If AI coding is lowering your code quality, you're not managing quality right
AI-assisted coding is polarizing engineers who see it either as a powerful productivity boost or as a fast path to unmaintainable “slop” that overwhelms reviewers. Many argue that code quality ultimately depends on human process and judgment — rigorous specs, testing, reviews, and clear guardrails for agents — while others counter that in real-world teams with weak management, AI mostly amplifies existing bad habits and technical debt. Underneath the debate is a broader concern that organizations are chasing speed and volume of code without the culture, incentives, or safeguards needed to keep systems reliable over time.
Impact on Code Quality
- Strong split: some say AI can produce code as good or better than their own; others insist AI code is inherently sloppier or unmaintainable.
- Critics argue many developers can’t reliably judge “good” code, especially architecture, so they overestimate AI quality.
- Supporters claim recent models, used correctly, produce high‑quality code, especially for details, refactors, and legacy modernization.
- Several note that AI mostly amplifies existing practices: good engineers get somewhat better; mediocre ones produce 10× more bad code.
Process, Guardrails, and Testing
- Many comments stress that quality comes from classic practices: clear specs, acceptance criteria, code review, linting, typing, complexity checks, and strong tests.
- Positive, prescriptive instructions (“always search for helpers before writing one”) work better than “never do X” rules.
- Post‑implementation “audit passes” by agents to find duplicates, naming issues, and style violations are reported as effective.
- Heavy emphasis on tests: unit, property-based/fuzzing, differential tests, golden masters, coverage. Some argue coverage % alone is a poor quality proxy.
Productivity vs Review Burden
- Reported gains range from modest to “10×” on greenfield solo work; in team settings, 50–100% improvements are seen as more realistic.
- Others say AI doesn’t actually save much time once serious review and redesign are included.
- Review workload is a major pain point: AI enables huge PRs that humans can’t fully understand, especially in big orgs and open source.
Management, Incentives, and Responsibility
- Many argue AI isn’t “destroying codebases” on its own; humans and management processes decide what gets merged.
- Others counter that unrealistic velocity expectations make thorough review impossible, effectively forcing low-quality AI use.
- There’s debate over whether it’s fair to blame users for “holding the tool wrong” instead of demanding better defaults and safer behavior from tools.
Agentic and Recursive Use
- Recursive, agentic workflows (AI repeatedly modifying its own output) are seen as particularly risky, tending toward “random walk” spaghetti without tight human steering.
- Some think current prompting/guardrail techniques are insufficient to safely let agents run unsupervised on complex, non‑boilerplate systems.
Attitudes and Craft
- A recurrent theme: caring about code as a craft vs “just a job.”
- Those who treat AI as an assistant and maintain high standards report better outcomes than those who “vibe code” and ship whatever passes minimal checks.