Coddy Tech's new developer survey doesn't show that AI coding tools are useless. It shows something more awkward: developers keep using tools they trust less because the career signal is getting harder to ignore.
You used to know when to stop. The code review queue would empty out, or you'd hit a bug at 9 p.m. that no amount of caffeine could push through. Either way, the work d itself. Coddy Tech's AI Coding Addiction Report, published on August 13, 2026 and later picked up by ZDNet, says that stopping point is disappearing. The company surveyed 305 developers who use AI coding tools at least weekly, and 80% said their AI use had felt more like dependence than advantage at least once.
The useful part of the survey is that it doesn't read like an anti-AI pamphlet. Developers are not saying the tools have no value. Forty-three percent said they'd kept coding with AI after hours when they meant to stop, 32% had put off sleep to keep going, and 39% said AI made it harder to switch off from work. Half said their employer now expects more output because of AI. The machine never looks tired. You still do.
The reward is real #
The promotion story is the part managers should read twice. Coddy Tech found that developers in the top third of weekly AI use, meaning 20 hours or more, were 74% more likely to have earned a raise or promotion than lighter users. Across the full sample, 74% said AI made them more ambitious. That is the bargain.
It also explains why the dependence number is so high. If your team is standardising on GitHub Copilot, Cursor, Claude Code, OpenAI Codex, you name it, choosing to sit out can look less like discipline and more like a career mistake. Junior developers feel that pressure first. JetBrains' 2025 Developer Ecosystem Survey found that 61% of junior developers described the job market as challenging, while 54% of senior developers said the same. That isn't a huge comfort gap, but it is enough to make early-career engineers think twice before becoming the person who says no to the tool everyone else is learning in public.
GitHub Copilot turns custom model access into a startup opportunity GitHub Copilot's support for custom endpoints gives enterprises more control over model choice, billing and security. It also opens a new distribution path for model startups trying to reach developers inside tools they already use. - how to use custom models with GitHub Copilot - building AI startup opportunities in developer tools market
The risk is not imaginary. Coddy Tech found that heavy users were 51% more likely to report burnout, and its Hooked group was far more likely than other developers to say work had become hard to disconnect from. Claude Code was the tool developers most often named as hardest to put down, at 35%. Codex users stood out on after-hours work, with 62% saying they kept going past the point they meant to quit.
The trust problem is worse #
This is where the story gets sharper. Developers are spending more time with tools they believe in less. Stack Overflow's 2025 Developer Survey found that only 29% of respondents trusted AI outputs to be accurate, down from 40% in 2024, while 46% actively distrusted them. That is not confidence. It is managed doubt.
You can see the cost in the daily workflow. A model gives you a function, a migration, a test suite, then you have to decide whether it is right enough to ship. Stack Overflow said the top frustration was AI answers that are almost right but not quite, cited by 66% of developers, and 45% said debugging AI-generated code takes more time. That is a very specific kind of exhaustion: not writing from scratch, but checking a machine's work while the machine keeps offering more work to check.
That friction used to help. Waiting for a teammate, getting stuck, walking away from the desk out of plain frustration, these were not productivity tools, but they created exits. AI assistants remove many of those exits. There's no queue. No wall. No natural stopping point. The next suggestion is always sitting there, cheap and close.
Frankly, companies that measure AI success by usage alone are asking the wrong question. More prompts can mean better output, but they can also mean longer sessions, weaker recovery, and more code that a tired person has to verify at the end of the night. The Coddy survey doesn't prove AI coding tools are bad for developers. It proves that usage is a lousy substitute for judgment.
Also read: Anthropic Agrees to Pay Nscale $45 Billion for AI Computing Power • MiniMax's Revenue Nearly Quadrupled While Its Losses Kept Growing Too • Z.AI Reveals It Secretly Built Ox Alpha, the Model That Beat DeepSeek