A survey of 305 professional developers published by the coding-education platform Coddy found that 71 percent have shipped code they did not fully understand, and that four in five describe their relationship with AI coding tools as closer to a dependence than an advantage. The study, fielded in July 2026 and picked up this week in an analysis by The New Stack, is one of the first attempts to measure what agentic coding costs developers behaviorally rather than what it delivers in throughput.
Key takeaways
- Coddy surveyed 305 developers who use an AI coding tool at least weekly; 80 percent said their use had felt more like dependence than an advantage at least once, and 59 percent said they feel that way often or sometimes.
- Unplanned after-hours coding split sharply by tool: 62 percent among OpenAI Codex users, 45 percent for Google Gemini, 40 percent for Claude Code and 36 percent for GitHub Copilot.
- 71 percent admitted shipping code they did not fully understand, rising to 77 percent among Gemini users and 79 percent among Gen Z respondents.
How dependent are developers on AI coding tools?
Respondents were recruited through Prolific and CloudResearch Connect, ranged in age from 18 to 74 with an average of 38, and all used an AI coding assistant at least once a week. Coddy sorted them by how many compulsive-use signals they displayed. Forty-one percent showed three or more and landed in the tier the report labels "Hooked."
The dependence figure is the study's anchor, but the stickiness numbers are harder to dismiss. Only 17 percent said they would stop using AI tools for a single month if asked, and 67 percent said they would be uncomfortable if the tools disappeared tomorrow. Eighty-two percent named at least one tool they find difficult to abandon.
The cost lands on calendars. Thirty-six percent said they had skipped breaks or time off to keep a session going. Thirty-two percent delayed or skipped sleep, a figure that climbs to 40 percent among senior developers. Twenty-eight percent missed meals, 26 percent dropped errands and 23 percent skipped exercise.
Why does the tool you use change the behavior?
The most striking split in the data is between products rather than people. Codex users reported unplanned after-hours coding at 62 percent, 26 percentage points above GitHub Copilot users at 36 percent, with Gemini and Claude Code in between. That is a wider spread than the gap between seniority levels or generations in the same dataset.
The rankings do not line up neatly. Claude Code was named the single hardest tool to put down by 35 percent of respondents, more than any other product, yet its users reported the second-lowest after-hours rate. Being hard to quit and keeping developers at the keyboard past dinner appear to be different properties, and the survey does not disentangle whether the difference comes from the tools' autonomy levels, their pricing models or the kind of work each attracts.
Who gets rewarded for the dependence?
The incentive picture is where the report gets uncomfortable for engineering managers. Among the "Hooked" group, 75 percent credited AI with raises or promotions, compared with 39 percent of the developers who described their usage as controlled. The heaviest users, those logging 20 or more hours a week with AI tools, reported the strongest career returns.
That sits alongside the 71 percent who have shipped code they did not fully understand. Coddy's breakdown puts Gemini users at 77 percent and Codex users at 75 percent on that question, with Gen Z respondents highest at 79 percent. The survey does not claim those developers shipped defects, only that comprehension and delivery have come apart.
Writing in The New Stack, Octopus Deploy's Steve Fenton argued that the reward structure resembles operant conditioning: tight generate-and-check loops supply frequent small payoffs, and managers then reward the developers who sit in those loops longest. His read is that organizations end up paying for volume of output while the decisions that actually determine software value get made by tired people.
What it means for engineering teams
The findings arrive as coding agents move from autocomplete to unattended execution. Terminal-native agents such as Meta's Muse Code are designed to run long stretches without a human in the loop, which shortens the review window that used to force comprehension. Coddy's data suggests that review step is already being skipped at scale.
The report's own recommendation is unglamorous and specific: the developers who reported the healthiest relationship with the tools were not the ones who quit, but the ones who set narrow limits, reserving AI for defined task types and protecting off-hours. Anthropic, OpenAI and GitHub have all shipped usage dashboards and session controls over the past year; none of them currently surface comprehension as a metric.
Outlook
Expect the comprehension gap to become a compliance question before it becomes a wellbeing one. Regulated industries already require that shipped code be attributable and reviewable, and a 71 percent figure is hard to reconcile with that. The more immediate test is whether any large engineering organization publishes its own version of these numbers.
FAQ
How many developers were surveyed in the AI Coding Addiction Report?
Coddy surveyed 305 developers in July 2026, recruited through Prolific and CloudResearch Connect. All respondents used an AI coding tool at least once a week, and they ranged in age from 18 to 74 with an average age of 38.
Which AI coding tool keeps developers working latest?
OpenAI Codex, by a wide margin. Sixty-two percent of Codex users reported continuing to code after hours when they had meant to stop, compared with 45 percent for Google Gemini, 40 percent for Claude Code and 36 percent for GitHub Copilot.
Does the survey prove AI coding tools cause burnout?
No. It is a self-report survey of 305 developers at a single point in time, so it measures correlation and perception rather than causation. It does not track health outcomes, and it cannot separate tool design from the workloads and deadlines each tool is typically used for.
