Tech CEOs are making ambitious claims about AI’s coding capabilities. In March, Anthropic CEO Dario Amodei said, “We’ll be there in three to six months – where AI writes 90% of the code.” Meta CEO Mark Zuckerberg predicted in April that for a project “over the next year, probably…maybe half of the development will be done by AI.” Executives from Amazon, Google, and Microsoft have also highlighted the ability of large language models (LLMs) to generate code.
AI tools have changed the way coding works for some developers and there are now fewer entry-level software engineers. But despite this dramatic talk, AI in software engineering may not mean a new era of automation.
After months on the sidelines, software engineer Colton Voege finally tried AI tools for his job. He remembers watching a podcast in which executives from tech startup incubator Y Combinator, which helped Voege launch a company, touted the “incredible productivity” of AI.
“It kind of confirmed what I was already feeling, which is that it’s really effective at shortening some things,” Voege said. “(AI) is great for writing small tools that you use once and then throw away.” But he did not see any long-term improvement in its effectiveness.
Interviews show that many, but not all, software engineers share Voege’s experience. Some talk about how they unraveled the AI-generated code their colleagues handed them, others talk about the pressure to invent work that claims to use AI to make superiors happy.
Yes, AI can write code… but it still needs people
In an interview with NPR, Boris Cherny, head of Anthropic’s Claude Code unit, said that “most of the code is written by Claude Code” but declined to provide a percentage. “We’re still working on the science to make sure we can give a really accurate number.”
Cherny also made it clear that no matter how much AI is actually used, “every line of code must be reviewed by an engineer.”
He added that this includes code generated by what Anthropic calls “agents.” Agents are a component of AI tools that can test and rewrite code without human intervention. The feature, hailed as much more autonomous than previous tools, was first introduced by Claude Code and was celebrated by many in the industry as a path to automation.
But agents don’t always work. “When these tools are at their best, they automatically solve a lot of problems. … When they are at their worst, they go into death spirals,” Voege said, describing situations in which the tools, far from solving the problem at hand, get stuck in a testing loop. “I’m like, oh no, you’re far away.”
Cherny did not hesitate to describe Claude as a fairly successful person. “I think a good mental model is that of an expert programmer sitting next to you,” he said. Although he says he can serve as a “thought partner” in some cases, “at the end of the day, it’s a human doing the work.”
Simon Willison, an independent AI researcher invited to preview OpenAI’s latest model, said it was plausible that AI could write a high percentage of code. However, Willison said, “I think they will still need the same number, if not more, humans involved in this process.” This is because AI cannot think about solving higher-level problems like engineers can. “Our job is not to type code into a computer. Our job is to provide systems that solve problems.” Willison said.
For Willison, known in the software world for tinkering with new tools and developing open source ones, the increase in productivity linked to the use of AI depends on choosing appropriate uses and relevant expertise. “I’m comfortable saying that for certain tasks, (for) experienced programmers – those who can work without AI – with AI they can get two to five times the productivity gain, sometimes more, depending on the specific task.”
Willison later wrote on his blog that the teams that benefit the most seem to adopt best practices in software development and collaboration.
But that’s not always what happens in real life. People use AI even when it is necessary, sometimes resulting in more work for their colleagues. Researchers coined the term “workslop” to describe this.
“A very senior engineer at the company was very excited about coding AI and thought he could do this very complex project very, very quickly on his own and tried to run some AI tools,” said an Amazon engineer who is not authorized to speak publicly.
“(It) produced some kind of messy code that didn’t work and no one figured it out. And what I’m working on now is just trying to do it the old-fashioned way.”
All of this could explain why studies show mixed results when it comes to the impact of AI. According to a study conducted by METR, an AI testing nonprofit, experienced open source software engineers who used LLMs ended up taking 19% longer to complete their tasks than their peers who did not, contrary to the engineers’ own expectations. In a national survey conducted in Denmark, software engineers reported saving 6.5% of their time using AI. This is the highest of the 11 occupations, averaging 2.5%.
“I would definitely take 3% any day. That’s like annual productivity growth in a typical profession,” said Anders Humlum, co-author of the Danish study. “It’s not nothing, but I would say it’s modest compared to the experiments.”
Engineers agree that AI shines most where precision matters less. “Making a quick illustrated prototype took me an hour (used to) and now it takes me five minutes because I could just tell Claude to break something (together),” Willison said. “The fact that he’s going to make mistakes and all that doesn’t really matter.”
AI has helped reduce some repetitive coding tasks at work, said Thomas Ptacek, a software developer at cloud infrastructure company fly.io. “Because I’ve written (a piece of code) many times before… I can also look at this code very quickly and know if it’s crazy or not.”
Little room for disagreement
The Amazon software engineer said management encouraged staff to use AI in all parts of their work, even when it was not justified. “It’s really a solution in search of problems most of the time,” the engineer said.
The engineer said all of this makes it seem like there is little room for disagreement. “You have to use it more and it has to make you more efficient. If you don’t, then you’re doing something wrong.”
That said, the engineer also said that the tangible effects of rhetoric on work and promotion vary within the organization.
Amazon isn’t the only tech company hypothesizing about the effects of AI on code. In a memo first reported by 404 Media and Wired, a Meta executive called on his department to achieve “5X productivity” in software as well as a wide range of other functions using AI. “The more we strive, the more we will unlock,” he wrote in the note. Two engineers were fired from an AI startup in San Francisco for not using an AI coding tool sufficiently, The Information reported.
The Amazon software engineer and Voege said they are not the only engineers who are skeptical of AI in their workplace. A recent Google survey of software engineers across countries and industries showed that while AI use is near-universal and improves some aspects of software development, about half of respondents trust AI’s response “somewhat,” with 30 percent choosing “a little” or “not at all.”
Many also worry that employers are replacing younger coders with AI, cutting off the supply of long-term talent needed to oversee the machines in the years to come. Others worry about the massive amount of electricity and human-generated content that the larger models consume, even as dramatic improvements in capabilities begin to seem elusive.
In a statement to NPR, Amazon spokesperson Tom Parnell said its internal AI tools “help engineers act faster, ship more secure code, and spend less time on busy work” and that, through surveys, “engineers tell us these tools help them save time and feel more productive.” He also said the company doesn’t mandate the use of AI and can’t say what percentage of its software engineers regularly use large language model tools, but it believes “these tools can help developers work more efficiently.”
Voege has since quit her job and is once again considering starting her own business. He was stunned to see how his former backer, Y Combinator, went from supporting different types of technology to focusing solely on AI.
He reviewed the organization’s latest startup application. “It’s just AI, AI, AI, five out of five.”
