AI Trust Among Developers: Writing Code vs Running Production
This is the finding I'd put on the whiteboard. Here is what developers who use AI actually use it for, as reported:
| Use | Share | Source |
|---|
| Generating code in familiar areas | 67% | Stack Overflow[1] |
| Debugging | 61% | Stack Overflow[1] |
| Analyzing data, logs, metrics, or query results | 41.9% | Virtualization Review[5] |
| Creating or configuring dev and test environments | 28.1% | Virtualization Review[5] |
| Deploying, operating, or troubleshooting production | About 20% | Stack Overflow[1] |
2 caveats before reading too much into the table. First, Stack Overflow's figures cover respondents who use AI. Virtualization Review reports the production figure as 19.9% across 30,903 valid responses, which looks like an all-respondents base, but I couldn't confirm the denominator.[5] Rows from different sources may not be measured against the same base. Second, some outlets report the code-generation and debugging figures as 69% and 64%. The direction holds either way; the precise decimals are softer.
The direction is clear, though. Usage falls off the closer the work gets to production. Roughly two-thirds use AI to write code; roughly one-fifth use it where a mistake pages someone at 3am.
Why the Line Falls Where It Does
The trust numbers explain the usage numbers. Only 6.6% would hand AI an important decision,[1] while about half trust output they can easily verify.[6] Developers trust what they can check.
That's my interpretation, and it fits every incident I've worked. Generated SQL is easy to verify: run it, diff the counts, read the plan. A production incident is the opposite. Half the evidence is missing, the logs lie by omission, and the fix you pick under pressure has a blast radius. I once spent 2 days chasing a pipeline that silently dropped records because an upstream team renamed a column and the job's fallback logic quietly defaulted it to null. No tool would have caught that from the stack trace, because there was no stack trace.
The 41.9% figure for analyzing data, logs, metrics, or query results is the most DE-shaped number in the whole survey.[5] It sits between writing code and touching prod. Plenty of people let AI help them read the evidence; far fewer let it act on that evidence.
Key takeaway: AI is trusted where its output is cheap to verify and avoided where mistakes are expensive. Production data work is the expensive end. That's where a Data Engineer's judgment is still the product.