Coders refuse to work without AI — but the hidden costs are piling up

In 2026, developers have made their stance clear: they will not code without artificial intelligence. Researchers at the respected AI safety lab METR discovered this firsthand when they attempted to replicate a 2025 study measuring how AI affects coding productivity. The problem? No developers were willing to participate in the control group — the one that required working without AI assistance.

METR’s failed experiment and the self-perception gap

METR’s original 2025 study had already delivered a counterintuitive finding: while developers felt more productive with AI, they actually took longer overall because of time spent correcting AI-generated errors, steering the tool, and waiting for responses. When METR tried to repeat the experiment with newer AI models, they couldn’t recruit participants. As the researchers confessed, developers “do not wish to work without AI” even temporarily for a controlled study.

Instead, METR published a survey in May 2026 allowing technical employees to self-report their productivity gains. Unsurprisingly, respondents perceived themselves as twice as valuable to their organizations. But recent developments suggest these self-assessments may not reflect reality.

Tokenmaxxing: the trend that already backfired

The practice of using AI token consumption as a proxy for productivity — dubbed “tokenmaxxing” — became a defining trend of early 2026. But it may already be collapsing under its own weight. Amazon shut down its internal token-tracking leaderboard, Kirorank, after employees gamed the system by running AI agents excessively, driving up costs without meaningful output, as reported by the Financial Times. Uber burned through its entire 2026 AI budget in just four months, and COO Andrew Macdonald acknowledged on a podcast that the spending had not led to measurable increases in projects or productivity.

The maintenance cost trap

Beyond cost overruns, evidence is mounting that AI-generated code introduces long-term liabilities. Programmer and author James Shore articulated the problem in a widely circulated blog post: “You write code twice as quick now? Better hope you’ve halved your maintenance costs. Otherwise, you’re screwed. You’re trading a temporary speed boost for permanent indenture.”

Independent research from Singapore Management University, published in April 2026, warned that “AI-generated code can introduce long-term maintenance costs into real software projects.” Code review platform Code Rabbit analyzed open source pull requests and found AI-produced code introduced 1.7 times more problems than human-written code. While these numbers come from companies with a commercial interest in code review tools, the pattern is consistent across multiple sources.

Why the productivity promise may be hollow

The disconnect between perceived and actual productivity is not new, but the scale of AI adoption amplifies its consequences. Developers love their AI assistants — they report feeling faster, more capable, and less frustrated. But objective metrics tell a different story: more bugs, higher token costs, and maintenance burdens that compound over time. A viral post from Aiswarya Sankar, CEO of reliability engineering startup Entelligence AI, claimed companies are spending 44% of their AI tokens on fixing bugs that AI itself generated.

What the experts recommend

Solutions vary depending on who you ask. Cognition CEO Scott Wu, whose company makes the AI coding agent Devin, argues that developers should simply use more AI to fix the problems AI creates. But even Wu admits Devin currently performs between a junior and mid-level programmer, depending on the task — not a hands-off solution.

The Singapore Management University researchers advocate a more human-centered approach: developers must understand which tasks AI handles well and which it doesn’t, as deeply as they know their programming languages. They recommend strong quality assurance systems designed specifically for AI-generated code, and careful human review of AI output — treating it like work from a junior developer. The researchers, and Wu, agree on one point: humans should retain responsibility for big-picture decisions like software architecture and security design.

Conclusion

The coding profession has reached an inflection point. Developers have integrated AI so deeply into their workflows that many cannot or will not work without it. But the evidence suggests this dependency comes with real costs — inflated budgets, hidden maintenance burdens, and a gap between perceived and actual productivity. The path forward likely involves not abandoning AI, but developing more disciplined practices around when and how to use it, and accepting that no tool eliminates the need for human judgment in software engineering.

FAQs

Q1: Why did METR fail to replicate its 2025 AI productivity study?
Developers refused to participate in the control group that required working without AI assistance, even temporarily for a research study. METR instead published a survey-based report.

Q2: What is tokenmaxxing and why is it controversial?
Tokenmaxxing is the practice of using AI token consumption as a proxy for employee productivity. Companies like Amazon and Uber found that employees gamed the system, driving up costs without corresponding increases in output, leading to the abandonment of token-tracking programs.

Q3: Does AI-generated code increase maintenance costs?
Multiple sources — including academic research from Singapore Management University, independent analysis by Code Rabbit, and commentary from industry experts — indicate that AI-generated code introduces more bugs and long-term maintenance burdens compared to human-written code.