When a firm replaces junior analysts, coders, or reviewers with AI tools and output holds steady, it is tempting to call the substitution a success. Wolfgang Rohde’s April 2026 preprint argues the opposite: the organization has not gained capacity, it has borrowed it from a shrinking pool of senior humans who still know how the work actually works. The paper calls this borrowed competence “capability masking,” and the slow, quiet drawdown that follows “capability erosion.” The distinction matters because one of these problems is visible on a quarterly report and the other is not.
The Efficiency Illusion: What “Capability Masking” Looks Like in Practice
Rohde defines two linked mechanisms. Capability masking occurs when AI-generated output creates the appearance that organizational capability has been replaced, even though skilled-human dependence remains intact underneath. The code compiles. The summary reads well. The slide deck is formatted correctly. Nobody notices that the person who could diagnose why the code is wrong, or whether the summary is missing a key contradiction in the source material, was the last one out the door six months ago.
Capability erosion is the slower process: the degradation of human skills that are difficult to restore once lost. It operates on a different timeline than masking. A team that stops doing code review because the AI handles first-pass checks does not lose its review muscles in a week. But after eighteen months, the tacit knowledge about what a secure module looks like, what a brittle abstraction pattern smells like, and where the sharp edges live in the legacy codebase has quietly atrophied.
The masking-then-erosion sequence is the core argument of Rohde’s paper, and it is worth pausing on the order. Masking comes first. It produces the short-term performance signal that justifies continued substitution. Erosion follows, slowly enough that causation is hard to prove by the time the damage is visible. The paper is a preprint, not yet peer-reviewed, and Rohde frames the mechanisms as a hypothesis supported by evidence rather than established law. The structural logic is sound regardless: if you stop exercising a capability, it degrades.
Why AI-Assisted Coding Still Needs Senior Humans
The paper cites evidence from AI-assisted coding studies to ground its claims in a domain where adoption is already widespread. The findings are consistent: generated code still requires substantial human verification, and its quality remains uneven across correctness, maintainability, and security. Repository-level studies point to further limits in handling broader codebase context, the kind of systemic understanding that distinguishes a working pull request from one that introduces a latent bug three modules away.
This is where the masking-erosion argument becomes concrete. If AI-generated code passes CI and merges without senior review, the masking condition is met: output volume holds, cycle time drops, and the dashboard looks healthy. But the erosion is already accumulating. The senior engineer who would have caught the thread-safety issue, or recognized that the generated abstraction mirrors a pattern that failed in production two years ago, has fewer opportunities to exercise that judgment. And the junior engineer who would have learned from the review comment explaining why the fix was wrong never sees that comment at all, because the review never happened.
The dependency on senior judgment does not disappear when AI enters the workflow. It goes latent. The paper’s argument is that organizations mistake the latency for absence.
The Junior Pipeline Collapse: Short-Term Savings, Long-Term Fragility
Rohde’s most consequential claim is about hiring incentives. Capability masking, he argues, can support hiring restraint, especially of junior roles, while the slower costs accumulate in the background. The mechanism is straightforward: if AI output is good enough to mask the absence of junior contributors, there is no immediate budget signal to hire them. The quarterly numbers look fine. The cost savings are real.
But the junior pipeline is not just a staffing line item. It is the mechanism by which organizations produce the senior judgment that AI still cannot supply. Junior engineers become mid-level engineers by doing review-worthy work and getting reviewed. Junior analysts become senior analysts by handling edge cases, building institutional memory about what failed last time, and developing the taste to distinguish a good model output from a plausible one. If those entry points close, the supply of future senior capability shrinks.
There is historical precedent for the pattern, though the scale differs. The World Bank’s 2019 World Development Report found that new industries and jobs in the technology sector outweigh the economic effects of workers displaced by automation, according to Wikipedia’s summary of automation research. But the same research notes that job losses blamed on automation have been cited as a factor in the resurgence of nationalist, protectionist, and populist politics in the US, UK, and France since the 2010s. The first-order economic metric (net job creation) hides the second-order distributional cost: who loses, whether they recover, and what political friction the dislocation produces. Rohde is making a structural version of the same argument applied inside the firm rather than across the economy.
