AI and Skill Atrophy - Anti-Pattern

We’re losing skills because we don’t use them, and we’re not growing the developers of the future. What do we give up to LLMs?

When you use an LLM to make a decision you could make without the LLM, you gain a sense of false confidence. The easier it is to prompt an LLM to produce an answer, the less we develop the skill. This affects junior people more than senior people, since they have less overall experience to base their skills on.

In the academic literature, this is called Deskilling. In radiology, AI assistants help detect abnormalities in medical images. In theory, these detect problems faster and don’t get tired. (In practice, there are open questions about their effectiveness.) Too much reliance on AI leaves the radiologist unable to detect on their own. Bad enough for the senior radiologists, but even worse for new doctors. When the AI Assistant was used in the learning process, the new doctor didn’t have as deep an understanding of what they were seeing.

This is a knowledge cliff, and the software industry is headed to one. (Hat tip to Brian Graham for the term.)

NASA has been through this recently; they want to build big rockets again. They have mountains of documentation (literally 2900 cubic feet), but they don’t have the understanding and tacit knowledge to use it.

Effects of Deskilling

• Team Impact: Reduced mentorship and knowledge sharing.
• Long-Term Risk: Novel or unexpected problems can’t easily be solved with AI, leaving people unprepared.
• Cognitive Atrophy: Over-reliance on GenAI increases confidence in the AI while diminishing critical thinking skills.

Brian Graham offers a question list: before automating with AI, ask:

• Knowledge assessment - do we understand the domain well enough to evaluate the AI output?
• Is this a Core competency vs Background noise?
• Failure Planning - Are we monitoring and prepared to reverse course if this decision fails?

The more we use AI, the more critical thinking becomes important.

Pilots have been through this before. With the advent of autopilot, instrument landing systems, automated flight management systems, etc., these systems have reduced the need for certain piloting skills. To combat this, they have increased the amount of training and simulator time for pilots to maintain their certifications.

I did not mention the 2026 Anthropic study: “How AI assistance impacts the formation of coding skills”, because there are enough questions about the design and methodology that I don’t see it as a reliable source.

Image: Agile Pain Relief Consulting, February 2026