Why AI Skills Don't Deteriorate—It's All About How You Use Them

AI tools are often blamed for eroding the very skills they’re meant to enhance. But a closer look at recent research suggests the problem isn’t the technology itself—it’s how we interact with it. A groundbreaking study published in January 2026 by Shen and Tamkin challenges the prevailing narrative that AI use inevitably leads to skill atrophy. Instead, the research points to a critical insight: the way we deploy AI tools directly shapes whether they preserve or diminish our expertise.

The Real Impact of AI on Skill Development

The study, titled How AI Impacts Skill Formation, tracked 52 junior software engineers learning to use Trio, an asynchronous Python library. Participants were divided into two groups: one received AI assistance via a sidebar chat tool, while the other coded manually. Both groups completed a warm-up task, implemented two features, and took a comprehension quiz—one they had been explicitly warned about in advance.

The results were striking. The AI-assisted group scored an average of 50% on the quiz, compared to 67% for the manual group—a 17-percentage-point gap (Cohen’s d = 0.738, p = 0.01). While the AI group finished slightly faster, the difference wasn’t statistically significant. The most alarming disparity emerged in debugging questions, where participants struggled to identify errors and understand failures in AI-generated code.

Three key observations emerged from the study that contradict common assumptions:

• Speed gains aren’t guaranteed. Many AI participants spent up to 11 minutes—30% of their allotted time—refining prompts, with some issuing as many as 15 queries. This suggests that for unfamiliar tasks, AI assistance doesn’t always deliver the expected productivity boost.

• Oversight skills degrade first. The quiz heavily weighted debugging, code reading, and conceptual understanding—precisely the competencies needed to validate AI output. This wasn’t a test of rote coding ability but of the critical thinking required to oversee AI effectively.

• AI use alone doesn’t determine outcomes. The study found that how participants used the tool shaped their retention. This nuance is often overlooked in broader discussions about AI’s impact on skills.

Who Is Most Affected—and When?

The research focused on junior developers tackling unfamiliar technologies, but the implications extend beyond this group. The study’s authors emphasize that their findings apply to learning mode—the phase where someone is acquiring new skills—not to experienced developers directing AI on familiar systems.

However, the risk isn’t limited to beginners. Even seasoned engineers can experience skill erosion if their workflows prioritize passive acceptance over active engagement. The authors note that their experimental setup—a chat assistant in a sidebar—may understate the impact of more advanced agentic tools, which could amplify these effects.

Critics argue the study’s artificial constraints—such as a timed quiz and an unfamiliar library—limit its real-world applicability. While the authors acknowledge these limitations, they maintain that the study’s directional findings hold merit, particularly in highlighting how interaction patterns influence skill retention.

Reversing the Trend: Practical Interventions

The study doesn’t just diagnose the problem—it offers a path forward. The authors found that certain adjustments to workflow design can mitigate skill degradation without sacrificing productivity. Their key recommendations include:

• Mandate structured review sessions. Instead of relying on ad-hoc code reviews, teams should enforce deliberate pauses where developers dissect AI-generated output line by line.

• Encourage prompt engineering as a skill. Crafting effective prompts isn’t just about speed—it’s about deepening understanding. Teams should treat prompt composition as a learning exercise, not a productivity shortcut.

• Introduce friction intentionally. Default workflows favor speed, but teams can embed checks—such as requiring manual validation steps—that force deeper engagement with the code.

These interventions don’t require abandoning AI tools. Instead, they reframe AI use as a collaborative process where human oversight remains central. The goal isn’t to reject AI but to design workflows that preserve—and even enhance—our ability to evaluate and improve its output.

The Path Forward: Building AI-Ready Teams

The conversation around AI and skill development has long been polarized between doom scenarios and blind optimism. The reality lies somewhere in between. AI tools are neither inherently harmful nor universally beneficial—their impact depends entirely on how we integrate them into our workflows.

For teams, the challenge isn’t just adopting AI but engineering its use. This means designing workflows that balance speed with rigor, automation with oversight, and efficiency with skill retention. The tools themselves won’t enforce these principles—teams must build them into their processes.

As AI becomes more embedded in software development, the question shifts from whether we should use it to how we use it. The evidence suggests that with the right approach, AI can be a force for both productivity and skill preservation—not a trade-off between the two.