Seventeen percent. That's how much worse software developers performed on conceptual assessments after delegating coding tasks to AI, compared to developers who learned without it. They could produce working code. They just couldn't explain how it worked or debug it when it broke.
That finding, from a 2026 study by researchers Shen and Tamkin, is part of a growing body of work drawing a sharp line between two different problems with AI dependency. Adults who offload cognitive tasks to AI experience what researchers call "cognitive atrophy" - existing skills weaken from disuse, but the underlying capacity remains. Children who grow up delegating thinking to AI face something potentially worse: "cognitive foreclosure," where foundational reasoning skills never develop in the first place.
Skills That Rust vs. Skills That Never Form
A 2025 study by Gerlich found that participants over 46 showed higher critical thinking scores with lower AI reliance, while those aged 17 to 25 showed the opposite pattern. The older group had decades of built-up reasoning skills that persisted even as they adopted AI tools. The younger group appeared to be substituting AI output for their own analytical development.
This maps directly to how people use tools like ChatGPT or Claude in daily work. An experienced marketer who asks AI to draft a campaign brief can evaluate whether the output is good because they've written hundreds of briefs themselves. A marketing student who has AI write every brief from day one never builds that evaluation capacity. They can't distinguish good output from bad because they have no internal benchmark to compare against.
The Homogenization Effect
There's a compounding problem beyond individual skill loss. Research by Sourati, Ziabari, and Dehghani in 2026 found that large language models (the technology behind ChatGPT, Claude, and similar tools) tend to homogenize reasoning and expression, converging toward what they describe as "Western, educated, mainstream norms."
This has practical consequences. A team of five people all using AI to draft strategy documents will produce more similar outputs than five people thinking independently. The diversity of approaches that makes collaboration valuable gets compressed.
Who Can Actually Check AI's Work?
The most actionable concern is what researchers call the "AI audit problem." Adults can evaluate AI output against existing expertise - a doctor reads an AI-generated diagnosis and catches an error because they've examined thousands of patients. Children and novices lack the foundation to perform that check.
This creates an uncomfortable reality for AI productivity tools. The people who benefit most from AI assistance - experienced professionals with deep domain knowledge - are exactly the people who need it least. The people who might seem to gain the most from AI doing their thinking - students, junior employees, career changers - are the ones most at risk of having their development stalled by it.
For anyone managing a team that uses AI tools daily, the research suggests a specific policy question: not whether to use AI, but which tasks should remain human-only for skill development purposes. Delegating execution is relatively safe for experienced workers. Delegating the learning process itself is where the damage compounds.