AI Is Writing Nearly a Third of All Software Code in the US as the Technology Takes Over Silicon Valley

Until just very recently, writing software was a purely human craft, a slow and grinding process of translating logic into a myriad forms of syntax. Any developer worth their salt needs to know Java, Python, and JavaScript as a bare minimum. But there are literally hundreds of libraries and frameworks that are available, some of which they need to master in order to perform well. But now, the most common programming language is English. Or Spanish. Or Japanese, why not?

If you’re not familiar with what I’m referring to, perhaps you’ve missed the vibe coding train. Vibe coding is an AI-assisted software development method where developers describe their desired application in natural language prompts, letting a large language model (LLM) generate the code. You can literally ask Claude Code, Google’s Gemini, ChatGPT’s Codex, or Cursor — have your pick, there are so many other services I haven’t mentioned — “make me an iOS app for my recipe book” or “build me a website with such-and-such sections and in such-and-such style.” Twenty minutes later the job is done and sometimes the AI “one-shots” it, meaning it can get it right on the first go.

Vibe coding is driving a new revolution that involves the democratization of software developers, allowing virtually anyone to build simple software tools now. In the hands of seasoned software developers, these AI writing tools are even more powerful, acting as massive force multipliers.

According to a new study published today in Science, generative AI hasn’t just entered the world of programming — it’s looking like it’s taking over the foundation.

In the United States, the share of new code written with AI assistance has skyrocketed from a mere 5% in 2022 to a staggering 29% by early 2025. This is a massive structural shift in how our digital world is built.

“We analyzed more than 30 million Python contributions from roughly 160,000 developers on GitHub,” says Simone Daniotti of the Complexity Science Hub (CSH), noting that the sheer scale of the data allowed them to track this global transformation in real-time.

A Digital Divide Between Nations

While the U.S. leads the charge, the rest of the world is following at different speeds. France and Germany are close behind, with AI-supported code reaching 24% and 23% respectively^{. India is the “fast-mover,” currently at 20% but catching up rapidly.}

However, the geopolitical landscape of AI is far from uniform. China (12%) and Russia (15%) significantly lag behind^{. This gap isn’t necessarily due to a lack of interest, but rather a lack of access. “Users in China and Russia have faced barriers to accessing these models, blocked by their own governments or by the providers themselves,” explains Johannes Wachs, a researcher at CSH and associate professor at Corvinus University of Budapest.}

Wachs notes that while VPN workarounds exist, the real game-changer might be domestic breakthroughs like China’s DeepSeek, which could close this gap in the coming year. DeepSeek’s initial open source R1 model was a game changer in the AI space, performing on par with top-tier, proprietary models (like OpenAI’s o1) while using significantly less computational power. By openly sharing its reasoning paths and post-training methods, R1 turned advanced reasoning, previously locked behind closed APIs, into an engineering asset that could be downloaded, distilled, and fine-tuned. We can expect non-Western countries currently lagging behind AI coding adoption to balloon in the future for this reason.

The Experience Paradox

Perhaps the most startling finding of the research is who actually benefits from these tools. On the surface, you might expect AI to be the ultimate equalizer — a “leveling up” tool for beginners. The data shows that early-career developers are indeed the heaviest users, with AI assisting in 37% of their code.

You know there’s a “but” here somewhere. The study found they aren’t seeing any productivity gains.

The real “force multipliers” are the seasoned veterans. Experienced developers use AI less frequently (27% of their code) but drive the entirety of the study’s documented 3.6% increase in overall productivity. Bear in mind these figures lead up to 2024, so today AI code is most likely even more pervasive.

When asked about this disparity, study author Frank Neffke suggests that effective AI use requires a “managerial” mindset rather than a “crutch” mindset. “Using AI effectively requires that a programmer interprets, evaluates and integrates its code suggestions,” Neffke told ZME Science.

“In a sense, the programmer manages a process, weighing alternative suggestions and deciding what to implement and how. More broadly, AI can only answer the questions you can ask”.

AI Code Versus Human-written Code

To find these AI-generated snippets within millions of lines of code, the researchers had to build their own “digital detective.” They trained a neural classifier — specifically a GraphCodeBERT model — to distinguish between human and machine.

