Recent studies have highlighted a growing concern for employers. Deskilling occurs when seasoned professionals lose proficiency through lack of practice. Never-skilling arises when beginners fail to develop expertise from the start, creating a tougher challenge since firms are already reducing entry-level hires.

The clearest data stems from a randomised trial conducted by Anthropic researchers Judy Hanwen Shen and Alex Tamkin, released in January.

They enlisted 52 mostly entry-level software engineers, provided an AI tool to half the group, tasked everyone with mastering the unfamiliar Python library Trio, and tested retention of key ideas shortly afterward.

The AI-assisted participants averaged 50 percent. The manual group reached 67 percent. Anthropic notes this difference equates to nearly two letter grades and proved statistically meaningful with a p-value of 0.01.

The expected time savings from the tool failed to appear. The AI group finished roughly two minutes quicker, a margin lacking statistical weight, as some spent as long as 11 minutes on prompts, consuming about a third of the session.

Participants gained less knowledge, saw no real speed gain, and performed poorest on debugging, the critical skill when tools falter. The unaided group encountered and fixed issues, mirroring standard debugging practice. The AI group avoided such encounters.

Medical fields have raised parallel concerns. A Nature Medicine Perspective from May, authored by experts at Duke-NUS Medical School alongside colleagues from Harvard, UCL, and King’s College London, described risks for trainees depending on AI in early clinical stages without building essential independent judgment.

It identifies a further issue with limited focus: mis-skilling, where learners adopt AI mistakes without question and treat them as accurate.

The authors remain measured, unlike some reports. They note direct proof from medical education is missing. Their case draws on learning principles and preliminary findings from other domains, including trials like the one from Anthropic.

They recommend a three-stage approach: develop skills without AI first, train calibrated doubt next, and add tools only with oversight.

Usage patterns determine outcomes more than adoption itself. In the Anthropic study, top performers sought conceptual insights or explanations with the code. Lower scorers offloaded entire tasks or relied on the tool for fixes.

Companies are adjusting hiring practices. Gartner forecasts that declining critical thinking will lead half of worldwide firms to demand AI-free evaluations through 2026, signaling distrust in submitted work samples.

Ford has started recalling engineers to correct AI-generated mistakes, illustrating the costs when error-catchers are absent from teams.

The trial has acknowledged constraints. The group was limited in size, the test checked immediate recall rather than long-term retention, and it featured a basic assistant instead of advanced agents. The team anticipates stronger effects from these factors.

The source adds context. Anthropic markets such tools yet released findings showing careless use harms performance. This reflects either notable openness or groundwork for improved training features.

The findings do not advocate banning AI for juniors. They indicate that quick fixes and genuine expertise follow separate paths, contrary to assumptions over the past two years.