Tech CEOs are making bold claims about AI’s coding prowess. In March, Anthropic CEO Dario Amodei predicted that AI would write 90% of code within three to six months. Meta’s Mark Zuckerberg echoed similar optimism in April, suggesting that for some projects, “maybe half the development” could be AI-driven within a year. Amazon, Google, and Microsoft executives have also touted large language models (LLMs) as game-changers for software development.
But on the ground, the picture is more complicated.
After months of hesitation, software engineer Colton Voege finally tried AI coding tools—prompted in part by Y Combinator leaders praising AI’s “incredible productivity.” He found the tools useful for quick, disposable scripts: “It’s really good at shortcutting certain things.” Yet he hasn’t seen a sustained boost in his overall efficiency.
His experience isn’t unique. Many engineers report mixed results. Some spend time untangling poorly generated code from overenthusiastic colleagues. Others feel pressured to incorporate AI—even when it adds little value—just to satisfy management expectations.
Boris Cherny, head of Anthropic’s Claude Code unit, told NPR that “most code is written by Claude Code,” though he declined to give exact figures. Crucially, he emphasized: “Every line of code should be reviewed by an engineer.” Even Anthropic’s so-called “agents”—AI systems that can test and rewrite code autonomously—aren’t foolproof. “When they’re at their worst,” Voege said, “they go into death spirals,” looping endlessly without solving the problem.
Cherny likens Claude to “an expert programmer sitting next to you”—a helpful collaborator, but not a replacement. “In the end, it’s a human that’s doing the work.”
Independent AI researcher Simon Willison agrees. While AI might generate a high volume of code, he argues that “you’ll still need the same—or even more—humans” because software engineering isn’t just about typing code; it’s about designing systems that solve real problems. For experienced developers working on the right tasks, AI can deliver 2x to 5x productivity gains. But those gains depend on skill, judgment, and disciplined engineering practices.
In practice, however, AI is often used indiscriminately—leading to what researchers call “workslop”: extra work created by poorly integrated or misunderstood AI output. One Amazon engineer described a senior colleague who tried to rush a complex project using AI, only to produce a “messy blob of code that didn’t work and nobody understood.” Now, the team is rebuilding it “the old way.”
Empirical evidence reflects this ambiguity. A study by nonprofit METR found that experienced open-source developers using LLMs actually took **19% longer** to complete tasks than those who didn’t—contrary to their own expectations. Meanwhile, a national survey in Denmark showed software engineers self-reporting a modest **6.5% time savings** with AI—the highest among 11 professions, but still far from transformative. “It’s not nothing,” said co-author Anders Humlum, “but I would call it modest relative to the hype.”
Engineers agree AI excels in low-stakes scenarios—like whipping up a quick prototype. “It used to take me an hour; now it takes five minutes,” said Willison. “Mistakes don’t matter much there.” Similarly, Thomas Ptacek at fly.io noted AI helps automate repetitive tasks he’s written many times before, allowing him to quickly spot errors.
Yet corporate pressure to adopt AI is mounting—even where it’s unnecessary. “It’s very much a solution in search of a problem,” said the Amazon engineer, who added that dissenting views are often unwelcome: “You have to use it more, and it has to make you more efficient. If you’re not, then you’re doing something wrong.”
Meta has reportedly pushed for “5X productivity” through AI, and at least one AI startup fired engineers for not using coding tools enough. Meanwhile, a Google survey found that while nearly all engineers use AI in some form, only half “somewhat” trust its output—and 30% trust it “a little” or “not at all.”
There are deeper concerns too: the displacement of junior developers (who traditionally learn by doing routine coding), the environmental cost of training massive models, and diminishing returns on capability improvements.
Amazon maintains that its internal AI tools help engineers “move faster, ship more secure code, and spend less time on busywork,” and claims engineers report feeling more productive. But the company won’t disclose usage rates and insists AI use isn’t mandatory.
For his part, Voege has left his job and is considering launching a new venture. He’s struck by how dramatically Y Combinator’s focus has shifted: “It’s just AI, AI, AI—five out of five” in their latest startup applications.
The message is clear: AI is changing how code gets written, but it’s not replacing engineers. At least not yet—and perhaps not ever in the way executives imagine. The real work—design, judgment, collaboration—still belongs to humans.