The « State of AI Code Quality in 2025 » report by Qodo, based on a survey of 609 developers, explores the evolving role of AI in software development. It highlights that while AI tools have become mainstream (82% daily/weekly usage, 59% use 3 or more tools), deep trust in their outputs remains elusive. AI tools significantly influence production code: 65% of developers report that at least 25% of their commits are AI-generated or AI-shaped.
Productivity vs. trust: the paradox
State of AI Code Quality in 2025
While 78% report productivity gains and 57% find their work more enjoyable, a major barrier persists: hallucinations. 25% of developers estimate that one in five AI suggestions contains errors, which weighs heavily on trust and adoption. This prevalence of hallucinations creates low confidence: 76% of developers facing frequent hallucinations are reluctant to ship AI code without human checks. Even among those with low hallucination rates, a majority (75%) hesitate to merge without manual verification.
Code quality and AI review: a key catalyst
Contrary to fears, increased productivity with AI often correlates with better code quality. 70% of developers reporting substantial productivity gains also report better code quality. AI-powered code review acts as a catalyst: 81% of fast-moving teams using AI for review report quality improvements, versus 55% without it. This automated validation helps maintain quality standards while accelerating delivery.
Context: a fundamental factor
The report identifies context as the #1 factor in perceived quality and trust. 65% of developers report that AI misses relevant context during refactoring, a more frequent problem than hallucinations themselves. Similar issues arise in test generation and code review. Developers overwhelmingly call for "better contextual understanding" from their AI tools. The report argues for persistent, automated context learning across the entire repository, as manual context selection is inefficient and frustrating.
The Confidence Flywheel
The report introduces the "Confidence Flywheel": a self-reinforcing cycle where context-rich suggestions reduce hallucinations, leading to correct code, increased developer confidence, faster delivery, and ultimately better examples fed back into the model. Only 3.8% of developers currently experience this ideal scenario, but they report higher quality gains and greater confidence.
Testing and confidence
Developers using AI for testing are 2x more confident in their test suites (61% vs. 27% for non-users), suggesting that full AI integration across the development cycle improves overall confidence.
Strategic conclusion
Qodo concludes that unlocking the full business value of generative AI requires bridging the gap between LLM capabilities and proven existing systems, with domain integration being critical. The report calls for an agentic code quality platform providing deep context awareness and integrating AI across the entire development cycle to strengthen code quality and developer confidence. This iterative approach, with technical guardrails, is highly effective but also reveals limits: human technical expertise remains essential to catch AI errors, and agents lack business context.