“Learning debt”: AI can make output look competent before capability e – Good Machine

TL;DR

A new survey of 1,200 US workers found 59% use AI at least sometimes for tasks they were not trained to do; 29% say they have delivered work they could not fully explain. Those are self-reports from a learning-platform vendor, not proof that AI causes poor performance. [TalentLMS, Tier 2; HR Dive, Tier 2]
The important idea is sound even if the headline statistics are treated cautiously: a completed task is no longer reliable evidence that the person who submitted it understands the work.
This is not an argument to ban AI or romanticise manual work. It is an argument to distinguish delegating toil from delegating the repetitions that build judgment.
For workers, the practical test is simple: can you explain, check, modify and recover the work without the tool? If not, the output may be useful, but it is not yet your capability.

The performance–capability gap

The new phrase doing the rounds in workplace-learning circles is learning debt: the backlog created when a job changes faster than the person doing it has time and support to learn.

TalentLMS’s July report gave the phrase its viral spark. In its June survey of 1,200 US employees aged 25–64, 41% said their role had evolved faster than their employer’s training could keep up. Thirty-seven percent said AI had made them look more competent than they were. Nearly half said they had kept quiet about not knowing how to do a work task. [TalentLMS, Tier 2; HR Dive, Tier 2]

The cleanest reading is not “AI makes workers incompetent.” The survey cannot establish that. It is vendor-sponsored, US-only and based on self-reporting. It measures perceptions and behaviour, not independently assessed skill or causation.

But it identifies a real workplace failure mode: performance is visible; understanding is not. Before generative AI, a person who could produce a spreadsheet model, code change, customer response or board summary had usually done enough of the work to reveal some of their underlying ability. Now a credible-looking result can arrive before the person has built a mental model of the task.

That is the debt. The deadline is paid. The learning bill is deferred.

What happened — and what it does not prove

The survey reports that 62% of respondents use workarounds when they lack the skills or training for a task, while 59% use AI at least sometimes for work they were not trained to do. Workers who said they were falling behind on learning also reported more AI use for untrained tasks. [TalentLMS, Tier 2]

Those figures matter as an early-warning signal, particularly because the workplace incentives are obvious. If a tool can produce a passable first draft in seconds, asking for training can feel slower, riskier and more conspicuous than simply submitting the draft.

Still, correlation is not causation. A worker may use AI for an unfamiliar task precisely because the work is changing quickly; AI is then a symptom of the training gap as much as a cause. The report’s claim that workers falling behind are nearly six times more likely to make preventable mistakes is also cross-sectional self-reporting, not a controlled estimate of AI’s effect.

That caution strengthens the analysis. The news is not that one survey has settled the deskilling question. The news is that the survey has put a sharp, usable name on a problem that independent research and labour-market evidence already make plausible.

Why this is more than a training issue

A useful distinction:

AI use pattern	Immediate result	Long-term effect
Automation of toil — formatting, transcription, routine retrieval, first-pass sorting	Faster work	Usually benign if outputs are checked and the underlying task has little developmental value
Augmentation of judgment — critique, counterarguments, simulation, explanation, practice	Better decisions and faster learning	Positive when the person remains accountable for reasoning and verification
Substitution for understanding — accepting a solution one cannot explain, test or adapt	Convincing output	Learning debt: apparent competence rises while independent capability may not

The third row is the risk. It is also the easiest one to miss in a dashboard. Output volume goes up. Cycle time falls. The review queue may look healthy. Then an edge case appears, the model is unavailable, a customer asks “why?”, or a regulator asks for the reasoning. The organisation discovers that it has automated the easy repetitions through which people used to learn to handle the hard cases.

This is why the viral claim deserves attention beyond HR. It touches quality assurance, risk management, professional accreditation, succession planning and security. A team that cannot independently inspect its AI-assisted work has not merely taken on a learning problem; it has taken on an operational dependency.

The evidence says design matters

The more useful research question is not whether AI helps or harms learning. It is what the system asks the human to do before, during and after assistance.

A 2025 study in Proceedings of the National Academy of Sciences found that unrestricted generative AI could improve students’ practice performance while harming learning on an unaided exam. A version designed as a tutor—offering hints rather than answers—avoided that penalty. [PNAS, Tier 1]

A separate three-month chess study, reported by Wharton researchers in 2026, found students with on-demand AI help gained less than half as much as students receiving controlled, timed assistance: 30% versus 64%. The mechanism was not mysterious. On-demand help displaced productive struggle: the effort of trying, failing, comparing options and correcting a mental model. [Wharton, Tier 2]

These studies are in education, not offices. Do not pretend that a classroom result automatically quantifies workplace deskilling. But they do establish an important design principle: the same model can either replace a learner’s thinking or structure it.

That principle fits labour-market evidence. The ILO says the transition is chiefly one of job transformation, not instant job extinction, and argues that whether AI improves work depends heavily on training, guidance and workers’ involvement in implementation. Its 2026 lifelong-learning report adds that employers increasingly seek rounded profiles: technical competence combined with digital literacy, critical thinking and social skills. [ILO, Tier 1]

The implication is blunt. A prompt library is not a skills strategy. Nor is access to an AI assistant.

Who gains, who carries the risk

Workers under time pressure gain immediate relief. AI lets someone get unstuck, produce a draft or complete an unfamiliar task. That can be genuinely empowering, especially where formal training is scarce.

Managers gain surface-level throughput and may lose sight of where coaching is needed. A high-quality-looking deliverable can suppress the very questions that would reveal a gap.

