AI Coding Dependency Deepens Even as Evidence Mounts It Slows Developers Down
Summary
- • Developers now refuse to work without AI even briefly for research studies
- • METR found AI actually slowed coders down despite their feeling more productive
- • Amazon and Uber both found heavy AI usage failed to boost measurable productivity
- • AI-generated code may create more bugs and long-term maintenance costs than human code
Details
METR could not repeat its 2025 productivity study because developers refused to work without AI
METR's original 2025 study found AI slowed developers down despite their self-reported productivity gains. When METR attempted a follow-up in February 2026, participants declined to work without AI tools even for a limited number of study tasks, forcing METR to shift to a self-reported survey instead.
METR survey found technical employees believe AI makes them twice as valuable, but this is self-reported
Because the controlled experiment was not feasible, METR published a May 2026 survey allowing technical employees to self-report productivity gains. Respondents perceived AI doubled their organizational value — a figure that cannot be independently verified and conflicts with objective findings from the 2025 study.
Amazon shut down its internal AI productivity leaderboard Kirorank after employees gamed it
Amazon had tracked AI token usage via an internal leaderboard called Kirorank as a productivity proxy. Employees responded by running AI agents excessively to inflate their scores, driving up costs without producing corresponding output gains. The Financial Times reported Amazon shut the program down.
Uber exhausted its entire 2026 AI budget within four months with no measurable productivity increase
Uber's COO Andrew Macdonald acknowledged on a podcast that the company's rapid AI spending had not led to a measurable increase in projects or productivity. The overspend occurred in the first four months of 2026, well ahead of schedule, illustrating the gap between AI investment and realized returns.
CodeRabbit analysis found AI-generated code produces 1.7x more problems than human-written code
Code-reviewing tool company CodeRabbit analyzed open source pull requests and found AI produced 1.7 times more issues than human developers. Separately, Entelligence AI CEO Aiswarya Sankar reported that companies spend 44% of their tokens fixing bugs their own AI generated. Both figures come from companies with commercial interest in AI code review, though Singapore Management University independently reached similar conclusions.
Singapore Management University warned in April 2026 that AI code introduces long-term maintenance costs
SMU researchers published independent findings warning that AI-generated code worsens over time as maintenance burdens compound. Their research aligns with programmer James Shore's viral argument that speed gains at code-generation time are offset by permanent increases in maintenance work: 'You're trading a temporary speed boost for permanent indenture.'
Research = study findings, Industry Update = corporate program changes, Financials = spending outcomes
What This Means
The AI coding boom has created a dependency loop: developers feel more productive with AI and refuse to work without it, but objective measurements suggest AI often slows them down and generates more bugs than human code. Companies that bet heavily on AI to accelerate engineering output — Amazon and Uber among them — are finding that raw AI usage does not translate to more shipped work or lower costs. The deeper risk is structural: as AI-generated code accumulates in codebases, the maintenance burden grows, meaning today's speed shortcuts may create compounding engineering debt for years to come.