Why 96 percent of developers distrust AI code, but only half of them review it
Almost every developer says they wouldn’t blindly trust AI-generated code. Yet only about one in two consistently reviews it before committing. A recent survey of more than 1,100 developers shows: >96 percent consider AI code not fully trustworthy, but only 48 percent verify it consistently before committing (Sonar, 2025). Between distrust and review behavior lies a gap that can’t be explained by a lack of awareness.
The reason runs deeper: it’s the speed itself. Anyone who generates working code prompt by prompt in a state of flow loses exactly the critical distance they otherwise believe is necessary. For engineering leaders in medtech, automotive, or financial services, this isn’t just an academic observation: every unreviewed line of code can later become an audit risk.
The flow moment: Why speed feels like progress
The flow moment: Why speed feels like progress
Practitioners describe vibe coding again and again in the same terms: “instant success,” “flow,” “addictive momentum.” In an analysis of practitioner reports, this experience was by far the most common theme, cited in 64 percent of cases (Fawzy, Tahir & Blincoe, 2025). A complementary qualitative study confirms this: developers actively seek out this state because it’s enjoyable and reduces cognitive load. They let the AI handle the syntax and focus on the idea instead (Pimenova et al., 2025).
That very relief is the mechanism that creates risk. Anyone who delegates syntax and detail to the AI also, often unnoticed, delegates the review. The widely cited METR study on developer productivity provides evidence for this: experienced open-source developers working with AI assistance were actually 19 percent slower, yet subjectively felt about 20 percent faster (Becker et al., 2025). Perceived speed and actual review depth diverge.
Automation bias: The blind spot speed creates
The phenomenon behind this has long been documented in research: automation bias, the tendency to question automated output less critically than work produced by oneself. The most comprehensive systematic review to date analyzed 74 studies and found that faulty automated decision support increased the risk of an incorrect human decision by 26 percent (Goddard, Roudsari & Wyatt, 2012). Vibe coding carries this effect into software development, only at a considerably higher cadence than in the contexts originally studied.
In practice, this shows up in two ways. First, in the verification gap already mentioned: 66 percent of surveyed developers say AI code “looks correct but isn’t reliable,” and 38 to 40 percent even find reviewing AI code more effortful than reviewing human-written code (Sonar, 2025). Second, among less experienced team members: a randomized study with predominantly young software developers found that the group using AI assistance scored 17 percentage points lower on a comprehension test than the control group, almost two full grade levels (Shen & Tamkin, 2026). Anyone who learns inside the flow without consciously understanding fails to build exactly the review competence they’ll later need.
For regulatory affairs and compliance leaders, this shift is especially sensitive: documentation requirements under IEC 62304 or ISO 26262 demand traceable review decisions, regardless of how quickly the code was produced.
Regaining critical distance: What works in practice
The good news: the flow effect can’t be switched off, but it can be deliberately contained. Three approaches have proven effective in practice:
- Limit diff size. A ten-month code review study at Cisco Systems, covering 3.2 million lines of code, shows: beyond 400 lines, the ability to find defects drops noticeably. In 200 to 400 lines reviewed over 60 to 90 minutes, by contrast, 70 to 90 percent of all defects are found (SmartBear, 2006). Large AI-generated diffs undercut exactly this threshold.
- Verbalize verification. Actively naming test runs, diffs, and assumptions, instead of silently accepting them, breaks the automatism of trust.
- Understand before adopting. The same study also shows: usage patterns aimed at understanding, rather than pure delegation of code, achieved over 65 percent comprehension scores versus under 40 percent for pure delegation (Shen & Tamkin, 2026). What matters less is whether AI is used, and more how.
sepp.med relies on exactly these principles in AI-assisted software testing: AI takes over routine tasks. Experienced testers validate critical test cases. And the review process stays documented and audit-ready.
Better experienced live than just read about
This exact tension, between the speed vibe coding promises and the depth of review that regulated software demands, is easy to describe in a blog article but hard to truly convey. Only when you watch live as code emerges from a prompt within seconds does the flow effect become tangible, with everything it brings in terms of opportunity and blind spots.
That’s exactly what’s at the center of the third session at the next sepp.med Afterwork Exchange on September 24, 2026, in Röttenbach: a live vibe coding session in front of an audience. Join the discussion on where, for your team, the line runs between productive speed and risky carelessness.
FAQ
No. The effect can’t be avoided, but it can be limited: small, manageable diffs, verbalized verification, and a deliberate usage pattern already make a significant difference.
No. According to Sonar’s data, the gap between distrust and actual review affects developers at every experience level. Early-career developers additionally build up less understanding along the way.
Automation bias is a phenomenon documented for decades in research on automated decision-support systems. Vibe coding carries the same effect into code creation, only at a higher cadence.
Code review data suggests an upper limit of around 400 lines. Beyond that, the ability to spot errors during review drops noticeably.
Sources
- Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the impact of early-2025 AI on experienced open-source developer productivity (arXiv:2507.09089). METR.
- Fawzy, A., Tahir, A., & Blincoe, K. (2025). Vibe coding in practice: Motivations, challenges, and a future outlook — a grey literature review (arXiv:2510.00328). ICSE SEIP 2026.
- Goddard, K., Roudsari, A., & Wyatt, J. C. (2012). Automation bias: A systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association, 19(1), 121–127.
- Pimenova, et al. (2025). Good Vibrations? Investigating developer experience and motivation in vibe coding (arXiv:2509.12491).
- Shen, J. H., & Tamkin, A. (2026, 29. Januar). How AI assistance impacts the formation of coding skills (arXiv:2601.20245). Anthropic.
- SmartBear. (2006). Best Kept Secrets of Peer Code Review (Cisco-Systems-Fallstudie über 3,2 Millionen Codezeilen).
- Sonar. (2025). Sonar data reveals critical “verification gap” in AI coding.