The team created a “ground truth” dataset by taking human code from 2018 (before the era of modern LLMs) and then using a two-step “synthetic cloning” process. They had one AI describe a human-written software function in plain English, and a second AI attempt to recreate that function from the description. This created pairs of code designed to do the same thing, but with different “DNA” — one human, one machine.

Interestingly, the researchers found no strong link between “AI slop” — unusually verbose or “unnatural” code — and their model’s ability to spot it.

“We did not directly try to look into the black box of our classifier, which uses code embeddings (similar to word embeddings) to describe software code. But when we analyze the false positives (instances where the model misinterprets human-written code for AI-generated), we don’t find any strong relations with things like how verbose the code is or how “natural” the flow of the code is. But this was not a main focus of our study,” Neffke told me in an email.

The $38 Billion Question

What does a 3.6% productivity boost look like when translated into the real economy? For the U.S. alone, the researchers estimate it adds between $23 and $38 billion in value annually^.

“The value estimates should be regarded as ballpark numbers from back-of-the-envelope calculations. We combine an estimate of the productivity boost we observe for individual programmers from using AI with our estimate of how intensively AI is used by US programmers at the end of 2024. Next, we estimate labor costs in the US economy that can be attributed to coding tasks. Putting all of this together leads to a rough guess of the value of the code that using AI generated,” Neffke said.

But will this mean fewer jobs for humans? Neffke points to the Industrial Revolution for a clue. When textile production was mechanized, many expected employment to crater. Instead, it “ballooned”. Because clothes became cheaper, people bought more of them — curtains, drapes, and multiple outfits became the new norm.

“Whether this value shows up in more code or in fewer hours-worked… depends on what economists call the ‘elasticity of demand’ for code,” Neffke explains.

Elasticity of Demand

“AI coding assistants reduce the time it takes to create code. So, on the one hand, you could reduce the number of programmers to create the same amount of code. Code then becomes cheaper (same code for lower wage costs). This, on the other hand, may lead to more demand for code (at the new price point, it becomes cost-effective to renew an existing code base, digitize new work processes, create new digital products/services, etc.) This would increase the demand for code to the extent that you actually end up needing more programmers.”

If making code cheaper leads to a massive wave of new digital products and services, we might end up needing more programmers than ever, not fewer. Yet long-term trend is unclear, even though the short-term effects seem to point toward some level of job loss among developers, particularly those starting out in entry-level positions.

The Washington Post reported in March 2025 that more than 25% of all computer-programming jobs in the U.S. disappeared in just two years. The journalists used data from the Current Population Survey from the Bureau of Labor Statistics. There were more than 300,000 computer-programming jobs in 1980. The number peaked above 700,000 during the dot-com boom of the early 2000s but employment opportunities have withered to about half that today. U.S. employment grew nearly 75% in that 45-year period, according to the Post.

There are now fewer computer programmers in the U.S. than there were when Pac-Man was first invented — years before the internet existed. The decline is heavily attributed to generative AI’s ability to handle “rote” or “routine” coding tasks. Tasks that used to require a junior or entry-level programmer to write boilerplate code are now being handled by AI agents.

Reshaping the Career Ladder

As AI becomes the backbone of our digital infrastructure, the “beginner’s gap” remains the most pressing concern for the future of the industry. If the next generation of coders is using AI as a crutch without gaining the experience needed to become “force multipliers,” the career ladder of the future might be missing its bottom rungs. Then there’s also the market forces involved. If companies would rather get rid of entry-level developers and programmers in order to replace them with generative AI code, where will the next generation of senior developers come from? Perhaps that’s a problem future CEOs need to worry about.

“For businesses, policymakers, and educational institutes, the key question is not whether AI will be used, but how to make its benefits accessible without reinforcing inequalities,” says Wachs.

In a world where even your car is essentially a “software product,” understanding how we learn to work with our new silicon colleagues is no longer optional — it’s essential. But look at the bright side, anyone can now start coding in plain English. Give it a try, and you might be surprised by the results.