Developers use AI daily and trust it only on checkable work

Developers now use AI every day, and the largest group among them trusts its output only when they can check it easily. In Stack Overflow's 2026 survey, 48% gave that answer, while 6.6% said they trust AI for many tasks, including important work decisions.

How much do developers use AI now?

Stack Overflow published its 2026 Developer Survey on 6 October 2026. Its methodology page gives 30,903 valid responses from 169 countries, collected between 23 June and 5 August 2026. The survey was promoted through Stack Overflow's own site, blog and email list, so the respondents are people already close to that community.

On the AI page, 10,926 respondents answered a question on daily AI use. Of them, 30.9% use AI four or more hours a day, 24.1% two to four hours and 25.3% one to two hours. That adds up to 80.3% using it at least an hour a day. A separate figure in Stack Overflow's own summary is that 73% of respondents who use AI coding assistants or agents use them daily. These are two different bases, and we use them as Stack Overflow states them.

JetBrains gives a second view. In its Developer Ecosystem Survey 2026, which it describes as globally representative and based on more than 15,000 professional developers, 90% used AI coding agents at work at least weekly in May to July 2026, and 68% used them daily. "Professional developers" in that report includes developers, AI and ML engineers, DevOps engineers, architects, data scientists and QA engineers. Google's 2025 DORA report, based on nearly 5,000 technology professionals, found 90% using AI at work in 2025. Three sources use three definitions of use, and all of them point in the same direction.

How much do they trust it?

The 6.6% figure is real, and it has a narrow definition. On the overview page of the 2026 survey, the question is "How much do you trust the output from AI tools or AI agents as part of your workflow?" There were 14,304 respondents. The largest group, 48.0%, chose "I trust it when I can easily verify the output". Two groups of 16.3% each followed: "I trust it for many tasks, but not important work decisions" and "I trust it for low-risk work tasks only". Another 12.3% chose "I do not trust it for most work tasks", 6.6% chose "I trust it for many tasks, including important work decisions", and 0.5% were not sure or did not use AI tools.

The 6.6% is therefore the share who extend trust to important decisions. It measures willingness to rely on AI for the weightiest choices, and says nothing directly about how accurate respondents think AI is or how many distrust it. The same overview page says that more than 87% had some level of trust, "a huge increase from 31% last year". The 87% is everyone except the 12.3% who do not trust it for most tasks and the 0.5% unsure or not using it, so it includes people who trust AI only for low-risk work. We could not tie the 31% to a 2025 question asked the same way, so we make no claim about a trend.

The 2025 survey asked something else. In the 2025 AI results, people were asked how much they trust the accuracy of AI output: 3.1% highly trust it, 29.6% somewhat trust it, 26.1% somewhat distrust it and 19.6% highly distrust it. Stack Overflow summarised that as 46% distrust and 33% trust.

An independent survey points the same way. Google's DORA report of September 2025 found that 24% trust AI "a great deal" (4%) or "a lot" (20%), while 30% trust it "a little" (23%) or "not at all" (7%). The Google Cloud announcement puts it as 30% reporting little or no trust in the code generated by AI. The rest sit in between.

JetBrains' 2025 survey, with 24,534 developers, lists the top concerns about AI in software development in this order: inconsistent quality of AI-generated code, limited understanding of complex code and logic, privacy and security risks, a possible negative effect on developers' own skills, and lack of context awareness. The post gives the ranking and no percentages for it.

Is "developers use AI more than they trust it" a fair summary?

In direction, yes. In detail, it needs care. Stack Overflow asks both questions in the same survey, but the usage question has 10,926 answers and the trust question 14,304, and the pages we read publish no cross-tabulation of the two. We cannot say how many of the daily users also trust AI for important decisions.

Our reading is that "trust" here is mostly a matter of task and checkability. Stack Overflow's own page says the sweet spot is trusting AI "when they understand what it's working on", for example familiar code (67% currently use AI with this) or debugging and refactoring (61%). People use AI heavily where they can see whether it is right. The figures do not show that they use it blindly, and they do not show that they distrust it everywhere.

What does this mean for code review?

JetBrains Research published a study in October 2026 on reviewing AI-generated code, based on a design study with 17 practitioners and a survey of 43 software professionals. That is a small sample, so we treat it as a view from researchers and give it less weight than a measurement. Its main message is that "the core problem with reviewing AI-generated code is trust calibration", meaning the ability to spend review effort in proportion to the risk of each part of a change when you cannot ask the author. The post adds that a language model presents every line with the same apparent confidence.

The survey data adds one more point. In Stack Overflow's 2026 results, 58.1% of respondents to its task question (6,915 answers) use AI to write or improve tests. Our reading is that when the same kind of tool writes both the code and the tests, a passing test run says less than it did when people wrote the tests independently. Neither survey tests that claim. It is a reason to keep a person responsible for deciding what must be tested, and to check what the tests actually assert. That is the work we describe on our QA and test automation page.

What should a buyer ask a software supplier?

This section is our addition. It turns the figures above into questions for a contract or a supplier meeting. It is not legal advice, and a lawyer should check the wording for a specific contract.

The first group of questions concerns where AI is used. Ask which tasks use AI assistants or agents, and which never do, since Stack Overflow's respondents most often name familiar code, debugging and answering technical questions. Then ask whether AI-written code goes through the same review and the same tests as other code, and who signs off. The supplier, not the tool, should remain accountable for a defect. A related question is who writes the tests and how you can see what they cover, so ask for the test plan and the results of each release.

The second group concerns data and change. Ask which AI tools see your code and data, and under what terms, and whether those terms exclude training on your material. Ask also for notice when the tooling changes. In JetBrains' data, use at work of Claude Code, which JetBrains calls the most widely adopted AI coding tool at work, rose from 18% of professional developers in January 2026 to around 39% in May to July 2026, so the tools named in a contract may not be the ones in use a year later.

The last question is where a named person makes the final call on important changes. The reason to ask is that only 6.6% of Stack Overflow's respondents trust AI with important work decisions.

If you want to see how we handle this on our side, our AI-native development page describes eval sets, regression checks and a person in the loop, and how we work describes the three steps of every project.

What we could not confirm

Stack Overflow's AI page text calls the 31% a share of "daily AI users" and says over 70% use AI daily, but its chart lists all 10,926 answers, including people who use AI less than daily. We report the chart's rows as shares of those who answered. Stack Overflow also publishes no cross-tabulation of daily AI use and trust on the pages we read, so we cannot say how many heavy users trust AI with important decisions.

Stack Overflow's task list gives n = 12,547 but its percentages do not match the counts over that base (6,915 is 55.1% of 12,547, not 58.1%), and its text quotes 67% and 61% where the list shows 69.3% and 63.8%. We report the figures as printed and name the source. We could not tie the overview page's "31% last year" to a 2025 question worded like the 2026 one, so we make no claim about a change over time.

The JetBrains 2026 posts we found report usage and tool adoption, and they give no trust or concern figures. The only JetBrains concern ranking we found is from the 2025 survey, without percentages. No second survey from 2026 asks the same trust question, DORA's figures are from 2025, and we found no 2026 DORA report. Stack Overflow and JetBrains describe their samples in their own words, and we have not checked how representative either one is.