Stack Overflow Survey: AI Coding Use Soars as Developer Trust Falls

The headline numbers​

The retrospective compares the 2024 and 2025 surveys and adds figures from a smaller pulse survey run in April 2026. Its year-over-year series counts developers who currently use AI tools:

Stack Overflow's July 2025 launch announcement used a different measure, so it gave different numbers. In that release, 84% said they use or plan to use AI tools in their development process, up from 76% in 2024. The two sets of figures don't conflict. One counts current users only, and the other adds people who plan to start. If you quote either one in a budget meeting, say which you mean.

The sample size needs a correction. Some coverage of the retrospective says the survey reached "more than 27,000 developers." That undercounts it. The press release says Stack Overflow received over 49,000 responses from 177 countries across 62 questions focused on 314 different technologies. The official developer profile page puts the exact total at 49,019. The 27k+ figure in the retrospective is the number of people who answered one question, about how many tools they use at work.

Section summary: Current AI use rose from 44% to 79% in two years. "Use or plan to use" reached 84%. The survey had about 49,000 respondents, not 27,000.

Sentiment is going the other way​

Across all respondents, favorable sentiment toward AI tools fell from 72.0% in 2024 to 59.7% in 2025. Negative sentiment went from 6.4% to 20.4%, more than three times higher.

Trust in accuracy looks worse. Stack Overflow found that 46% of developers said they don't trust the accuracy of the output from AI tools, a significant increase from 31% last year. Stack Overflow has published two different "trust" figures for 2025, so be careful which one you use:

• The official AI results page says 33% trust AI output accuracy.
• A February 2026 Stack Overflow blog post said only 29% of 2025 respondents said they trust AI, down 11 percentage points from 2024.

The difference probably comes from how answers are grouped or which respondents are counted. Either way, distrust clearly outweighs trust.

The survey also shows why. Two-thirds (66%) report that AI answers are "almost right but not quite," and 45% said they lose significant time debugging AI-generated code. Experienced developers are the most cautious group. The survey page records their lowest "highly trust" rate at 2.6% and their highest "highly distrust" rate at 20%. It reads that as a need for human verification in roles that carry accountability.

Section summary: AI use went up by double digits while favorable sentiment fell about 12 points. Distrust (46%) now clearly outweighs trust (roughly 29–33%).

Learners show the sharpest drop​

People learning to code show the widest gap between use and sentiment:

• Usage: up from 63.0% in 2024 to 73.3% in 2025. That includes 39.5% who use AI daily, 18.7% weekly and 15.1% monthly.
• Favorable sentiment: down from 72.2% to 52.8%. The 2025 figure breaks down as 19.3% "very favorable" and 33.5% "favorable."
• Negative sentiment: up from 6.4% to 26.3%, roughly four times higher.
• Moderate or high trust in accuracy: down from 49.1% to 37.4%.
• Favorable view of AI on complex work: down from 41.5% to 33.7%. Another 18.6% said they avoid AI for advanced tasks or aren't sure it's reliable.
• Agents: 28.4% of learners used them, and 44.1% said they had no plans to. Across all respondents, 37.9% had no plans.

Two caveats matter. First, the 2025 trust question added a neutral option, which moved some answers away from the trust categories. Part of the trust decline may come from that change in wording. Second, these are separate annual snapshots of different people. The survey did not follow the same learners over time.

Stack Overflow's own summary of the learner data says that usage might be accelerating but "healthy skepticism remains." That fits the numbers: students use the tools but don't simply believe what they produce.

The retrospective also flags a concern for how people learn. Among everyone who answered the AI frustrations question, not only learners, 20.0% said AI made them less confident in their own problem-solving. Another 16.3% said it was hard to understand how or why the generated code worked. Stack Overflow says directly that these answers don't prove AI weakens learning. Still, getting working code and being able to explain it are now separate skills, and new developers need both.

Section summary: Learners adopt AI quickly but trust it least. The data doesn't show learning getting worse. It does show that "the code runs" and "I understand the code" can now diverge.

Agents: "nearly doubled" needs context​

The most striking figure is agent adoption. The 2025 annual survey found 31% of respondents using AI agents. In the April 2026 pulse survey, 59% said they used them, and Claude Code usage rose from 41% to 55%.

The retrospective itself calls the pulse survey smaller. According to Stack Overflow, agent use in the pulse was concentrated among daily AI users and people in more executive roles. A small survey weighted toward heavy users and executives can't be compared directly with a 49,000-response annual survey. The rise is probably real, but the doubling is likely overstated. The 2026 annual results should give a better comparison.

For reference, the 2025 baseline was modest. A majority – 52% of developers – say they don't use agents or stick to simpler tools, and 38% have no plans to try AI agents anytime soon. Developers who did use agents reported mixed results. 69% say AI agents improved their workflow, 70% report reduced time on specific tasks. However, there's a caveat: only 17% believe AI agents have improved team collaboration, making it the lowest-rated impact.

Section summary: Agent use likely grew a lot. The 59% figure, however, comes from a smaller pulse survey weighted toward heavy users and executives.

Where Microsoft and GitHub stand​

Among 2025 respondents using ready-made AI assistants, the retrospective lists usage as follows. These are multi-select answers, not market share:

GitHub Copilot is clearly one of the two main entry points. Microsoft Copilot, the general-purpose assistant, sits at 31%, behind Gemini and Claude Code. Stack Overflow's 2025 press release reported that Visual Studio Code and Visual Studio were still the top IDEs. AI-first editors are gaining alongside them: Cursor at 17.9%, Claude Code at 9.7% and Windsurf at 4.9% in the work-tools data. For collaboration, Microsoft Teams grew from 25.6% in 2020 to 53.1% in 2024, while Zoom and Slack both declined.

The retrospective's conclusion is that AI tools are being added alongside the existing stack rather than replacing it. Docker (71.1%), PostgreSQL (55.6%) and Jira (46.4%) are still widely used.

What Windows and dev teams should do with this​

This is my own analysis, not survey findings:

1. Build a review step into AI use. 66% of developers see "almost right" answers and 45% lose time debugging AI code. Tests, code review and static analysis should apply to AI-generated code exactly as they do to human code. Treat AI output as untrusted until it has been checked.
2. Limit where agents can act. Developers themselves are cautious here: 76% don't plan to use AI for deployment and monitoring, and 69% don't plan to use it for project planning. Keep agents away from production pipelines until you have audit trails and rollback in place.
3. Coach juniors to explain their code. If a new hire can't explain what their AI-assisted pull request does, it isn't ready to merge, even if it passes the tests.
4. Measure your own results. Survey sentiment tells you how people feel, not how productive they are. Before renewing Copilot or agent licenses, look at defect rates and cycle times on your own team.
5. Keep people available for questions. In a hypothetical future where AI handles most coding, 61.3% of respondents said they would still ask a person when they wanted to fully understand something, and 58.1% to learn best practices.

What to watch in the 2026 results​

Stack Overflow says the 2026 data will test four questions:

• whether the gap between AI use and favorable sentiment persists
• whether more agent use comes with developers handing over more tasks and decisions
• how control over one's work relates to job satisfaction, separately from remote or office arrangements
• how AI affects community use, including visits for help with AI-related problems

Stack Overflow is not a neutral party here. It sells human-curated knowledge and data licensing, and its launch release argued for a trusted human intelligence layer in the age of AI. A trust gap supports its business case. Even so, the figures are consistent across three years and several ways of asking the question. Developers are using AI more, checking its output closely, and still relying on other people when it counts.