What cognitive process is AI replacing? Is it removing routine search and formatting—or analysis, retrieval, judgment and reflection? Those are not equivalent forms of automation.
Is the removed process part of the capability we want to develop? If employees will later need to make the judgment independently, they need opportunities to practice making it.
Can AI challenge before it answers? Design interactions in which employees first attempt, predict, explain or decide before receiving AI assistance.
Can competence survive without the tool? At selected points, assess whether employees can transfer what they learned to new situations with reduced assistance.
This is not about testing employees to prove that humans are superior to machines. It is about understanding what capability actually remains inside the organization.
What humans still need to learn
For years, L&D leaders have asked how technology can make learning more accessible, personalized and efficient. Those remain important questions. But generative AI introduces another one: When technology becomes capable of performing part of the thinking itself, which parts of thinking do employees still need to practice?
There will be no universal answer. Organizations should probably automate some cognitive work aggressively. Other capabilities—particularly judgment, critical thinking, decision-making, problem framing and the ability to recognize when an answer is wrong—may become more valuable precisely because AI is becoming more capable.
The responsibility of L&D is shifting. It is no longer enough to provide knowledge, improve performance during training or simply determine whether employees can use AI.
The deeper challenge is deciding what humans still need to learn well enough to do, understand and judge for themselves. AI can make work easier. Good learning design must determine when that ease supports capability—and when learning still requires the learner to do some of the work.
