Does use of language models degrade cognitive thinking skills? Some studies might suggest this happens.
Researchers found that using language models for code generation can degrade some human coding skills. The effect might happen elsewhere as well.
AI also can also shift where real mastery sits, moving value away from syntax and toward design, debugging, and systems thinking.
Some might point to prior innovations having similar impact:
Studies on calculators show frequent reliance can weaken basic arithmetic and number sense, even as it improves performance on brute calculation tasks.
Similar work on AI tools finds that heavy users tend to offload thinking to the tool, which correlates with weaker independent reasoning and critical thinking.
In a controlled coding study, developers allowed to use AI scored significantly lower on conceptual questions about the library they were using, suggesting inhibited skill formation when too much is offloaded.
So what impact might AI have on programmers?
Routine skills: remembering APIs, writing idiomatic boilerplate, and simple algorithmic patterns are prime candidates for decay, analogous to mental arithmetic.
Conceptual depth: if developers mostly paste and run code, they get fewer reps in reading, tracing, and understanding unfamiliar code paths, which are key to deep fluency.
Error tolerance and debugging: non‑AI users in experiments made more errors but improved debugging skills by fixing them; AI users avoided some errors but learned less from the process.
So value can shift:
Problem framing and specification: the scarce skill becomes formulating precise requirements and constraints that drive the generator toward useful solutions. This parallels how good calculator use depends on setting up the right equation.
Code review and validation: human experts may specialize in reading AI‑generated code for security, correctness, and architecture, rather than writing every line themselves.
System design and abstraction: as low-level implementation is automated, comparative advantage grows in designing architectures, protocols, data models, and failure modes.
Meta‑skills around AI: mastery can shift into knowing when not to offload, how to structure prompts, how to test and monitor generated code, and how to integrate these tools into a development process without hollowing out the team’s competence.
Lessons from calculators and education:
Calculator research suggests the impact depends heavily on how the tool is integrated: thoughtful use can support higher‑order problem‑solving, but uncritical dependence weakens fundamentals.
The same pattern appears with AI coding tools: developers who actively interrogate and adapt AI output retain more skills than those who passively accept suggestions.
Educational responses emphasize sequencing: first build core skills, then introduce tools in ways that force learners to choose when to offload and when to compute or reason themselves.
The bottom line is that some programming skills (though possibly not critical thinking) could diminish as AI is used to generate code.
But we might also note that students in many cases are no longer are taught cursive writing, either, so that skill has atrophied. Is that, on balance, a negative thing or simply a change?
For many working programmers, syntax‑level fluency may atrophy, much as long‑division skills did for most adults, while higher‑level engineering judgment becomes the main skill shift.
A smaller group might deliberately maintain low‑level mastery (in critical infrastructure, compilers, verification) much as some mathematicians maintain strong manual skills, because they need detailed internal models to catch AI’s failures.
