AI assistants promise to speed up coding. But they are making junior developers worse at debugging.
A randomized trial with 52 mostly junior engineers revealed a sharp drop in code comprehension. Developers using AI scored 50% on a quiz about their own recent code. Those who hand-coded scored 67%. The difference was statistically significant with a p-value of 0.01.
The biggest gap was in debugging. The AI group didn’t catch errors during debugging, unlike the control group. Speed gains were negligible. The AI group finished two minutes faster, but that difference wasn’t significant.
This suggests juniors lean on AI-generated code without learning how to reason through problems. A May 2026 medical perspective warned trainees relying on AI might never develop the independent reasoning needed for safe practice.
Employers are aware. Gartner predicts half of global organizations will require “AI-free” skills tests by 2026. Ford has begun rehiring engineers specifically to fix AI-induced errors.
AI Hallucinations Fuel Security Risks
AI coding models hallucinate—meaning they confidently invent code and package names that don’t exist. This opens a new attack vector called slopsquatting. Malicious actors register fake packages that AI recommends, injecting malware into supply chains.
Hallucination rates vary from 50% to 82% in some AI models. Even GPT-4o has a minimum hallucination rate of 23%. Analysis of 576,000 code samples found 19.7% contained hallucinations. Proprietary models hallucinate four times less than open-source ones.
Reported vulnerabilities linked to these fake packages are growing at 98% annually—far outpacing the 25% growth in open-source packages. These malicious packages can lurk undetected for months or years, spreading malware widely.
More than 40% of code commits now include AI assistance, and 72% of AI users code daily. Double-checking AI-suggested packages is the only reliable defense against slopsquatting.
Experts Weigh In
Phil Chen, former OpenAI and Google DeepMind engineer, puts it bluntly: “AI models get better at anything you can write a loss function for.” But he warns, “The valuable work of the next decade is everything that can’t be graded within the span of model training.”
Chen urges developers to focus on meaningful problems and build proven excellence. “Relentlessly prioritize your time so that whatever you work on…you focus on problems that you find meaningful,” he says. After all, humans excel at selecting problems for AI to solve and allocating capital wisely.
Anthropic, maker of the AI assistant used in the trial, published research showing careless AI use worsens job performance. The trial itself had limits—small sample, immediate quiz after coding, and use of a sidebar assistant rather than a fully autonomous coder.
Still, the warning is clear: AI coding tools are no substitute for developing core debugging skills. The convenience comes with a cost. Junior developers might be losing the very foundation that makes them good programmers.
Based on
- Never-skilling: the research says juniors using AI never learn to debug — thenextweb.com
- These career skills matter the most in the AI era, says former OpenAI and Google DeepMind employee | Business Insider Africa — africa.businessinsider.com
- Forget typosquatting; slopsquatting is the software supply chain threat created by AI coding tools | VentureBeat — venturebeat.com
