An investigation into 4 billion GitHub commits reveals the speed and scale of an industry transformation

Software developers don’t like admitting they need help. Yet buried in GitHub’s public archive, a different story emerges. Between January 2020 and August 2025, mentions of AI assistance in commit messages jumped from 9,656 to 1.67 million—a staggering 17,212% increase that nobody saw coming.

The real number is almost certainly higher. Much higher.

This investigation is based on the analysis of over 4 billion GitHub commits. It reveals not just the explosive growth of AI coding tools. It also uncovers a more troubling picture. The industry is racing ahead without understanding what it’s leaving behind.

The numbers don’t lie, But they don’t tell the whole truth either

Here’s what I found in the raw data:

In 2020, developers mentioned AI tools in their commits exactly 9,656 times. It was statistical noise, 0.0025% of all commits. Most developers hadn’t even heard of AI coding assistants.

By August 2025, that number had exploded to 1,671,648 mentions across just eight months. The adoption curve isn’t just steep; it’s nearly vertical.

But here’s the catch: these numbers only count developers who explicitly credit AI in their commit messages. Think about that for a moment. How often do you mention your IDE in a commit? Your coffee maker?

My guess is the real usage is 10 to 50 times higher. If I’m right, we’re looking at 5-10% of all code being touched by AI. Today. Right now.

June 2025: The month everything hanged

Something extraordinary happened in June 2025. After steady but manageable growth through the spring, AI-assisted commits suddenly doubled. In a single month.

The numbers are jarring:

• May 2025: 105,583 AI commits
• June 2025: 219,797 AI commits
• July 2025: 399,462 AI commits

That’s a 108% jump in June, followed by another 82% surge in July. By August, developers were creating over 13,000 AI-assisted commits every single day.

What triggered this explosion? The data doesn’t say. The timing could coincides with several factors. Did Claude’s dramatic improvements have made an impact? Did GitHub Copilot’s aggressive pricing changes also play a role? Maybe, a viral blog post or YouTube video by an engineer claims AI had made him “10x more productive.” and got shared many thousands of times?

The genie was out of the bottle, no matter what caused it.

The great tool wars: How Claude ate everyone’s lunch

The most shocking revelation in my data isn’t the growth—it’s who’s winning.

In 2024, the landscape looked predictable:

• ChatGPT/GPT dominated with 511,028 mentions (74% market share)
• GitHub Copilot held second with 116,341 mentions (17%)
• Claude trailed with just 56,645 mentions (8%)

Fast forward to 2025:

• Claude: 954,832 mentions (69% market share)
• GitHub Copilot: 284,906 mentions (21%)
• ChatGPT/GPT: 114,361 mentions (8%)

Claude didn’t just win; it obliterated the competition.

In eight months, it went from distant third to absolute dominance, capturing nearly 70% of explicit mentions.

This isn’t normal market behavior. Tools don’t typically gain 60 percentage points of market share in eight months. Something fundamental shifted in how developers work, and Claude was perfectly positioned to capitalize on it.

Or-and this is where it gets interesting Claude users simply be more to credit their AI assistant. The data can’t tell me which.

What developers actually do with their AI Overlords?

When developers do credit AI, what are they using it for? I analyzed 690,807 commits from 2024 to find out:

The Breakdown:

• Debugging and problem-solving: 244,569 commits (35.4%)
• Building new features: 140,730 commits (20.4%)
• Code improvements and refactoring: 95,959 commits (13.9%)
• Unspecified: 209,549 commits (30.3%)

Stop and think about that first number. More than a third of AI usage is for fixing problems. Developers are increasingly outsourcing their debugging to machines. The implications are staggering.

Ten years ago, not even ten, about four years ago, debugging was how junior developers learned. They’d spend hours tracing through code, understanding logic flows, and discovering edge cases. Now? They ask ChatGPT or Claude.

“It’s like GPS,” one veteran engineer told me few weeks ago. “Incredibly convenient, but now nobody knows how to read a map.”

The skills crisis nobody wants to talk about

The elephant in the room isn’t whether AI makes developers more productive—it clearly does. The question is what happens to human skill when machines handle the hard parts.

