What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
There is a credible reason to ask whether AI coding tools can weaken learning through lack of practice, but there is not evidence that software engineers as a workforce are undergoing cognitive atrophy. A small randomized study found that developers who used AI while learning an unfamiliar Python library scored lower on an immediate comprehension quiz than developers who hand-coded. It did not test long-term retention or experienced engineers’ abilities over time. The practical issue is how developers use assistance—and whether their workflow still makes room to learn, diagnose, and verify.
What does “cognitive atrophy” mean here?
In this context, the phrase is best treated as a concern about skills becoming weaker through disuse, not as a diagnosis established for software engineers. If an assistant routinely supplies code, diagnoses errors, or makes design decisions, a developer may get fewer chances to practise those activities independently. That is a plausible risk; it is not proof of permanent or general cognitive decline.
It also helps to separate two different uses of AI:
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute- Applying an existing skill: using a familiar language or API to complete work the developer already knows how to do.
- Acquiring a skill: learning an unfamiliar concept, library, debugging method, or way to reason about code.
The strongest direct evidence discussed here concerns the second case. It cannot tell us whether AI reduces learning in every task, or whether experienced developers lose skills over time.
What did the developer study find?
A short experiment on learning Trio
Anthropic’s January 2026 randomized study involved 52 mostly junior software engineers. Participants had used Python at least weekly for more than a year, but were unfamiliar with Trio, the Python library used for the task. They worked with two Trio features and then took a quiz covering debugging, reading code, writing code, and conceptual understanding.
The AI-assisted group averaged 50% on the quiz; the hand-coding group averaged 67%. Anthropic reported that the difference was statistically significant (Cohen’s d = 0.738, p = 0.01). The quiz came only a few minutes after the coding task, so this is evidence about immediate performance—not how much participants remembered days or months later.
The AI group was not significantly faster
Participants using AI finished about two minutes faster on average, but the study did not find a statistically significant difference in task time. Some participants spent as many as 11 minutes—30% of their allotted time—composing up to 15 queries. Those figures describe some participants in this particular experiment, not typical usage across software teams.
Recommended Free Tools
The result is therefore not simply “AI makes developers faster” or “AI makes developers worse.” In this learning task, the assisted group’s average quiz score was lower, while its apparent time advantage was not statistically established.
Debugging showed the largest quiz gap
Anthropic reported the largest difference between groups on debugging questions. That matters because developers remain responsible for code they ship, including code generated by an assistant. A developer who can produce a plausible solution but cannot inspect it, explain its behavior, or locate a defect has less effective oversight.
The study also analyzed how participants interacted with AI. Participants whose use was characterized by delegating code or letting AI lead debugging tended to score poorly; those who asked conceptual questions or requested explanations tended to score better. Anthropic cautioned that this qualitative analysis does not show that one interaction style caused the difference. People who already understood the material, for example, may have been more likely to ask conceptual questions.
What the evidence does—and does not—establish
Direct evidence is narrow
The Anthropic experiment supports a limited conclusion: in one short exercise learning an unfamiliar library, AI assistance was associated with lower immediate quiz performance than hand-coding. Its participants were mostly junior engineers, and the study did not measure long-term skill retention, repeated practice, workplace performance, or cognitive change across a career. The authors identify long-term skill development as an open question.
The reviewed evidence does not establish how common skill loss is among software engineers, how quickly it might happen, or whether any decline is durable or irreversible. A result from one task should not be turned into a population-wide estimate.
Developers set different boundaries for different work
Microsoft Research’s July 2026 mixed-methods study examined 448 professional developers’ views about AI autonomy. Most accepted AI producing work under their oversight, but acceptance varied. Developers were less willing to delegate identity-defining, human-facing, and design-oriented work. This study is useful context for how developers draw boundaries around delegation; it did not measure learning, retention, or atrophy.
Human-first assistance is a promising idea, not a proven coding fix
In a September 2026 article, independent author and design practitioner Christopher Noessel described deskilling as a possible result of losing practice in tasks that AI takes over, and overreliance as accepting AI output without adequate scrutiny. Noessel wrote: “When AI assistants absorb tasks humans once performed themselves, they may eliminate the very activities that maintain competence.”
Noessel also reported small exploratory navigation studies. In one cohort of 15 participants using conventional navigation assistance over several sessions, unassisted navigation performance was reported to degrade by approximately 48%. Noessel described this as an order-of-magnitude estimate, not a statistically significant result. In a second small cohort using a “Human Goes First” sequence, the unassisted outcome was roughly 19% better than that cohort’s own assisted baseline. Neither finding is a software-engineering result, and the second does not prove that the approach prevents coding skill loss.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →The design idea is to have the person form and commit to an independent assessment before revealing the assistant’s answer. In Noessel’s words: “In the ‘Human Goes First’ pattern, the AI’s output is withheld until the user has formed and committed to an independent assessment.” Applied to coding, that might mean writing an initial diagnosis or sketch before asking an assistant to critique it. Whether that improves programmers’ learning remains untested in the evidence described here.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to use AI without giving up the learning opportunity
For unfamiliar tools and concepts
- Try first, then ask: write an initial approach, identify what you do not understand, or make a first debugging hypothesis before requesting a complete solution.
- Ask for explanations and critique: use the assistant to clarify an unfamiliar API, explain an error, compare approaches, or review your attempt. These uses preserve more room for active reasoning than handing over the entire task, though the study’s interaction findings are not causal proof.
- Keep debugging in your practice: inspect failure cases, trace relevant code, and test your own diagnosis instead of accepting the first generated fix. The immediate quiz’s largest group difference was on debugging.
- Verify what you own: read generated code, run relevant tests, and make sure you can explain how it behaves and what could go wrong. Assistance does not transfer responsibility for the result.
For familiar work
When a task mainly applies a skill you already have, delegation may save effort. The learning study does not establish that using AI for familiar work causes skill loss. The useful check is whether the workflow still leaves you able to review, debug, and take responsibility for the result—especially when the work is design-oriented or human-facing, areas where Microsoft Research found lower acceptance of AI autonomy.
How should a team judge an AI-assisted workflow?
Speed and output quality are not enough to show that a workflow is supporting developers’ skills. Teams comparing approaches should keep learning outcomes distinct from short-term delivery outcomes.
- Task type: distinguish learning a new concept or API from applying a familiar one.
- Independent practice: note whether developers form an approach or diagnosis before seeing a complete AI answer.
- Comprehension and debugging: check whether developers can read, explain, test, and repair the resulting code.
- Time and quality: measure actual completion time and the quality of the work rather than relying on perceived speed alone. In Anthropic’s experiment, the time difference was not statistically significant.
- Retention: use delayed checks and repeated practice if the goal is to understand learning over time. An immediate quiz cannot answer that question.
- Oversight by task: decide how much autonomy is appropriate based on the work and the developer’s ability to assess the result, rather than treating every task as equally delegable.
For managers, the key distinction is between near-term output and skill formation. Reducing friction may help someone finish a task, but that alone does not show whether they learned the underlying material. Measuring both is more informative than assuming assistance should be either unrestricted or banned.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
