Developers Refuse to Work Without AI, Studies Question Gains
By 2026, AI coding tools have become so embedded in developers' workflows that many refuse to work without them, even for research purposes. But a growing body of evidence suggests that while AI speeds up code generation, it may not improve overall productivity and could create long-term headaches.
The METR Revelation
In February 2026, the respected AI research lab METR published a surprising finding: most developers would not participate in a productivity study that required them to work without AI assistance, even for a limited set of tasks.
METR had hoped to update groundbreaking research from 2025 that measured how long open source developers took to complete tasks by hand versus with AI. That earlier study found that while developers believed AI made them more productive, it actually slowed them down. The time saved in writing code was eaten up by fixing errors, steering the AI, and waiting for it to complete tasks.
When METR tried to repeat the experiment to account for advances in both AI and coder proficiency, they could not find willing participants. "Because they do not wish to work without AI," the researchers confessed.
Instead, METR published a survey in May allowing technical employees to self-report their AI productivity gains. Unsurprisingly, respondents perceived that AI made them twice as valuable to their organizations. But recent headlines about the cost of so-called tokenmaxxing, along with other research, cast doubt on those self-perceptions.
Tokenmaxxing and Budget Blowouts
Tokenmaxxing, or using the number of tokens a person consumes as a proxy for productivity with AI, has been a major trend in 2026. But it may already be over.
Amazon shut down an internal token-tracking leaderboard called Kirorank after employees were gaming the system by using AI agents excessively and running up costs, the Financial Times reported this week. The employees proved that heavy AI use does not automatically translate to higher productivity.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Uber blew through its 2026 AI budget within the first four months of the year, according to The Information. Chief Operating Officer Andrew Macdonald said on a podcast that such spending had not led to a measurable increase in projects or productivity.
Long-Term Maintenance Costs
AI-generated code may not reduce ongoing maintenance needs and could even increase them, programmer and author James Shore argued in a blog post that went viral on Hacker News. "You write code twice as quick now? Better hope you've halved your maintenance costs," he wrote. "Otherwise, you're screwed. You're trading a temporary speed boost for permanent indenture."
Other evidence supports his view. Aiswarya Sankar, founder and CEO of reliability engineering agent startup Entelligence AI, tweeted that companies spend 44% of their tokens on fixing bugs that their AI generated. Meanwhile, code-reviewing tool company Code Rabbit analyzed open source pull requests and found that AI produced 1.7 times more problems than human-written code.
Those statistics are self-serving, coming from companies that sell AI code-reviewing tools. But independent researchers have found similar issues. Researchers at Singapore Management University published a report in April warning that "AI-generated code can introduce long-term maintenance costs into real software projects."
Is There a Solution?
Some vendors argue that developers can simply use AI coding agents to fix the code as fast as AI generates it. Scott Wu, founder and CEO of Cognition and creator of the AI coding agent Devin, suggests that approach. But even he admits that Devin's skill level currently falls somewhere between a junior and mid-level programmer, depending on the task. It is not a hands-off solution.
The Singapore Management University researchers recommend a more human approach. Programmers should understand what tasks AI does well and what it does poorly as deeply as they know their favorite coding languages. Organizations need strong quality assurance systems designed for AI-generated code, and developers must carefully review AI output as they would a junior developer's work.
The researchers and Wu agree on one point: humans should still handle big-picture work like software architecture and security design.