AI coding assistants are making experienced developers more productive. Early-career developers? Not so much.
That's the finding from a rigorous study published in Science this week, and it has profound implications for how AI might reshape not just software development, but career ladders across knowledge work.
What the research found
The researchers tracked thousands of professional software developers before and after they adopted AI coding tools. They measured productivity, code quality, and - crucially - whether developers expanded into new domains of software development.
Senior developers showed clear gains. They completed tasks faster, maintained code quality, and more readily ventured into areas of development outside their primary expertise. The AI effectively lowered the barrier to trying new things.
Junior developers showed no significant benefits.
Not slower gains. Not smaller improvements. No measurable benefit at all.
Why the gap?
This makes intuitive sense if you think about what AI coding assistants actually do. They're autocomplete on steroids - pattern-matching systems trained on vast amounts of existing code. They're excellent at generating syntactically correct code for common patterns.
But here's the thing: experienced developers already know the patterns. What takes them time isn't remembering syntax or looking up API documentation - it's architectural decisions, understanding trade-offs, debugging edge cases, and maintaining systems over time.
AI can speed up the mechanical parts while senior developers handle the judgment calls.
Junior developers, on the other hand, are still learning what good patterns look like. They need to understand why code is structured a certain way, not just that it works. AI-generated code can actually obscure the learning process - it gives you an answer without showing the reasoning.
It's like trying to learn mathematics with a calculator that shows you the answer but not the steps.
The widening gap
What worries me most isn't the productivity difference - it's what this means for career development.
Software development has traditionally had a reasonably clear progression: you start as a junior doing more routine work, gradually taking on more complex problems, learning from code reviews and mentorship, eventually becoming senior yourself.
But if AI can handle much of the routine work that juniors used to cut their teeth on, where do they develop expertise?
We might end up with a bimodal distribution: highly productive senior developers who can leverage AI effectively, and a missing generation of mid-career developers because the learning path disappeared.
Not just a coding problem
This pattern likely extends beyond software. Consider:
• Junior lawyers who learn by drafting routine documents • Analysts who develop judgment by working through basic datasets • Researchers who build expertise by doing literature reviews • Writers who develop their voice through editing and revision
In each case, AI can handle the routine work - potentially eliminating the very experiences that build expertise for the next generation.
The researchers note this could "widen skill gaps and reshape future career ladders." That's academic speak for: we might be inadvertently dismantling the pathways through which expertise develops.
What comes next
This doesn't mean AI in the workplace is inherently bad. But it does mean we need to think carefully about how we integrate it, particularly for people still building their skills.
Maybe that means deliberately limiting AI assistance during learning phases. Maybe it means redesigning training to focus on the judgment and architectural thinking that AI can't do. Maybe it means finding new ways to create the deliberate practice that builds expertise.
What it definitely means is that we can't just assume AI tools benefit everyone equally. The same tool that makes experts more productive might be actively harmful to novices - not because of malice, but because expertise develops through struggle, and AI can make things too easy.
The universe doesn't care about our career ladders. But if we want to maintain the ability to develop expertise in an AI-augmented world, we need to understand these dynamics before we optimize ourselves into a corner.
