# Echoes of AI: Investigating the Downstream Effects of AI Assistants on Software Maintainability

**Authors:** Markus Borg, Dave Hewett, Nadim Hagatulah, Noric Couderc, Emma Söderberg, Donald Graham, Uttam Kini, Dave Farley  
**arXiv:** [2507.00788](https://arxiv.org/abs/2507.00788)  
**Date:** 2025-07-01  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2507.00788v2/

## Key points

- Study by Markus Borg et al. examines impact of AI assistants like GitHub Copilot and Cursor on software engineering
- Research aims to determine if co-development with AI assistants affects software maintainability and ease of evolving resulting source code
- Two-phase controlled experiment conducted with 151 participants, 95% professional developers
- Phase 1: Participants added new feature to Java web app with or without AI assistance
- Phase 2: Randomized controlled trial where new participants evolved solutions without AI assistance
- Results from Phase 2 showed no significant differences in subsequent evolution regarding completion time or code quality
- Bayesian analysis suggested minimal and highly uncertain improvements in speed or quality when using AI assistants
- Observational findings from Phase 1 indicated that utilizing an AI assistant reduced completion time by a median of 30.7%, habitual users experienced estimated speedup of 55.9%
- No systematic advantages or disadvantages in maintainability found when other developers evolved code co-developed with AI assistants
- No consistent warning signs indicating degraded code-level maintainability within the scope of experiment's tasks and measures used
- Future work should focus on examining potential risks such as code bloat resulting from excessive code generation and cognitive debt as developers rely more on AI assistants for mental effort offloading

## Summary

The study "Echoes of AI: Investigating the Downstream Effects of AI Assistants on Software Maintainability" by Markus Borg et al. examines the impact of AI assistants like GitHub Copilot and Cursor on software engineering. The research aims to determine if co-development with AI assistants affects software maintainability and how easily other developers can evolve the resulting source code. A two-phase controlled experiment was conducted with 151 participants, 95% being professional developers. In Phase 1, participants added a new feature to a Java web application with or without AI assistance. In Phase 2, a randomized controlled trial was performed where new participants evolved these solutions without AI assistance. Results from Phase 2 showed no significant differences in subsequent evolution regarding completion time or code quality. Bayesian analysis suggested minimal and highly uncertain improvements in speed or quality when using AI assistants. Observational findings from Phase 1 supported previous research showing that utilizing an AI assistant reduced completion time by a median of 30.7%, with habitual users experiencing an estimated speedup of 55.9%. Overall, the study found no systematic advantages or disadvantages in maintainability when other developers evolved code co-developed with AI assistants. Within the scope of the experiment's tasks and measures used, there were no consistent warning signs indicating degraded code-level maintainability. The researchers suggest future work should focus on examining potential risks such as code bloat resulting from excessive code generation and cognitive debt as developers rely more on AI assistants for mental effort offloading.