GenAI Skill Erosion Is a Governance Problem, Not a Technology Verdict
Over-reliance on generative tools can dull writing, coding, and critical thinking — but the evidence points to use conditions, not the models themselves, and that is where governance must work.
The programme director asked a question the assessment rubric could not answer: if a student used GenAI on the capstone, which parts of the submitted work still reflected their reasoning? Several people looked at the policy line about "original work," then at the syllabus that permitted open tool use, and nobody could point to a method that would verify human reasoning separately from polished output. That gap is where this piece lives.
Most institutional responses to GenAI skill erosion still treat the tool as the independent variable — ban it, detect it, or bless it with a principles paragraph. The literature is moving elsewhere. Conditions of use increasingly look like the variable that matters, which means the honest policy question is not "GenAI yes or no" but "what can we verify about human skill development when assistance is universal?"
Skill Erosion Tracks Substitution, Not Access
It helps to separate two things that get collapsed in public debate. Generative AI can amplify cognition when use is scaffolded — guided problem-solving, argumentation-based writing, reflective cycles that keep the learner doing the reasoning work. It can substitute for cognition when use is passive: paste the draft, accept the answer, skip the verification beat. A recent systematic review of 89 studies (2024–2026) on higher-order cognitive skills documents both pathways and finds that over-reliance, reduced analytical autonomy, and cognitive offloading cluster where GenAI use is unstructured.
If that distinction is new, the one-sentence version worth keeping is: skill erosion tracks substitution, not access. The same review is explicit that documented risks "are not inherent to the technology but to the conditions of its use" — which is exactly the sentence a governance framework should be built to operationalise, not quote on a slide.
Granted, some tasks genuinely benefit from automation of low-level steps. The governance error is treating all GenAI use as one category — as if a scaffolded lab exercise and an unchecked overnight draft submission were the same intervention.
Productivity Can Mask Learning Decline
Practice, across disciplines, is messier than a single moral headline.
In higher education, a systematic critical review of 54 studies (2023–2025) alongside international policy documents describes a field of fragmented rules, uneven detection, and disciplinary divergence — productivity and accessibility gains sitting beside risks to originality, critical thinking, and epistemic justice. The picture is not "students are ruined"; it is "institutions are governing the wrong object."
In professional and general populations, cognitive offloading — delegating memory, analysis, or verification to external tools — mediates the relationship between frequent AI use and weaker critical thinking in survey and interview work with 666 participants. Younger cohorts show higher dependence; higher educational attainment moderates the pattern. Age-blind, one-size prohibition is poor institutional design even if you accept the correlation.
Coding offers a sharper behavioural signal. A quasi-longitudinal study of five years of graduate submissions in an unchanged cloud-computing assignment reports that after ChatGPT's release, final submissions grew longer while average edit distances between revisions increased and average score improvements fell — patterns consistent with over-reliance and weaker learning trajectories, though individual misuse cannot be proven from behaviour alone. In plain terms: the repository got bigger; the learning curve got flatter. Employers still grading "output quality" may be certifying polish, not craft.
Writing and evaluation face a related failure mode researchers call metacognitive laziness — learners delegating not just the task but the self-monitoring that tells them whether they understand the answer. If that term is load-bearing here: it names the habit of skipping the internal check, not merely asking the model for a faster draft. That is a different erosion from spelling or syntax. It is the atrophy of the reflex to doubt.
Verification-First Governance — What Institutions Can Test
The OECD's education work on AI and skills frames the comparison correctly: policy needs to ask where AI complements human skill and where it replaces it, and reshape assessment accordingly — including forward-looking measurement such as the planned PISA Media and AI Literacy assessment. UNESCO's global GenAI guidance pushes toward human-centred regulation, data protection, age limits, and pedagogical validation of tools in education and research. Its 2024 AI competency frameworks for students and teachers add the competence layer: critical thinking, ethics, agency in relation to AI systems.
In plain terms: governments are sketching the guardrails; institutions still have to build the attestation chain inside the classroom and the hiring pipeline.
Institutions can translate that into mechanisms a sceptical auditor could inspect:
- Friction by design. "Write first, critique second" protocols — produce a human draft before model assistance — recreate the verification beat that passive offloading skips. The point is not nostalgia; it is to make substitution visible in the workflow.
- Assessment of competence, not artefact. Timed explanation, oral defence, in-environment problem solving, and version-history review measure what a polished PDF cannot. If your integrity model is detection-only, you are governing outputs, not skills.
- Structured access tiers. Course-level rules for when tools are permitted, with what logging, mirror structured access thinking in frontier-model governance — controlled reach with terms a reviewer can inspect. If that phrase is unfamiliar: external parties get defined capability under agreed rules, not a blank licence to the whole stack.
None of this requires pretending GenAI away. It requires specifying which cognitive work must remain human-attributable and building evals that test that claim under conditions harder to fake with a one-shot prompt.
Where Governance Designs Fail — a Stress Test
Any design worth adopting should survive ordinary failure, incentive pressure, and drift.
Ordinary failure: An "AI-allowed" policy with no scaffold defaults to substitution because substitution is faster.
Empty permissiveness is not neutral; it selects for offloading.
Incentive bends: When throughput is rewarded — publication counts, ticket closure rates, pass rates — staff and students rationally delegate reasoning. Governance that ignores throughput metrics governs fiction.
Reality drift: A syllabus approved in September meets agentic tools by January. Static bans age into irrelevance; static permission ages into scope creep. Programmes need review cadences tied to capability change, not annual PDF refresh.
The fair objection here is equity: scaffolded, friction-rich GenAI use takes time and sometimes paid access. That objection is right about cost. Where it stops holding is the assumption that a ban is costless — it often externalises cost onto hidden use and weaker learners without institutional support.
What Institutions Should Verify Next
Skill erosion from GenAI over-reliance is real enough in the record to take seriously — writing, coding, design thinking, critical evaluation, and the meta-skill of checking your own understanding. It is not uniform, not destiny, and not fixed by rhetoric.
Most programmes do not need a manifesto. They need an honest scope review of where substitution is currently default, an assessment method somebody will defend under scrutiny, and one verifiable pilot that proves human competence can still be measured in the age of universal assistance.
Which course or team in your organisation still certifies skill by polished output alone — and what would a verification-first redesign of that assessment actually test?