The AI Coding Paradox: Developers Won't Work Without It, But It May Be Slowing Them Down

A quiet rebellion is underway in software engineering. Developers are refusing assignments that ban AI coding assistants, a stance that disrupted a controlled productivity study this spring. Researchers at METR couldn't recruit participants for a trial comparing AI-assisted and unaided coding—engineers simply wouldn't participate without their tools. The refusal, reported by TechCrunch on May 29, 2026, highlights a deepening dependency on platforms like Cursor, GitHub Copilot, and Claude-powered editors.

Yet the data tells a different story. METR's 2025 randomized trial found experienced programmers took 19% longer to complete real open-source tasks when using frontier AI models. Participants spent extra time steering prompts, reviewing outputs, fixing hallucinations, and waiting for suggestions. Despite objective evidence, they still believed AI made them 20% faster—a perception gap that has only widened. A May 2026 survey found developers now report feeling twice as valuable with AI, even as hard metrics contradict that sentiment.

Enterprise rollouts reveal similar patterns. Amazon's internal leaderboard, KiroRank, designed to encourage AI adoption, backfired. Engineers assigned trivial work to AI agents to climb the rankings—a phenomenon dubbed "tokenmaxxing." Costs soared with no productivity gain, and Amazon shut it down. Uber exhausted its 2026 AI budget in four months, with flat throughput. Salesforce expects to spend $300 million on Anthropic tokens this year alone, according to The Next Web.

The downstream effects compound. Pull requests surged 98% with AI assistance, but review times jumped 91%. Faros AI's 2025 engineering report documents this dynamic: more code arrives, but it demands extra scrutiny for subtle errors and duplicated logic. Maintenance costs also climb. James Shore, a veteran programmer, warned in The Next Web: "You write code twice as quick now? Better hope you've halved your maintenance costs… Otherwise, you're screwed."

Specialized tools confirm the trend. Entelligence AI found 44% of tokens in some environments went toward bug fixes. CodeRabbit determined AI-produced code created 1.7 times more problems than human-written contributions. Singapore Management University researchers reached similar conclusions. The tools accelerate initial drafting but shift effort downstream.

Despite this, developers won't relinquish them. New graduates expect AI autocomplete as standard. Mid-career engineers describe flow states that vanish without intelligent suggestions. Even researchers can't recruit control groups anymore. The industry faces a paradox: individual velocity feels higher, but system velocity does not.

Some organizations are tightening guardrails, routing prompts through intermediary layers to control cost and quality. Others invest in better evaluation frameworks tracking defects, review cycles, and long-term ownership burden. A few are experimenting with AI usage caps on critical paths. The moves acknowledge a central tension: feeling productive and delivering faster are not the same.

Stack Overflow's 2025 developer survey captured the split. 84% of professionals use or plan to use AI tools. Only 16% say the assistance makes them substantially more productive. The largest cohort reports modest or negligible gains. Those numbers align with METR's controlled data.

Resolving the paradox demands sharper measurement. Teams need telemetry capturing full-cycle outcomes—defect density, onboarding speed, and total cost of ownership over quarters. They must distinguish tasks where AI shines from those requiring human judgment. Most of all, they must resist optimizing for proxies like tokens consumed or pull requests merged. Those proxies failed Amazon. They will fail others.

Until better data arrives, the refusal to work without AI will likely harden. The tools have become part of the craft. Whether they ultimately raise the profession's output or simply redistribute effort remains the open, expensive question facing technology leaders in 2026.