AI Is Deskilling the People Who Need It Most

Strich et al. (2021) found that substitutive AI systems can deskill qualified employees while allowing less qualified ones to mask inexperience. I read that sentence three times. Substitutive AI deskills the people who have the knowledge and empowers the people who do not. That pattern is not a side effect. It is the logical endpoint of a trajectory that Cohen and Levinthal saw coming thirty-five years before ChatGPT.

The deskilling argument is not new. Braverman made it about industrial machinery. The IS field has been circling it since Markus and Robey (1988) listed skill enhancement and deskilling as possible outcomes of technology implementation. What is new is the scale and the asymmetry. Generative AI does not just displace physical tasks. It offloads cognitive work, the judgment calls, the diagnostic reasoning, the pattern recognition that professionals spent years building. And it does so while making the person receiving the output feel like they are still in charge.

Benbya, Pachidi, and Jarvenpaa (2021) laid out the problem directly: automation can render human experts redundant or deskill them. They cited Endsley and Kiris (1995) on the loss of cognitive skills when automation takes over decision-making. The concern is not theoretical. It is empirical. Endsley's work on situational awareness in aviation showed that pilots who relied heavily on autopilot lost the ability to manually fly the plane when the automation failed. The skill degraded not because the pilot chose to forget, but because the cognitive muscles atrophied from disuse. The plane was fine while the automation worked. The pilot became dangerous when it did not.

Kattnig et al. (2024) documented automation bias in AI-assisted decision-making. People follow the automated recommendation even when a reliable source tells them it is wrong. This is not a mistake. It is a systematic tendency. The automation becomes a heuristic replacement for vigilant information processing. Participants without automation aid outperformed those with it. Let me say that again. People who did not have the AI made better decisions than people who did, because the people with the AI stopped thinking for themselves.

This is where absorptive capacity becomes more than an adoption theory. Cohen and Levinthal (1990) defined it as the organizational ability to recognize the value of new external information, assimilate it, and apply it to commercial ends. The critical mechanism is prior related knowledge. You cannot absorb what you cannot recognize, and you cannot recognize it without prior knowledge in that domain. Absorptive capacity is path-dependent and cumulative. Organizations that invest in building knowledge can absorb new knowledge faster. Organizations that never built the knowledge base cannot absorb new information regardless of how much money they spend acquiring it. I wrote about this mechanism in the context of why expensive platforms fail to deliver value.

Now apply that logic to AI deskilling. The organization that offloads diagnostic judgment to an AI system stops exercising the cognitive muscles that built its absorptive capacity in that domain. The prior knowledge does not disappear overnight, but it degrades. The radiologist who stops reading films because the AI flags anomalies loses edge detection speed. The financial analyst who stops building valuation models because the AI generates them loses the feel for when a model's assumptions are wrong. The junior consultant who never learns to structure a problem because the AI provides the framework never develops the capacity to evaluate whether that framework fits. This is not speculation. Strich et al. found that AI systems allow less skilled employees to mask inexperience by referring to the AI, while the AI simultaneously deskills the experts who actually have the knowledge to override it. The system rewards inexperience and penalizes expertise.

I need to double-check this, but my recollection is that Natali, Marconi, and Cabitza (2025) published a systematic review in Artificial Intelligence Review documenting AI-induced deskilling in medical contexts, finding that over-reliance on clinical decision support erodes clinicians' diagnostic abilities over time. I also need to double-check Gerlich (2025), which I recall as finding that cognitive offloading to AI reduces critical thinking self-assessment, though I do not have these papers locally and cannot verify specifics.

The mechanism is what matters, and it is the same one absorptive capacity names. Prior knowledge enables recognition. Recognition enables evaluation. Evaluation enables application. When you remove the need for recognition by letting the AI do it, you do not just skip a step. You atrophy the capacity that made the step possible. The organization that once could evaluate whether an AI recommendation was any good gradually loses that ability. And the loss is cumulative in the wrong direction. Cohen and Levinthal warned that failure to invest in an area of expertise forecloses the future development of capability in that area. Deskilling is deliberate disinvestment, except nobody calls it that. They call it efficiency.

Liu et al. (2025) observed a related finding in their study of human-AI delegation dynamics. High-willingness managers, those ready to delegate to AI, were also the ones who adjusted their willingness when the AI underperformed. They had the knowledge to evaluate, so they could recalibrate. Low-willingness managers just stayed hesitant regardless. The people who could evaluate the AI were the ones already equipped with prior knowledge. The people who lacked that knowledge either trusted blindly or distrusted blindly, because they had no basis for discrimination. This is the two-sided trap of AI delegation that I discussed in why AI pilots fail to scale. The people who should be evaluating the AI are the ones being deskilled by it. The people who should not be delegating to it are the ones who trust it most.

The organizational version of this problem is worse. Zahra and George (2002) reconceptualized absorptive capacity into potential and realized components. Potential absorptive capacity is the ability to acquire and assimilate external knowledge. Realized absorptive capacity is the ability to transform and exploit it. An organization can have high potential capacity, it buys the tools, subscribes to the platforms, hires the talent, and still have low realized capacity because it never built the routines to convert absorbed knowledge into organizational action. I traced this in the AI scaling gap. Most organizations can acquire AI. Few can exploit it.

Deskilling attacks realized capacity directly. When professionals stop exercising diagnostic judgment because the AI handles it, the organization loses the routines that convert absorbed knowledge into applied knowledge. The evaluation step, deciding whether the AI's output is any good, requires the very expertise that deskilling erodes. The organization becomes dependent on a system it can no longer assess. This is not the productivity paradox repeating itself. It is something more specific. The productivity paradox asks whether IT spending produces measurable output gains. Deskilling asks whether AI adoption actively destroys the organizational capacity that makes IT spending productive in the first place.

Gopal et al. (2025) acknowledged the skill atrophy risk directly in the context of research: accepting model-suggested constructs and specifications without justification optimizes metrics over meaning. The shortcut feels productive. The work gets done faster. The capacity to do it without the shortcut quietly disappears. Berente et al. (2021) named deskilling and labor substitution as major ethical frontiers of AI, noting that machines now replace mental functions, not just physical ones, which makes this wave of automation qualitatively different from earlier ones.

I think the IS field has half of the diagnosis. We have absorptive capacity for explaining why organizations fail to learn. We have automation bias for explaining why people trust systems they should not. We have deskilling research for documenting that expertise erodes when it is not exercised. What we do not have is a unified account of how these three mechanisms combine into one downward spiral. Offloading to AI reduces exercise of judgment. Reduced exercise atrophies expertise. Atrophied expertise reduces the capacity to evaluate AI output. Reduced evaluation capacity increases over-reliance. Over-reliance further atrophies expertise. The loop tightens with each cycle.

Cohen and Levinthal's path dependence was meant to explain why organizations with more knowledge learn faster. It also predicts, by inversion, why organizations that stop learning fall further behind. They cannot see what they are missing, so they do not invest in what would help, so they miss more. AI deskilling accelerates that cycle by making the offloading feel productive. The report is written. The diagnosis is generated. The decision is recommended. The work looks done. The capacity to do it without the AI is what disappears.