Who Benefits: Managerial Cost Incentives and Platform Concentration Risk
Rohde identifies two forces driving substitution: managerial cost incentives and national competition. Neither requires the AI to be good enough to replace the worker entirely. It only needs to be good enough to justify the headcount decision on a spreadsheet. If the output quality degradation from lost human review is delayed by 12 to 18 months and difficult to attribute when it arrives, the cost savings are realized now and the capability debt is somebody else’s problem later.
The paper also flags concentration risk. As organizations become more dependent on AI systems they did not build and cannot fully inspect, the locus of capability shifts from the firm to the platform provider. IBM’s automation framing describes intelligent automation as a combination of AI, business process management, and robotic process automation to “streamline and scale decision-making.” Rohde’s critique is that this framing treats the human expertise needed to verify and correct automated output as an externality. The vendor sells the streamlining. The customer absorbs the erosion.
There is an additional wrinkle. arXiv, where Rohde’s paper appears, announced in November 2025 that it would no longer accept computer science review articles and position papers that had not been vetted by a peer-reviewed journal or conference, specifically citing an increase in AI-generated research submissions. The venue is both the distribution channel for this kind of analysis and a case study in the masking-erosion dynamic it describes: the volume of submissions increased, the average quality signal became harder to read, and the platform responded by tightening the gate.
What Capability-Preserving AI Adoption Actually Looks Like
Rohde does not argue against AI adoption. The paper’s contribution is identifying the hidden cost structure, not prescribing abstention. A capability-preserving approach would need to do three things the current substitution model tends to skip.
First, maintain deliberate human review of AI-generated output, even when the output passes automated checks. This is expensive in the short term. It is also the only mechanism that keeps verification skills alive inside the organization.
Second, preserve and fund junior-to-senior development paths that involve producing work, receiving critique, and iterating. If AI takes over the production step, the critique step disappears with it, and the learning pathway collapses. Some organizations are experimenting with AI-assisted mentorship where juniors review AI output and explain their reasoning to a senior reviewer. Whether this produces equivalent skill development is an open empirical question.
Third, track capability debt as explicitly as financial debt. If a team has reduced headcount by relying on AI-generated output, the organization should be able to quantify what human review capacity has been lost, what institutional knowledge is no longer being maintained, and what the replacement cost would be if the AI system degrades, changes pricing, or introduces a reliability regression.
None of this is free. The paper’s point is that the current framing treats it as if it were.
Frequently Asked Questions
How does Rohde’s framing differ from the standard AI job-loss forecasts?
Most displacement coverage, including estimates from OpenAI and Microsoft, focuses on raw headcount projections or the binary debate over whether AI augments or replaces workers. Rohde models the specific pathway by which organizational know-how degrades, independent of headcount. His framework applies to firms that retain all their staff but reassign them away from judgment-intensive work, a scenario headcount estimates miss entirely.
Does the masking-erosion cycle apply outside software engineering?
Rohde uses coding as his evidence domain because adoption there is furthest along, but the mechanism is domain-independent. Legal document review, medical imaging triage, and financial risk modeling all involve AI producing output that passes surface checks while requiring expert judgment to verify. The erosion timeline varies by feedback speed: code breaks in production within days, but a legal reasoning error or a misclassified scan may not surface for months or years, extending the masking window.
What evidence would weaken or disprove the capability-erosion argument?
The hypothesis weakens if longitudinal research shows that practitioners who learn by reviewing AI output develop verification skills equivalent to those who learn by producing work and receiving human critique. No such study had been published as of May 2026. A secondary falsification path: if AI output improves to the point where human verification is genuinely unnecessary in a given domain, the erosion mechanism stops applying there. Rohde does not model this threshold, leaving open what correctness level would make human review redundant.
Could better AI models solve the erosion problem on their own?
Improved models could reduce the volume of human verification required, but Rohde’s framework suggests this may accelerate erosion rather than slow it. If AI output improves from 85 to 97 percent correctness, the remaining errors become harder to catch because reviewers grow accustomed to trusting the system. The paper links this dynamic to platform lock-in: organizations that have ceded verification capability to an external provider lose the internal expertise needed to evaluate whether switching providers or insourcing the work is even feasible.