A survey of 305 AI-using developers finds 71 percent shipped code they did not fully understand

Coddy, the coding-education startup, ran a survey of 305 developers who use AI tools at least once a week and published the results on 10 September. The New Stack wrote it up on 17 September. Two of the numbers are worth reading twice: 71 percent said they had shipped code they did not fully understand, and heavy users (20 or more AI hours a week) reported burnout at a rate 51 percent higher than the lighter-use group.

The data was collected in July 2026, with an average respondent age of 38: 51 percent Millennial, 25 percent Gen Z, 20 percent Gen X, 4 percent Baby Boomer. Coddy classifies 41 percent of the sample as "Hooked", meaning they showed signs of compulsive use, and 43 percent said they keep coding after hours despite meaning to stop. Kevin Spektor, Coddy's co-founder and CTO, is the named author of the report.

The tool split

Codex users were the most likely to say they got into a "just one more loop" habit, at 79 percent. After-hours coding rates split by tool: Codex 62 percent, Gemini 45 percent, Claude Code 40 percent, GitHub Copilot 36 percent. Coddy's own writeup puts Claude Code first for "hardest to put down", at 35 percent. The New Stack's coverage is more direct: managers are rewarding volume, and the reward looks real. Heavy users were 74 percent more likely to earn a raise or promotion than lighter users in the same period, and 50 percent said their employer now expects more output than before AI.

What people give up for the extra output

The list of things respondents said they delay or skip while in an AI coding session: sleep 32 percent, breaks and time off 36 percent, meals 28 percent, errands 26 percent, exercise 23 percent. Senior developers sacrificed sleep at 40 percent. Sixteen percent said AI use had strained a personal relationship. Eighty percent said their AI use had felt like dependence, and 67 percent said they would be uncomfortable if the tools disappeared. Only 17 percent said they would be willing to stop for a month.

Volume as the incentive

One number carries the whole finding: 71 percent shipped code they did not fully understand, and heavy users were 74 percent more likely to earn a raise or promotion in the same window. Two facts pointing the same way say something about how the work is measured. When code is generated faster than any human reads it, the review step is where quality lives now, and staffing the team to hit an output target that was set before agents will produce the pattern Coddy measured.

Coddy is an education vendor with a commercial interest in developers thinking hard about their own AI use, so read the qualitative framing with that in mind. The raw numbers come from a self-report survey, so the 71 percent means "developers admit to shipping code they did not read carefully". A code audit would probably find a higher rate, because the pattern people are aware of and willing to state is smaller than what they actually do. The tool split is small-sample and Coddy does not report confidence intervals, so treat the Codex-versus-Copilot gap as directional.

Source

• Coddy: AI Coding Addiction Report, the primary report from 10 September 2026.
• The New Stack: Developers are addicted to AI, and managers are making it worse, 17 September, by Steve Fenton.