Developers Refuse to Code Without AI, but the Data Says It's Making Them Slower and Costing Billions

METR's subsequent attempt in August 2025, expanded to 57 developers and over 800 tasks, only deepened the crisis. The results were statistically inconclusive—showing an estimated slowdown of -4% with a confidence interval so wide it crossed zero (-15% to +9%). Critically, between 30–50% of participants admitted they were self-censoring, choosing not to submit tasks they couldn't complete with AI assistance. The sample was irreversibly biased toward the small pool of developers still willing to work without AI . The lab eventually scrapped the experiment entirely, calling its data "unreliable" .

In less than a year, the research narrative shifted from "AI makes experienced developers slower but they can't tell" to a more profound finding: "developers won't even attempt work without AI." The dependency had become too deep to measure .

The Tokenmaxxing Gold Rush and Its Spectacular Collapse

As the METR study was falling apart, a parallel mania was sweeping through Silicon Valley. Tokenmaxxing—the practice of maximizing raw AI token consumption as a proxy for developer productivity—became the defining trend of early 2026 before imploding under its own weight.

This culture was actively gamified. Meta employees reportedly competed on an internal dashboard called "Claudeonomics," striving for titles like "Token Legend" and "Session Immortal" based on how many tokens they burned . Nvidia CEO Jensen Huang made headlines by stating he would be "deeply alarmed" if one of his $500,000 engineers wasn't heavily consuming tokens . Across the industry, token budgets became a badge of honor, signaling an employee's embrace of AI and, it was assumed, their innovative output.

Amazon Shuts Down KiroRank

The backlash hit a fever pitch in late May 2026. Amazon deprecated its internal "KiroRank" leaderboard on the Kiro developer platform after discovering employees were creating "pointless AI agents" solely to inflate their usage scores. Senior Vice President Dave Treadwell delivered a direct message to staff: "Please don't use AI just for the sake of using AI" . An Amazon spokesperson confirmed the leaderboard was a "beta dashboard" that was "not a formal or approved tool," but the damage from the compute costs it generated was already done . The company is now shifting toward a metric it calls "normalised deployments" to measure meaningful work instead .

Uber Burns Its Annual AI Budget in Four Months

Even more dramatically, Uber's AI push became a cautionary tale. The company gave its roughly 5,000 engineers broad access to Anthropic's Claude Code in December 2025. Adoption surged from 32% to 84% in months, and by April 2026, the company had exhausted its entire annual AI budget. According to CTO Praveen Neppalli Naga, 95% of Uber's engineers now use AI tools monthly, with 70% of committed code being AI-generated—the highest publicly reported percentage at any major tech company .

Yet this stunning adoption had no clear payoff. COO Andrew Macdonald publicly admitted in a late May interview that the company cannot draw a line between its colossal AI spending and meaningful business outcomes. "That link is not there yet," he said. "It's getting harder to justify" the costs . Internally, Uber executives had begun referring to the problem by its name: "tokenmaxxing" .

The Wider Reckoning

Meta and other large firms have also scrapped or reviewed their AI usage rankings, and major corporations across the board are now reassessing runaway AI spending that hasn't produced a proportional return . The trend of tokenmaxxing, Fortune concluded, "is dead" .

The Hidden Costs: More Code, More Bugs, More Technical Debt

The backlash isn't just about budgets. A growing body of evidence suggests that AI-generated code is planting time bombs in software projects.

• The productivity illusion is real. The perception of speed is contradicted by controlled research. The focus on token volume and lines of code generated obscures the hidden cost of debugging, refactoring, and maintaining AI-written software. A GitClear report from January 2026 found that code churn—a measure of lines deleted versus added—had skyrocketed by 861% in environments with high AI adoption .
• Technical debt is accumulating. The rapid output from AI tools is producing code that requires more rework and introduces security vulnerabilities. Analysts warn that the resulting maintenance burden can easily offset any short-term speed gains, threatening the long-term health of codebases .
• The measurement itself is broken. The entire episode has revealed that activity metrics—lines of code, token consumption, number of AI agents invoked—are a deeply flawed substitute for measuring actual output. As one analysis put it, "subjective impressions were a poor guide to measured impact even as the effect size itself moved."

The Emerging Playbook: How to Actually Balance AI and Human Oversight

The 2026 reckoning is producing a new, more sober consensus on how to integrate AI into software engineering.

• Abandon token-based metrics. Companies are scrapping leaderboards and moving toward engineering intelligence that tracks shipped features, code quality, and defect rates rather than AI consumption volume. Atlassian's $1 billion acquisition of developer experience platform DX exemplifies this investment in measuring meaningful outcomes rather than raw activity .
• Mandate rigorous human review. The METR study showed that without careful oversight, AI tools can make experienced developers less productive. Code review and architectural planning remain critical, not optional, activities .
• Cap AI spend by demonstrated value, not usage. Uber's experience demonstrated that uncapped, usage-based AI spending is unsustainable. Budgeting must be tied to measurable returns, not adoption targets .
• Redesign the evaluation culture. The goal is to shift the core question from "who uses AI the most?" to "what useful outcomes did AI enable for our team?" This means rewarding judicious and effective use of AI, not just high-volume consumption .

The lesson from 2026 is clear: AI coding tools have created an unprecedented psychological and operational dependency before they have reliably proven their economic value. The companies that navigate this paradox will be those that treat AI as a tool to be mastered with discipline, not a god to be fed ever-larger token sacrifices.