Experienced practitioners become the invisible backstop. They review, correct and recover work that less experienced colleagues can now generate more quickly. Their expertise becomes more valuable—but harder to replenish if junior staff skip the foundational work.

Early-career workers face the sharpest trade-off. AI can provide access to explanations and practice that earlier cohorts did not have. It can also remove the lower-stakes repetitions through which confidence and judgment are built. The outcome depends on whether their work still contains supervised practice, not on whether they use AI at all.

Training vendors benefit from the learning-debt frame. That does not invalidate the diagnosis; it means the source must not be allowed to grade its own prescription.

The non-obvious connection: AI governance is becoming competence governance

Most AI controls are framed around data access, privacy, model risk and hallucinations. Those matter. Yet a quieter control question is emerging: who in the workflow can detect a wrong answer without asking the same model that produced it?

This is especially consequential in accounting, banking, professional services and software. US job-posting data show fast growth in demand for AI skills in those sectors, including Microsoft Copilot, prompt engineering, retrieval-augmented generation and MLOps. [Bipartisan Policy Center/Lightcast, Tier 2]

The value of these skills is not “can operate the tool.” It is “can specify the task, inspect the result, recognise failure modes and take responsibility for the decision.” That is a higher bar than tool adoption, and it makes human domain knowledge more—not less—important.

Re-aim: what workers and managers should do

For individual professionals

Use the explain–check–change test. Before treating AI-assisted work as a skill, be able to explain its logic in plain language, check its key claims against source material, and change a meaningful part without re-prompting from scratch.
Keep one deliberate no-AI repetition. For a capability you need in your career—financial modelling, writing, coding, analysis, client diagnosis—perform a bounded version unaided each week. The point is calibration, not purity.
Ask AI for friction, not just answers. Prompts such as “ask me three questions before proposing an answer,” “show the assumptions and failure modes,” and “give me a practice case, not the solution” turn the tool into a coach rather than a substitute.
Keep an evidence log for consequential work. Record the prompt, sources checked, material edits and the final human decision. This improves both learning and accountability.

For managers and team leads

Do not infer capability from deliverable quality. Add lightweight demonstrations: a five-minute walk-through, a counterexample request, a short live modification or a review of sources and assumptions.
Map tasks by learning value. Automate repetitive work with little developmental value. Preserve supervised exposure to recurring decisions that build judgment, particularly for early-career staff.
Train at the moment work changes. The training cannot arrive six months after a new tool or process becomes normal. Pair rollout with role-specific examples, practice tasks and an escalation path.
Measure capability separately from usage. Tool adoption, completed modules and output volume are not evidence of understanding. Use scenario-based assessments, quality sampling and error-pattern reviews.
Reward early disclosure. If asking for help carries a competence penalty, people will hide gaps and use AI to paper over them. Treat a surfaced unknown as operational intelligence, not a performance failure.

Signal vs. noise

The viral version of this story says: “AI is making everyone fake competence.” That is too broad.

AI also enables rapid feedback, personalised tutoring, accessibility support and useful first drafts. The independent learning evidence does not support a ban; it supports constraints and pedagogy. In well-designed systems, AI can preserve effort where effort is formative and remove it where effort is merely wasteful.

The sturdier claim is narrower: when a system rewards completion but never tests comprehension, AI makes it easier for an organisation to confuse output with capability. That is not a future problem. It is a design choice being made now.

Uncertainty ledger

The survey: TalentLMS surveyed 1,200 US employees in June 2026. Its numerical findings are self-reported and sponsor-funded; they should not be read as causal estimates or global prevalence rates.
Workplace evidence: Direct, long-term causal evidence on AI-assisted deskilling in real workplaces is still limited. Education studies offer a mechanism, not a one-for-one forecast for every profession.
Skill change can be healthy: Some old skills should decline. The relevant question is whether the displaced skill is low-value toil or the judgment required to supervise, adapt and recover the work.
What would change this analysis: Rigorous longitudinal workplace studies that independently assess unaided performance, quality, retention and incident recovery across different AI-workflow designs.

Bottom Line

AI can make work look competent before the worker is competent. That is useful in a pinch and dangerous as a system of development. The winning organisations and careers will not be those that use AI least; they will be those that make AI-assisted output auditable, challengeable and convertible into real human judgment.

Sources

Tier 1 — International Labour Organization, Lifelong learning and skills for the future (May 2026); Generative AI at work: What it means for jobs in Europe and beyond (2025).
Tier 1 — Proceedings of the National Academy of Sciences, “Generative AI without guardrails can harm learning: Evidence from high school mathematics” (2025).
Tier 1 — OECD, Generative AI and the SME Workforce (2025); Artificial Intelligence and the Labour Market in Korea (2025).
Tier 2 — TalentLMS, Learning Debt Report 2026 (15 July 2026). Primary survey source; 1,200 US employees, self-reported results.
Tier 2 — HR Dive, “AI may conceal growing ‘learning debt’ for fast-changing roles” (16 July 2026).
Tier 2 — Wharton School, “When Does AI Assistance Undermine Learning?” (2026), reporting a three-month randomised chess-club study.
Tier 2 — Bipartisan Policy Center, “Industries with the Fastest Growth in Demand for AI Skills” (15 July 2026), using Lightcast job-posting data.
Tier 2 — AI & Society (Springer Nature), “AI deskilling is a structural problem” (published online 5 November 2025).