AI Replacing Software Developers? What the Latest Research Actually Shows
“AI will write 90% of code by tomorrow.”
“Developers will be obsolete in five years.”
Tech Twitter loves a good doom spiral. Engagement farms thrive on fear. Scary sells.
But here is what the data actually shows: AI is not a replacement. AI is amplification, and that amplification cuts both ways.
AI is chewing through boilerplate, dumb bugs, documentation, and tests. Developers who use it well are shipping faster. Some teams are getting more output from fewer people. That is happening.
But a wholesale wipeout of the profession? The numbers do not support that. The Bureau of Labor Statistics still projects software developer jobs growing 15–18% through 2034, well above the national average. If AI were gutting the field, government forecasts would not look this calm.
Goldman Sachs research from 2023 found that roughly 25% of work hours in advanced economies could be automated. But only around 6–7% of jobs would disappear entirely. The rest would be augmented or shifted. Developer work falls in the “exposed but survives” bucket. Judgment, context, stakeholder wrangling, and defining what the actual problem is? AI still struggles with all of that.
The Entry-Level Picture Is Harder
Juniors who spend most of their time translating clear specs into code are feeling this disruption most. AI plus a senior developer handles much of that translation work now. Hiring managers are reporting that junior slots are shrinking. Early-career tech employment dipped roughly 20% from 2022 peaks in some markets.
The flip side: PwC found that AI-exposed roles are upskilling 66% faster. Developers fluent in AI tools are grabbing 50%+ wage bumps and getting promoted twice as quickly. The profession is not dying; it is splitting. Seniors who use AI well are winning big. Everyone else is either catching up or falling behind.
The Real Productivity Math
Self-reported data is optimistic, almost by definition. The Stack Overflow 2025 Developer Survey found 84% of developers are using or planning to use AI tools, with about half using them daily. Many claim 30–60% time savings. It feels good to say.
Based on our own internal survey, 80% of code is generated by AI, and it has been that way for over six months.
The controlled data tells a different story.
METR ran a randomized controlled trial in 2025 on experienced open-source developers, people working in large codebases they knew well. Using frontier tools (Cursor with Claude 3.5 and 3.7), these developers were 19% slower than the control group. Not faster. Slower.
Before the study, they predicted they would be 24% faster. After, they estimated they had been 20% faster. The actual measurement was a 19% drag. The culprits: prompt wrestling, reviewing plausible-but-wrong output, debugging hallucinations, and integrating generated code that did not actually fit the codebase. The overhead ate any per-task gains.
Benchmark performance looks better, but benchmarks use toy problems. LLMs solve 84–89% of those correctly. In real class-level code generation on production repositories, accuracy drops to 25–34%. Adding RAG or documentation retrieval moves the needle only 4–7%.
My honest take: AI is automating around 20–40% of routine work today. Boilerplate, simple fixes, documentation. That automation frees up time for architecture, hard debugging, and cross-system problem-solving, but only if you have mastered the workflow. If you have not, AI is costing you time, not saving it. A big test is a year ago if we took AI away from our developers, they would have been really annoyed. Today they would revolt.
Who Is Winning and Who Is Not
The developers struggling right now fall into a few categories: those doing pure spec-to-code translation, juniors without AI skills, and teams that have not updated how they work.
The developers doing well are the seniors who treat AI like a turbocharged intern. Real accounts are reporting 2–2.5x output gains, not because AI is doing the thinking, but because AI handles the grunt work while they focus on architecture. New roles are also emerging: AI orchestrators, agent wranglers, output evaluators, and people building infrastructure to keep teams from drowning in generated code that needs validation.
Team sizes are changing in some organizations, from two-pizza to one-pizza. The senior-to-junior ratio is climbing. And the bar for what qualifies as a junior developer is rising fast.
What to Do Right Now
If the research points to anything, it is that the window for early adoption is still open, but it is closing. The developers pulling ahead right now are not smarter; they are just making more deliberate choices about where to invest their time. Here is where I would focus.
- Go deep on AI, not surface-level. Tab completion is table stakes. Prompt engineering combined with context structuring and ruthless output review is where the actual productivity gains live. Cursor, Claude, and Codex agents are worth real investment.
- Double down on the human parts of the job. System design. Cross-layer debugging. Business translation. Stakeholder communication. AI consistently underperforms in these areas. Own them.
- Move fast. Early adopters are being promoted twice as often and earning 18–35% more. The window will not stay open as everyone catches on. The next 12–18 months matter more than most people realize.
- Measure your own results. METR’s developers felt faster while actually being slower. Track your velocity with AI and without it. Perception is unreliable. Your own data is not.
The Bigger Picture
AI is not erasing software engineering. AI is the next step in a long line of tools that change what engineers do. Assembly coders did not vanish when higher-level languages arrived; they got more done with less effort. IDEs raised the bar for what one developer could produce. Cloud infrastructure lets small teams build things that used to require enterprises.
The same progression is playing out with AI. Routine work gets automated. The hard parts stay human. Entry-level pressure is real. Workflows are changing. Skills are, too.
The arc of this profession has always been the same: new tools raise the ceiling, early movers benefit most, and the profession expands into what it can now do.
The question is not whether jobs are disappearing. The question is whether you are adapting. AI won’t replace programmers; programmers using AI will replace programmers not using AI.
What is AI actually doing to your workday? Faster? Slower? Just more complicated? Send me a message on X (@chadmichel) to continue the conversation.
References
- Bureau of Labor Statistics, Occupational Outlook Handbook: Software Developers (2023–2033 projections)
- Goldman Sachs, “The Potentially Large Effects of Artificial Intelligence on Economic Growth” (2023)
- Stanford HAI, “AI and the Labor Market” (2025)
- PwC, “Global Workforce Hopes and Fears Survey” (2025)
- Stack Overflow Developer Survey (2025)
- METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”
- “LLMs for Class-Level Code Generation in Real-World Repositories”
