Echoes of AI: AI does not worsen code maintainability

# Echoes of AI Study: AI Assistant-Generated Code Maintainability Matches Traditional

The "Echoes of AI" study shows that code generated with AI assistants like Copilot or Cursor does not degrade maintainability. In a randomized controlled experiment with 151 developers, code from the AI group was solved 13% faster, with comparable CodeHealth and test coverage. For experienced users, AI code quality was even higher.

Experiment Design

The experiment was divided into two phases. In the first, 151 participants (95% practicing developers) refined a Java/Spring Boot app called RecipeFinder (~2000 lines with code smells and bugs). Task: add recipe filtering by cooking time.

Participants were divided into groups:

• AI-devs: used AI assistants (Copilot, ChatGPT, Cursor).
• !AI-devs: no AI.

Stratified distribution by experience. Geography: global, demographics balanced.

In the second phase, 75 new developers (no AI) extended code from phase 1—adding filtering by portion cost. Participants didn't know the code's origin. Success checked via acceptance tests in GitHub Actions.

Maintainability metrics:

| Metric | Tool | Description |

|---------|------------|----------|

| Completion Time | Self-reported + CI | Phase 2 task time |

| CodeHealth (CH) | CodeScene | Aggregate of 25+ factors (nesting, God Classes, duplicates, bumpy road; scale 1–10) |

| Test Coverage (TC) | JaCoCo | Line coverage by tests |

| Perceived Productivity (PP) | SPACE survey | Subjective rating (Likert scale) |

Analysis: frequentist (Wilcoxon, t-test, bootstrap) + Bayesian (linear/logistic regression).

Phase 1 Results: Development Acceleration

AI sped up the task:

• All AI-devs: median +30.7%.
• Habitual AI users: average +55.9%.

Tasks felt realistic (median 4/5 similarity to production).

Phase 2 Results: Code Maintainability

AI code matched traditional:

| Metric | AI vs !AI Difference | Significance |

|---------|-------------------|------------|

| Completion Time | +13% faster | Not significant (moderate Bayesian evidence) |

| CodeHealth | +0.3 points | Not significant overall, significant for habitual (+ moderate Bayesian) |

| Test Coverage | No difference | Not significant |

| Perceived Productivity | No difference | Moderate Bayesian positive |

Code from habitual AI users had better CodeHealth. AI generates idiomatic, predictable code—this boosts maintainability.

Risks: Code Bloat and Cognitive Debt

Metrics miss:

• Code bloat: AI eases generation, risk of excess volume → complexity.
• Cognitive debt: reduced deep code understanding, skill atrophy, Gell-Mann amnesia effect (less critique of weak languages).

Ties to DORA and Practices

DORA 2025: AI amplifies practices. Winning teams have:

• Small batches.
• Fast iterations (tests, reviews, refactoring).
• Modular architectures.
• End-to-end focus.
• Autonomy.

90% use AI daily, 80% see productivity gains, but stability drops without practices.

Key Takeaways

• AI code is no worse to maintain than traditional, sometimes better (habitual users).
• Phase 1 speedup: 30–55%, phase 2: ~13%.
• CodeHealth correlates with business metrics (2× iterations, 15× fewer defects).
• Risks: bloat and cognitive debt need discipline.
• AI boosts strong teams (DORA).

— Editorial Team