Why so many programmers still resist AI coding tools

AI coding tools are everywhere. Most professional developers now have some form of AI assistant wired into their editor, their IDE, or their browser. Yet at the very same time, trust in those tools has never been lower. On the surface, programmers complain about bugs, security, and bad study results. Underneath, something much bigger is going on.

The paradox: everyone uses AI, few actually trust it

Recent surveys show that around 84% of professional developers now use AI in their workflow. That’s the highest adoption rate ever recorded. But when you ask those same developers whether they trust AI to be accurate, only about 29–33% say yes. More developers actively distrust the tools they use every day than trust them.

This creates a strange reality: AI is becoming mandatory in practice, but emotionally and professionally, many developers are still pushing back. To understand why, it helps to look at the five most common reasons programmers give for resisting AI—and what those reasons are really hiding.

Reason 1: “AI code is almost right, which is worse than wrong”

The top frustration in recent developer surveys is simple: AI-generated code that’s almost correct, but not quite. About two-thirds of developers say this is their biggest issue with coding assistants. Nearly half say debugging that code often takes longer than writing it from scratch.

Large-scale analysis backs this up. One study of over 200 million lines of code from 2020 to 2024 found:

• Duplicated code blocks increased roughly eightfold.
• Code churn—new code rewritten within two weeks—rose from about 3% to nearly 6%.

The pattern: AI doesn’t always reduce work. It shifts it. Instead of spending time writing code, developers spend more time reviewing, refactoring, and cleaning up what the AI produced. For senior engineers who care about long-term maintainability, that can feel like a step backward.

If you’re interested in how this tension plays out in real engineering practice, it pairs well with the argument that AI won’t replace real software engineering, but will change how serious coding gets done.

Reason 2: Security vulnerabilities baked in by default

Security researchers have been raising alarms about AI-generated code for a while. Empirical tests on popular tools like GitHub Copilot found that a large share of suggested snippets contained known security weaknesses. In some studies, nearly 30% of Python snippets and about a quarter of JavaScript snippets had issues like:

• Insufficient randomness
• Code injection vulnerabilities
• Cross-site scripting (XSS)

These aren’t obscure edge cases. They’re classic vulnerabilities that any experienced developer is trained to avoid. Yet the AI tools can surface them as “good enough” defaults.

By mid-2025, AI-generated code was estimated to be introducing more than 10,000 new security findings per month—a tenfold increase in just six months. For teams shipping production systems, the logic is clear: why would you rely on a tool that makes it easier to ship vulnerabilities by accident?

Reason 3: Fear of a generation of “permanent beginners”

Another concern, especially among senior engineers, isn’t about today’s bugs—it’s about tomorrow’s skills. Several studies suggest that heavy reliance on AI for code generation can weaken understanding.

One research effort found that developers using AI scored 17% lower on comprehension tests when learning new libraries. Another study of junior engineers showed a split:

• Those who used AI mainly for conceptual questions scored above 65%.
• Those who delegated actual code writing to AI scored below 40%.

The fear is that AI can help people produce working code without ever really understanding how or why it works. Over the long term, that could create a generation of developers who are productive only as long as the AI is functioning—and helpless when it fails.

In that world, the rare engineers who can still design, debug, and reason deeply about systems without constant AI hand-holding become the exception. Their skills turn into a kind of luxury good inside the industry.

Reason 4: Intellectual property and licensing worries

Some of the resistance to AI coding tools never makes it into public blog posts or conference talks, because it lives in legal memos and internal policy documents. But it’s real.

In 2022, a class action lawsuit was filed against GitHub, Microsoft, and OpenAI. The core allegation: they trained Copilot on billions of lines of open source code without properly honoring the licenses attached to that code—licenses like GPL, MIT, and Apache that require attribution or impose conditions on reuse.

The plaintiffs argue that Copilot effectively strips attribution and sells derivative code back to companies, including code that may have been authored by the very developers now using the tool. While many of the claims have been dismissed, some, including a Digital Millennium Copyright Act claim with potentially huge damages, are still alive.

For developers who have spent years contributing to open source under specific licenses, this isn’t an abstract legal puzzle. It feels like their work may have been quietly absorbed into a commercial product, with no credit or consent.

