AI-induced skill erosion

AI-induced skill erosion is the slow loss of practiced judgment, craft, confidence, or professional identity when AI absorbs enough of the work that humans stop exercising the capabilities the system still depends on. The construct is narrower than automation-complacency: complacency describes reduced attention or effort in the moment; skill erosion asks whether repeated reliance changes the worker’s capability over time.

2026-ehsan-future-workers gives this topic its clearest longitudinal evidence. In a year-long radiation-oncology study, the authors describe “intuition rust”: clinicians approved AI-supported work more quickly, practiced manual skills less, and worried about becoming “bystanders” in their own practice. The important claim is not simply that AI can make experts lazy. It is that short-run operational gains can hide asymptomatic changes in expertise, agency, and identity until they harden into chronic harms.

2025-kosmyna-brain-chatgpt-cognitive-debt adds adjacent early evidence from educational writing rather than workplace expertise. In repeated essay-writing sessions, LLM users showed weaker EEG connectivity, more homogeneous language, lower ownership, and poorer recall of text they had just written than brain-only writers. That does not prove professional skill erosion, but it sharpens the possible accumulation mechanism: repeated offloading may reduce task engagement and memory before any durable capability loss is visible.

2026-shen-ai-impacts-skill-formation supplies a more direct short-horizon causal test of skill formation. In a randomized programming task, AI assistance reduced post-task skill evaluation scores by 17% without significantly improving completion time; the largest gaps appeared in debugging, which is exactly the capability workers need if they are expected to supervise AI-written code. This does not show long-run erosion by itself, but it gives the topic a plausible immediate precursor: AI-mediated work can produce the artifact while weakening the learning episode that would otherwise maintain future oversight capacity.

2026-edwards-developer-productivity gives a qualitatively distinct senior-practitioner account of that precursor. Some interviewed developers reported forgetting basic syntax and worried that AI’s leverage could reduce the foundational practice needed to direct, evaluate, and recover from generated code; other accounts described AI as a learning partner that supported reflection and higher-level architectural thinking. The source therefore reinforces that offloading is not itself erosion: the risk depends on what practice disappears and whether the redesigned role retains active specification, verification, and learning.

2025-randazzo-cyborgs-centaurs-self-automators adds a workflow-mode precursor to skill erosion. Abdicated/Self-Automator consultants produced fast, polished work by handing over much of the problem-solving and evaluative work to GenAI, but they did not develop either task/domain expertise or GenAI-related expertise during the episode. The warning is not that delegation is always bad; it is that output quality can hide whether the human side of the workflow is still practicing the capabilities future oversight depends on.

2026-lovett-tragedy-cognitive-commons shifts the unit of analysis from an individual worker’s capability loss to the profession’s capacity to regenerate expertise. Its Cognitive Commons framework proposes that independently rational firms can remove entry-level developmental work and thereby deplete the shared reservoir of Internalized Mastery on which later AI validation, crisis response, and succession depend. The article’s Validation Tether makes the bridge explicit: Distributed Mastery in orchestrating AI can produce plausible output, but substantive review still rests on the deep domain knowledge that developmental pipelines are meant to build. This is a conceptual prediction, not evidence that profession-level depletion has occurred.

2026-ruttenberg-cognitive-debt-ai-research extends the cognitive-debt mechanism into a threshold model. Skill erosion risk is not only a slow average decline in capability; it may remain invisible while outputs stay high, then surface when attentional erosion, effort displacement, and affective depletion cross a capacity threshold. 2026-ruttenberg-neurodivergent-expertise-ai-work adds the individual-differences contribution: autistic and neurodivergent professionals may detect unsustainable AI offloading earlier because they already monitor sensory, attentional, and recovery constraints that neurotypical work systems often ignore.

The paper’s response frame is also useful. “Dignified Human-AI interaction” treats agency, expertise, and self-worth as design requirements, not soft afterthoughts. “Sociotechnical immunity” names the protective system: early-warning signals, containment actions, and recovery routines that help workers detect, interrupt, and recover from AI-induced erosion while still serving quality goals.

2026-mehra-agents-that-teach supplies a developer-specific intervention hypothesis rather than erosion evidence. It calls the hidden gap between AI-executed code changes and a developer’s understanding Knowledge Debt, then proposes preserving practice through contextual, selective, asynchronous learning prompts and closed-loop comprehension checks. This gives the topic a more operational countermeasure than generic exhortations to keep humans involved: retain the reasoning, alternatives, and learning feedback that make later debugging and recovery possible.

Connections

• Relates to ai-mediated-work-experience because skill erosion is one pathway through which AI changes competence, identity, autonomy, and meaning at work.
• Relates to automation-complacency because reduced scrutiny can be the short-run behavior that accumulates into long-run skill loss.
• Relates to human-ai-collaboration because collaboration designs that keep humans nominally accountable but remove meaningful practice may weaken the human capacity they rely on.
• Relates to work-redesign because workflows need deliberate practice, feedback, escalation, and recovery routines if they are supposed to preserve expertise rather than merely route work through humans.
• Relates to critical-thinking because stewardship and verification are not just output-quality behaviors; they are also ways workers continue to exercise judgment.
• Relates to human-ai-co-creation-modes because some co-creation modes preserve practice and skill-building while others let the human bypass both.
• Relates to cognitive-debt because debt names the accumulating liability between repeated AI bypass/offloading episodes and later durable capability loss.
• Relates to automation-complacency because cognitive debt may begin as reduced effort or ownership in the immediate task before it becomes measurable skill erosion.
• Relates to digital-work-strain because capability loss can be entangled with overload, recovery failure, silent burnout, and productivity systems that keep outputs visible while cognitive sustainability declines.

Contradictions & open questions

• 2026-edwards-developer-productivity makes both sides visible in a small senior-developer sample: AI can support reflective learning and architectural thinking while reducing practice with foundational implementation. It cannot establish whether these reports become measurable long-run capability loss or whether specific workflows preserve both leverage and fluency.
• 2026-ehsan-future-workers is strong qualitative longitudinal evidence from a high-stakes expert domain, but it does not estimate the causal rate of skill decay or show how broadly the mechanism generalizes across jobs.
• Skill erosion and productive offloading can look similar in the short run. The open measurement problem is how to distinguish healthy delegation from loss of capability before failure becomes visible.
• The wiki needs evidence on which design interventions actually preserve expertise without imposing wasteful friction or romanticizing manual work that can safely be automated.
• 2025-kosmyna-brain-chatgpt-cognitive-debt uses an educational writing task and a small fourth-session crossover sample, so its strongest contribution is mechanism generation rather than direct evidence about AI-mediated professional skill decay.
• 2026-shen-ai-impacts-skill-formation is causal but short-horizon and task-specific; it shows impaired skill formation after one AI-supported programming task, not yet durable workplace skill erosion over months or years.
• 2025-randazzo-cyborgs-centaurs-self-automators does not show long-term erosion; it shows that one mode of GenAI use can fail to build expertise during the task, which is a plausible but not sufficient precursor to erosion.
• 2026-ruttenberg-cognitive-debt-ai-research and 2026-ruttenberg-neurodivergent-expertise-ai-work are theoretical and adjacent-evidence-based; their threshold models are useful for measurement design, but they should not be treated as demonstrated AI-induced erosion among knowledge workers until the proposed longitudinal studies exist.
• 2026-mehra-agents-that-teach proposes a plausible software-development countermeasure but supplies neither a validation of Knowledge Debt nor evidence that its SHIELD prototype preserves expertise, reduces later recovery costs, or avoids surveillance and interruption burdens.