The cost of an answer that is almost correct
The selected problem responses are nearly correct solutions at 66%, longer debugging at 45.2%, reduced confidence in problem solving at 20%, and difficulty understanding generated code at 16.3%. Respondents could select several answers.
The percentages count people who report a frustration. They do not measure the proportion of generated outputs that contain an error, the time spent correcting one, or the severity of a failure. Those would require different observations.
Define a check before accepting the output
Citation's practical reading is to decide how an answer will be checked before treating it as complete. A code change needs behavior that can be exercised. A research claim needs a source that actually supports it. A calculation needs inputs that can be independently recomputed.
For a generated paragraph, choose one consequential claim and trace it to the original evidence. Check the population, date and measure. A plausible explanation can still connect the wrong number with the wrong group, just as plausible code can still implement the wrong behavior.
Preserve understanding during revision
When an answer fails, identify the smallest part that needs correction and explain why. Replacing the entire output repeatedly can make it harder to see what changed. A short explanation of the correction helps the person using the tool retain ownership of the result.
This process is a proposed response to the kinds of friction in the chart. The survey does not establish that any one review method eliminates those problems. A team would need to observe its own error patterns and revision effort to evaluate the approach.
How to read the data
There were 31,476 responses to the frustrations question. We show the four concrete problem categories, excluding other answers, infrequent use and no problems. The original labels are retained in the CSV. Multiple selections mean that the displayed values must not be added to count distinct respondents.
The original report provides the published figures and question wording. The chart and downloadable data use the same values; labels are shortened for readability.
Questions about the findings
Does 66% mean that most AI answers are wrong?
No. It is the share of respondents who selected the nearly correct solutions response. The question does not count individual generated answers.
Why are some responses absent?
This chart focuses on specific problems. The original survey also includes other responses and options describing no problems or infrequent use; the linked source contains the complete chart.