Reason 5: The study that “proved” AI makes developers slower

The most famous argument against AI coding tools came from a study by the AI evaluation lab METR. In a randomized trial, 16 experienced open-source developers worked on 246 tasks in repositories they knew well. The headline result: AI tools made these developers 19% slower, not faster. Even more unsettling, the developers believed they were 20% faster while actually being slower.

That chart went viral. Developers screenshot it, shared it in Slack, and sent it to managers. It became the go-to citation for anyone arguing that AI coding tools are hype: the data was in, and AI didn’t just fail to help—it made things worse while fooling people into thinking it helped.

The twist: when the strongest argument falls apart

Months later, METR quietly revisited its own study. They didn’t fully retract it, but they admitted a major flaw: a selection problem in the methodology.

The developers who benefit most from AI tools simply refused to participate in the “no AI” control condition. They wouldn’t accept being paid $50 an hour to work without their tools. The people who stayed in the study were, statistically, the ones already skeptical of AI.

That means the 19% slowdown wasn’t a clean measurement of what AI does to experienced developers in general. It was a measurement of what happens when you ask AI-skeptical developers to use AI on their own codebases for a few weeks. METR is now redesigning the study from scratch.

But the correction barely spread. The original viral chart still circulates; the update has a fraction of the attention. And that gap—between what people share and what they quietly ignore—points to something deeper than a single flawed paper.

Why resistance persists even as the evidence shifts

Even if you remove the METR study from the conversation, the resistance to AI coding tools isn’t going anywhere. Developers still use AI at record levels, and trust is still at record lows. The gap is widening, not closing.

On paper, the five reasons—bugs, security, skill atrophy, IP concerns, and questionable productivity—are all real. They’re backed by data, and they’re frustrating in day-to-day work. But they’re also socially acceptable reasons. You can say them in a meeting, post them on LinkedIn, or bring them up in a code review without exposing anything personal.

What’s harder to say out loud is the emotional core: for 15 years, “learn to code” was sold as the safest, most future-proof career path. Now, the same companies that hired a generation of developers are building tools that look, from the outside, like replacements.

The labor market shock underneath the debate

To see why the resistance is so intense, you have to look at the job market, not just the tools:

• Hiring of new graduates at the 15 largest US tech companies has fallen by more than half since 2019.
• Employment for software developers aged 22–25 is down nearly 20% from its 2022 peak.
• There were around 80,000 tech layoffs in the first quarter of 2026 alone, with roughly 55,000 explicitly attributed to AI.
• For the first time since the dot-com crash, computer science enrollment in major public university systems is starting to decline.

These are not small adjustments. They’re signs of a structural shift in how many developers the industry thinks it needs—and what kind of developers those are.

In that context, arguments about buggy code or flawed studies become a kind of shield. They let people express discomfort with AI without having to say, “I’m worried my career is being automated away.” The resistance becomes the only psychologically available response to a profession that feels like it’s being rewritten in real time.

The real question: what does the industry owe its developers?

Framed this way, the debate over AI coding tools isn’t really about whether AI is “good” or “bad” for programmers. The truth is messier:

• The bugs are real.
• The security vulnerabilities are real.
• The risk of skill atrophy is real.
• The IP and licensing issues are real.
• The productivity impact is still genuinely unsettled.

But treating those arguments as the cause of resistance misses the point. While we argue over benchmark charts on social media, the labor market is quietly being restructured underneath us.

The honest question isn’t “Will AI replace programmers?” It’s: what does the software industry owe the people it encouraged to invest years of their lives into this career, just as AI was about to change the rules?

No number of debugging anecdotes or benchmark studies can answer that. Until companies, universities, and policymakers grapple with it directly, the resistance will keep finding new studies, new horror stories, and new screenshots to rally around—because the resistance was never just about the tools.

Looking ahead, many expect a new hierarchy to form: a small group of highly skilled engineers who can design systems and reason deeply without AI, and a much larger group who mainly orchestrate and supervise AI-generated code. In that world, coding without AI becomes a luxury skill—and a powerful bargaining chip.

For developers trying to navigate this shift, it’s worth exploring both how to work effectively with AI and how to keep your core engineering skills sharp. Comparing how different models behave in real workflows, like in hands-on tests of ChatGPT and Claude, is one practical way to stay grounded in what these tools can and can’t do today.