Consider this: In 2025, we’re seeing over 50,000 AI-assisted commits every month for debugging alone. That’s 50,000 instances where a developer chose to ask AI rather than solve the problem themselves.

The optimists argue this frees developers to focus on higher-level thinking. The pessimists worry we’re creating a generation of developers who can’t operator without their AI crutch.

Both can be right, but personal experience tells me the latter, not that I’m a pessimist.

Several tech companies have started reporting a troubling trend. New hires who learned to code alongside AI assistants struggle with basic debugging. This occurs when the tools aren’t available. They can build features quickly, but they can’t explain how their code works.

“We had a candidate. He built an impressive React demo app in the interview,” a friend of mine, a hiring manager at a Fortune 500 company shared with me. “When we asked him to explain a particularly clever algorithm, he admitted ChatGPT wrote it. He had no idea how it worked.”

The measurement problem that changes everything

Here’s the dirty secret about our data: it’s almost certainly wrong.

My analysis only captures developers who explicitly mention AI tools in commit messages. That’s like measuring coffee consumption by counting people who mention Starbucks in their emails. The real number is orders of size higher.

Consider the incentives:

• Many companies prohibit developers from mentioning AI use
• Some developers view AI assistance as “cheating”
• Others simply don’t think to mention it
• Commits made through AI-powered tools may not reference AI at all

If my detected 0.237% adoption rate shows just 2-5% of actual usage, the real adoption is already at 5-10% of all commits. At that scale, AI isn’t just helping with software development-it’s fundamentally changing what software development means.

The corporate adoption curve: from banned to mandatory

In 2023, most Fortune 500 companies banned AI coding tools. By 2025, many mandate them.

The reversal happened almost overnight. Once companies realized their competitors were shipping features twice as fast, or is it? data doesn’t tell me that either, it tells me only about the number of commits. Regardless everyone’s belief at this point is that, so the choice became simple: adopt or die.

But adoption without understanding is dangerous. Companies are deploying AI-assisted code into production without knowing:

• Who’s liable when AI-generated code fails?
• How to review code that no human fully understands?
• What happens to institutional knowledge when it’s outsourced to AI?
• How to keep AI-generated codebases long-term?

I believe “We’re essentially running a massive, uncontrolled experiment on the world’s software infrastructure.” We won’t know the results for years.

The coming reckoning

This is a the part 3 of MI: Dead reckoning

The data tells us AI adoption in software development is accelerating beyond anyone’s predictions. What it doesn’t tell us is whether that’s a good thing.

Yes, developers are shipping code faster. But faster isn’t always better. The technical debt from poorly understood, AI-generated code will haunt the industry for decades. We’re already seeing early signs:

• Codebases becoming increasingly difficult to manage
• Security vulnerabilities in AI-generated code going unnoticed
• Dependencies and architectural decisions that no one can explain

The 17,212% growth in AI-assisted commits isn’t just a statistic—it’s a warning. We’re transforming an entire profession at breakneck speed without stopping to ask whether we should.

What will happens next?

If current trends continue, AI-assisted commits will reach 2-3% of all GitHub activity by year’s end. The real number, including unacknowledged usage, already be there.

This isn’t the story of a tool being adopted. It’s the story of an industry being transformed, ready or not.

The software developers of 2030 will work in ways that would be unrecognizable to their 2020 counterparts. Whether they’ll be more capable or merely more dependent remains an open question.

What’s certain is that the revolution isn’t coming. It’s here. It’s in the 1.67 million commits that admitted it, and the millions more that didn’t.

The question now isn’t whether to adopt AI coding tools, that ship has sailed. The question is whether we can keep human skill and understanding in an increasingly AI-dependent industry.

Based on the data, I’m not trying very hard to find out.

---

My Approach and Limitations

This analysis examined 4.1 billion GitHub commits. The timeframe was from January 2020 through August 2025. It focused on searching for explicit mentions of AI coding tools. These tools included Copilot, Claude, ChatGPT, and related terms. The data captures only commits where developers explicitly acknowledged AI assistance—representing 2-10% of actual AI usage.

Raw data sourced from GitHub Archive via BigQuery. All statistics and trends cited are based on direct analysis of commit messages. These statistics do not significantly represent true adoption rates.