TL;DR
- Recent research gives two readings of AI at work. One finds people producing more while their unaided skill weakens. The other finds people attempting work they could not have attempted before. Both readings are supported, and both describe the same tools.
- The variable that separates them is rarely the technology. Analysis suggests four conditions decide the direction: who owns the consequences of an error, how fast errors surface, whether AI replaces work a person was still learning or extends them into work they never had access to, and whether anyone has tested what happens when the tool is unavailable.
- Erosion of capability is a form of epistemic debt. It accrues in people rather than in records, it shows no warning signal, and it is discovered at the moment it is needed.
- Organizations can check the four conditions without new metrics. Most never do, because productivity dashboards report only one side of the ledger.
Two readings of the same evidence
A recent essay in Open magazine by Dr Vijayakumar Parameswaran Unnithan, Output Is Easy to Count, But Capability Isn’t, makes a careful case that AI can raise performance without raising capacity. It draws on a 2025 Microsoft and Carnegie Mellon study of 319 knowledge workers, in which greater confidence in generative AI was associated with less reported critical-thinking effort. It cites a multicentre study in which experienced endoscopists detected fewer adenomas without AI assistance after AI became routine. It notes a McKinsey survey in which most respondents reported individual productivity gains while a minority reported any effect on earnings. Its central line is hard to dispute: organizations are getting better at measuring what people produce with AI than what they can still do without it.
Data
| Item | Value |
|---|---|
| Individual productivity gain | 80% |
| Positive effect on EBIT | 37% |
The essay introduces useful vocabulary. Developmental safety names the conditions under which capability continues to form. Capability return names what remains when AI assistance is withdrawn. A five-part model distinguishes augmentation, compensation, substitution, atrophy and incorporation.
A second reading of the same period is equally well grounded. Practitioners outside a technical specialty now configure network tunnels, administer servers and build software that would previously have required a specialist. Their output rises, and so does the range of work they can understand, discuss and supervise. The Microsoft and Carnegie Mellon finding that thinking shifts from generation toward verification and integration can be read as a loss. It can also be read as a promotion: verifying and integrating work is what supervisors do.
Both readings are correct. The useful question is what separates them.
Where the erosion account is strongest
The erosion account is strongest when AI performs the work through which a person would otherwise have developed. A junior analyst who never drafts a first-pass model, a new clinician who never builds a differential unaided, a developer who never traces a failure by hand: each delivers output on schedule while skipping the struggle that produces judgment. Unnithan's point that the same task can be tedious for a senior practitioner and formative for a novice is the most actionable observation in the debate. Automating a task is a different decision at different career stages.
The erosion account is also strong on detection. A productivity metric cannot tell AI-assisted performance from human capability, so a decline in the second is invisible until the tool is unavailable, wrong, or facing a case it was never trained on. Automation bias, the documented tendency to accept incorrect AI recommendations when correcting them takes effort, makes this worse. Evaluating an answer is a skill, and it weakens when it is not exercised.
Where the erosion account is incomplete
Three gaps appear when the erosion account is applied to practice.
First, it models AI as removing work that a person would otherwise have done. A large share of AI use does the opposite. It extends people into work they would never have attempted. No developmental experience was lost, because none was available. What is gained includes a durable map of what is possible: knowing that a technique exists, why it is preferable to the alternative, and roughly how its parts fit together. That map survives withdrawal of the tool even when procedural recall does not.
Second, not all unaided skill is worth preserving. Calculators, GPS and spell checking each reduced unaided performance in tasks that society then reasonably stopped requiring. The erosion account needs a way to separate skill that matters as a fallback or as a foundation for judgment from skill that has simply become obsolete. Without that distinction, every decline in unaided performance reads as a loss.
Third, the gap between individual productivity and organizational earnings has explanations other than declining capability. General-purpose technologies typically show a lag between adoption and measurable returns, a pattern economists describe as a productivity J-curve. Time saved by individuals is often absorbed rather than redeployed. These explanations do not refute the erosion account, but they mean the earnings gap cannot serve as evidence for it.
The four conditions that decide the direction
Unnithan's essay states that the outcome "is determined by the conditions of use" without specifying the conditions. Patterns observed in practitioners who report expanding rather than eroding capability suggest four.
Ownership of consequences. When the person using AI also bears the cost of its mistakes, verification becomes a skill that grows. Each confident error that causes a failure teaches what to distrust. When AI output is checked on behalf of someone else's metric, with no personal consequence for a missed error, verification becomes a formality and the skill fades.
Speed of feedback. Capability forms where errors surface quickly: a deployment fails, a calculation does not reconcile, a customer replies. It erodes where AI errors surface months later or never. The endoscopy finding sits in a domain where a missed lesion produces no immediate signal to the practitioner, which is exactly the condition under which reliance deepens unnoticed.
Expansion versus replacement. AI applied to work a person has already mastered, or could never have reached, tends to extend capability. AI applied to work a person is still learning tends to replace the learning. The task is the same in both cases. What differs is where the person stands in relation to it.
A tested fallback. Infrastructure teams do not assume a backup works; they restore from it. Few organizations apply the same discipline to people. A practitioner who can perform a meaningful share of routine work unaided, even slowly, holds a fallback. A practitioner who has never tried does not know whether one exists. Unnithan's proposed capability withdrawal protocol is a version of this test, and it is the condition most often skipped.
| Condition | Tends to expand capability | Tends to erode capability |
|---|---|---|
| Ownership of consequences | The user bears the cost of an AI error | Output is checked for someone else's metric |
| Speed of feedback | Errors surface in minutes or days | Errors surface months later, or never |
| Expansion vs. replacement | Work already mastered, or never reachable before | Work the person is still learning |
| Tested fallback | Unaided performance checked recently | Never tried without the tool |
Capability erosion as epistemic debt
Epistemic debt describes the gap between what an organization believes and what it can still prove, a liability that appears on no ledger and is discovered when a belief meets reality. Capability erosion follows the same structure with one difference. Epistemic debt accrues in records: the evidence behind a decision decays. Capability debt accrues in people: the ability to evaluate the evidence decays.
The two interact in a way that matters for AI adoption. AI is effective at paying down the first kind. It keeps evidence findable, drafts decision rationales and maintains documentation that would otherwise lapse. Used without the four conditions, it accrues the second kind at the same time. An organization can hold a complete, well-written rationale for every decision while the number of people able to judge whether those rationales are sound shrinks. The audit question for epistemic debt asks whether the organization can still prove why it believes something. Under AI, it needs a companion: can anyone still evaluate that proof without the tool that produced it?
A related failure appears when a strong AI practitioner's working method is captured as a playbook for others. The playbook records what the practitioner does. It does not transfer the judgment of what to distrust, which formed through consequences the playbook's readers never experienced. Documentation transfers the record without the reader.
What the conditions imply
The debate about whether AI makes people more or less capable is unlikely to resolve, because it asks the question at the wrong level. The same tool, used by the same person, can extend capability in one domain and replace it in another within the same afternoon. The level at which the question becomes answerable is the condition of use.
Organizations that want to know which direction their adoption is taking can examine four things without building new metrics: who absorbs the cost of AI errors, how long errors take to surface, whether AI is concentrated in work people are still learning, and whether unaided performance has been tested recently. Productivity gains from AI are real and measurable. Whether they are being paid for with capability that will be needed later is equally measurable, but only for organizations that look.