Developer trust in AI: why review still matters

In Stack Overflow's 2025 AI accuracy chart, 26.1% of respondents somewhat distrust AI output and 19.6% highly distrust it. Together, those published values total 45.7%. Somewhat trust accounts for 29.6% and highly trust for 3.1%, totaling 32.7%. The difference between the two combined shares is 13 percentage points.

These answers describe confidence in the accuracy of AI output within development workflows. They do not measure whether a particular generated answer passes a test. Someone can use a tool frequently and still decide that every consequential result needs review.

Citation's interpretation is that a general feeling of confidence is less useful than a defined check for the work in front of you. For a code change, the check might involve reproducing the reported problem and exercising the affected behavior. For a research paragraph, it might involve opening the original source and checking what the quoted figure actually measures.

The review should follow the claim. If a draft says a proportion doubled, inspect both values and confirm that the population and period are comparable. If a summary describes a study as causal, look at its design. If a generated citation looks complete, verify that the publication exists and supports the nearby statement.

This makes review an observable action. A colleague can see which claim was checked, which source resolved it and which uncertainty remains. The process does not depend on whether the prose sounds confident or whether the reviewer generally likes the tool.

A suggested heading can be judged quickly for clarity and relevance. A table that feeds a published conclusion needs a more deliberate check of definitions, calculations and missing values. Both deserve attention, but they ask different questions of the reviewer.

Before using an output, identify what would change if it were wrong. Then choose a check that could expose that failure. A calculation needs recomputation. A factual claim needs evidence. A proposed interpretation needs an examination of alternatives. Repeating the original request without adding independent evidence is not a substitute for that check.

The source displays four trust responses in its All Respondents chart. We reproduce those values without filling in the undisplayed answers or renormalizing the bars. Their total is 78.4%, so the table deliberately has no 100% total row. The original percentage scale remains intact.

This is a useful reading habit for any chart: first establish whether its categories are exhaustive. A table of selected responses can answer a focused question without describing every possible answer. Adding an assumed category would make the presentation look complete while introducing a label the displayed source does not establish.

The accuracy question has 33,244 responses. Stack Overflow fielded the survey from May 29 to June 23, 2025, recruiting mainly through its own channels. This approach can favor people more engaged with Stack Overflow, so the findings should not be treated as a random sample of all developers.

See the original AI accuracy results and survey methodology for the question and recruitment details.