The Addiction Is Real. The Results Are Murky.

Here's where we are in 2026: Developers are so hooked on AI coding tools that researchers literally cannot run experiments without them.

In February 2026, METR — a respected AI research lab — tried to update a landmark productivity study. Simple ask: code some tasks by hand so we can measure the difference. Developers said no. According to TechCrunch, METR's own researchers admitted devs refused to participate "because they do not wish to work without AI even just for the study."

They won't go without it for a few hours. For science.

The Original Data Was Already Damning

The study METR was trying to update found AI actually slowed developers down.

Developers reported feeling more productive. The stopwatch said otherwise. AI generated code faster, sure — then developers spent extra time finding errors, fixing hallucinations, steering the tool, and waiting on it to finish. Net result: slower output, not faster.

When METR ran a self-reported survey in May 2026 — because they couldn't get anyone to turn the AI off — developers said AI made them twice as valuable to their organizations. Self-perception versus measured reality reveal a significant gap.

Companies Are Finding Out the Hard Way

The real-world corporate data is brutal.

Amazon built an internal leaderboard called Kirorank to track employee AI token usage as a productivity metric. Employees gamed it — burning through AI agent compute just to climb the rankings, not to produce anything useful. According to the Financial Times, Amazon shut it down.

Uber blew through its entire 2026 AI budget in four months. COO Andrew Macdonald said on a podcast — reported by The Information — that the spending hadn't led to a measurable increase in projects or productivity. ZERO. They spent the year's budget in a third of the year and got nothing concrete to show for it.

This phenomenon even has a name now: tokenmaxxing — using AI token consumption as a proxy for productivity. TechCrunch reported it may already be a dead trend, killed by its own absurdity.

The Maintenance Debt Nobody's Talking About

Programmer and author James Shore wrote a blog post that went viral on Hacker News making a crucial observation that the AI hype machine keeps ignoring.

"You write code twice as quick now? Better hope you've halved your maintenance costs."

Code doesn't end when it ships. Every line has to be read, understood, debugged, and updated by human engineers — potentially for years. AI-generated code doesn't reduce that burden. It may increase it.

A developer writing code understands why every decision was made. AI-generated code is a black box. It gives you an answer. It cannot explain the reasoning. It cannot participate in an architecture review. When it's wrong, there's no learning moment — just noise and cleanup.

The Senior Engineer Problem

Writer Devi Green interviewed 100 developers for Medium and found a consistent pattern at large, established companies: most engineers aren't using AI tools at all. Many haven't tried them.

The reason isn't Luddism. Senior engineers are professionally trained skeptics. Their entire job is to NOT trust things until proven. Then an AI tool shows up that confidently produces code that looks right but is subtly, dangerously wrong.

Green's own experience with GitHub Copilot: she spent more time debugging its suggestions than it would have taken to write the code herself. The review burden is real — a junior developer writing bad code is one thing, but an AI producing bad code at ten times the speed is a different category of problem.

Dev.to contributor david duymelinck put it bluntly: "The marketing is you will do all the exciting stuff. The reality is that we will be doing more boring stuff — like reviewing, and reviewing, and reviewing."

What Mainstream Tech Coverage Gets Wrong

Most tech media frames this as a battle between AI skeptics and AI enthusiasts — old guard versus new wave. The actual story is about measurement.

Companies adopted AI tools at scale before anyone built real accountability systems. Productivity claims were self-reported. Costs were not tracked against outcomes. The assumption that more AI usage equals more value was baked in from the start — and nobody demanded proof.

Now the bills are coming due. Amazon's Kirorank fiasco and Uber's budget blowout aren't anomalies. They're early signals that the industry skipped the validation step entirely.

There's also a security and IP story being undercovered. According to Green's reporting, the number one blocker at large enterprises isn't developer resistance — it's legal and security teams asking whether proprietary code is being sent to AI companies for training. For a long time, the honest answer was "maybe." That's a legitimate concern, not corporate paranoia.

What This Means for Regular People

If you're a business owner paying for software development — ask harder questions. "We use AI" is not a productivity guarantee. Demand measurable output metrics, not token counts.

If you're a developer early in your career, the dependency risk is real. The engineers raising the alarm about skill atrophy aren't afraid of technology. They're pointing out that you can't guide a tool you don't understand. And when the AI is wrong — which it is, regularly — someone has to know enough to catch it.

The developers refusing to code without AI are betting their careers on a tool whose ROI the industry has NOT actually proven at scale. That's not a bold bet. That's an unchecked assumption with a big price tag.

Amazon and Uber are learning that lesson right now. With their money. Not the AI vendors'.