AI Coding Assistants Are Making Developers Worse (And That’s Fine)

Here’s a data point that should make you uncomfortable: experienced developers using AI coding assistants are actually 19% slower than those coding without them. That’s not a typo. A rigorous July 2025 study by METR found that developers with an average of five years of experience on their specific repositories completed tasks slower with Cursor Pro than without it—despite believing they were 20% faster.

The AI coding assistant market is worth $7.37 billion. GitHub Copilot has 20 million users writing 46% of their code. Cursor hit $500 million ARR in 17 months with zero marketing spend. And yet trust in AI-generated code has dropped from 40% to 29% year-over-year, while code quality metrics show AI-assisted code has 1.7x more issues than human-written code. Something doesn’t add up—until you realize that AI coding tools are optimizing for a different variable than we thought.

The evidence for skill atrophy is real

Let’s not pretend this is just curmudgeonly resistance to change. A 2025 study by Microsoft and Carnegie Mellon found that the more developers leaned on AI tools, the less critical thinking they engaged in. High confidence in AI led people to “take a mental backseat.” Researchers described this as “deterioration of critical thinking.”

The numbers are stark. CodeRabbit’s 2025 analysis found AI-generated code has 1.75x more logic and correctness errors, 1.64x worse maintainability scores, and 1.57x more security vulnerabilities. Veracode reports that 48% of AI-generated code contains security vulnerabilities. XSS vulnerabilities are 2.74x more likely in AI code. That’s not a minor quality trade-off.

The employment data is equally concerning. Software developer hiring for ages 22-25 has declined 20% from its 2022 peak. Entry-level tech hiring dropped 25% year-over-year in 2024. Dario Amodei, Anthropic’s CEO, said entry-level jobs are “squarely in the crosshairs” of automation. Marc Benioff announced Salesforce would stop hiring new software engineers in 2025, citing AI productivity gains.

Why this is fine (mostly)

Every new programming tool in history has triggered the same panic. Calculators would make students forget math. IDEs with autocomplete would make developers forget syntax. Stack Overflow would turn coding into copy-paste. High-level languages would disconnect programmers from the machine. None of these concerns materialized as catastrophically as predicted. Each tool shifted the skill ceiling upward while raising the skill floor.

AI coding assistants are following the same pattern. Syntax memorization was never the valuable part of programming. Boilerplate generation was never the creative part. What matters—architecture, system design, understanding trade-offs, debugging production systems—remains firmly in the human domain. If anything, AI tools free developers to focus on these higher-order concerns.

The democratization benefits are substantial. Gartner projects 70% of new corporate applications will be built with low-code or no-code tools by the end of 2025. Development cycles are shortening by up to 80%. Non-programmers can now create functional applications. Y Combinator’s Winter 2025 batch included startups where 25% of founders had codebases that were 95% AI-generated. Andrej Karpathy coined “vibe coding” for this phenomenon, and Collins English Dictionary named it Word of the Year 2025.

The real problem isn’t skill atrophy

The genuine concern isn’t that individual developers will become worse programmers. It’s the “hollowed-out career ladder” problem. If one senior engineer plus AI can do the work of several juniors, companies hire fewer juniors. This creates a pyramid with plenty of seniors at the top, AI tools at the bottom, and a missing middle layer of developers learning the craft.

Software architecture, system design, and production debugging require years of accumulated experience. That experience comes from writing bad code, fixing bugs, and learning from mistakes. If AI handles the grunt work before developers can learn from it, where does the next generation of senior engineers come from? The pipeline that produces experienced developers depends on the grunt work we’re automating away.

This is the same pattern that hit other industries. Automated trading didn’t eliminate finance jobs—it eliminated the entry-level jobs that trained future traders. Design tools didn’t eliminate graphic designers—they eliminated the junior positions that trained future creative directors. The risk isn’t unemployment; it’s a skills gap that takes a decade to manifest.

What actually matters

Google Chrome engineering lead Addy Osmani offers the most pragmatic advice: “Use AI to amplify your abilities, not replace them. Let it free you from drudge work so you can focus on creative and complex aspects—but don’t let those foundational skills atrophy from disuse.” He adds a critical warning: “Never merge code you don’t understand.”

Kent Beck, the creator of extreme programming and TDD, distinguishes between “vibe coding” (accepting whatever AI produces) and “augmented coding” (using AI to accelerate while maintaining quality). He calls AI tools an “unpredictable genie”—powerful but requiring constant supervision.

The METR study revealed something important: developers felt 20% faster even when they were 19% slower. AI-assisted coding had more idle time and felt easier. The cognitive load dropped even as total time increased. Developers aren’t stupid for using these tools despite the productivity paradox—they’re optimizing for sustainable effort, not raw output.

That’s the real trade-off: AI coding assistants make development less exhausting at the cost of making developers less sharp. For many projects and many developers, that’s a trade worth making. For anyone building critical infrastructure or training for senior roles, it’s not.

The uncomfortable conclusion

AI coding assistants are making some developers worse at some aspects of programming. The evidence is clear: code quality degrades, critical thinking diminishes, and the career ladder is hollowing out. And for most use cases, in most contexts, that’s an acceptable trade-off for the productivity gains and reduced cognitive load.

The developers who will thrive aren’t the ones who reject AI tools or embrace them uncritically. They’re the ones who understand the trade-off and choose consciously when to use AI as a force multiplier and when to practice the fundamentals that AI can’t teach. Osmani’s advice cuts to the core: “Don’t worry about AI replacing you; worry about not cultivating the skills that make you irreplaceable.”

The 84% of developers using AI coding tools aren’t wrong. The 46% who actively distrust AI output aren’t wrong either. Both can be true simultaneously. The tools are genuinely useful and genuinely risky. The skill atrophy is real and mostly fine. The career ladder problem is serious and largely unaddressed. Welcome to the messy reality of technological transition.