AI Code Review Gap: Developers Distrust AI Code but Skip Checks

Distrust without verification

Developers are deeply skeptical of AI-generated code, and that skepticism is not translating into review. Sonar's 2026 State of Code survey of more than 1,100 professional developers found that 96% do not fully trust what AI tools produce. Only 48% say they always verify AI-assisted code before committing it. 1 The problem grows with volume: the same survey found that 42% of committed code is now AI-generated or AI-assisted. 1

The usual framing treats trust as the core problem. On these numbers, the bigger issue is behavior. A team that distrusts its tools but reviews their output is running a working quality-control process. A team that distrusts its tools and commits anyway is relying on hope.

One reason review lags is effort. In Sonar's survey, 38% of developers said reviewing AI-generated code takes more work than reviewing a colleague's. 1 That fits a pattern many engineering leads will recognize. Machine output often looks plausible and polished, but it does not carry the context a human author brings to a code review discussion.

Trust is falling as usage rises

Other large 2026 developer surveys show the same tension. Between 84% and 90% of developers now use AI coding assistants every week. Trust in AI output has fallen to roughly 29–33%, down from 40–43% a year earlier. 3 The trust figures differ from Sonar's because of how the question is worded. "Not fully trusting" is a lower bar than active distrust. Still, both sources point the same way: adoption keeps climbing while confidence slides. 13

That divergence is unusual for a developer tool category. Normally, familiarity breeds trust. With AI assistants, heavier use appears to expose more failure modes.

What slips through

The security research suggests developers' distrust is justified. In a Spring 2026 study, Veracode ran 80 coding tasks across four languages. Roughly 45% of AI-generated samples introduced an OWASP Top 10 flaw when no explicit security guidance was given. Cross-site scripting was handled securely only 15% of the time, and log injection just 13%. 1

A broader Veracode evaluation of more than 100 models found that AI-generated code carries 2.74 times the vulnerabilities of human-written code. It fails secure-coding benchmarks 45% of the time. 2 Notably, the 2026 update found that newer, larger models scored no better than smaller ones. 2 That undercuts the common assumption that model improvements will fix the problem on their own.

Package hallucination adds another risk. Researchers from the University of Texas at San Antonio, Virginia Tech, and the University of Oklahoma tested 16 code-generating models across 576,000 samples. They found that 19.7% referenced a package that does not exist. 2 Attackers can register those phantom package names, so unreviewed AI code becomes a supply-chain entry point.

The supply chain is already under pressure. In March 2026, an attacker backdoored the GitHub Action for Trivy, a widely used open source security scanner. The compromised action ran inside the CI/CD pipeline of LiteLLM, a gateway that routes model requests for frameworks including CrewAI, DSPy, and Microsoft GraphRAG. There it harvested the credentials LiteLLM uses to publish to PyPI. 2 That incident was not caused by AI-written code. It does show how much damage a single unvetted link in an automated pipeline can do.

Why the reviewer matters too

Cybersecurity Insiders argues that the same AI that writes code may be poorly suited to reviewing it. 2 The logic is straightforward. If the model that produced the code also checks it, its blind spots may be correlated with its errors. Teams that respond to the review burden by adding an AI reviewer from the same model family may get the appearance of oversight rather than real oversight.

Competition shifts to trust

For vendors, this changes how the market competes. After 2026 repricing, Cursor, Copilot, and Devin Desktop/Windsurf all cluster around similar $20, $60–100, and $200 tiers. 3 Claude Code is pulling ahead in satisfaction as a favorite among developers. 3 With prices converging, the argument goes, differentiation will come from closing the trust gap rather than winning feature comparisons. 3

The sources point to a slightly different conclusion. Trust is a symptom. The underlying problem is a review process that has not scaled with AI's share of the codebase. The tools that matter most in 2026 may be the ones that make verification cheap enough for developers to actually do it. That could mean independent analysis, dependency validation, and security checks built into the commit path.

Until then, the most telling figure is the gap between 96% and 48%. 1