Evidence Record

Evidence summary

Sixteen experienced open-source developers completed 246 real issues in repositories they knew well. AI-allowed tasks took 19% longer, despite developers believing AI sped them up.

Exact claim supported or contradicted

AI availability does not guarantee speedup; developer perception can diverge from observed task time.

Source provenance

Primary study report and linked paper from an independent nonprofit.

Relevant excerpt or data

Developers expected 24% speedup; observed 19% slowdown; after the study they still estimated 20% speedup.

Interpretation

Repository familiarity, task realism, and verification overhead plausibly moderate effects.

Limitations

Only 16 developers, selected open-source projects, early-2025 tools, self- reported time, and uncertain generalization to other experience levels.

Counterevidence

EV-AICODING-2026-10A1 found a large speedup on a bounded greenfield task.

Reproduction or verification notes

The authors report estimator robustness but acknowledge representativeness risk. Later recruitment developed treatment-avoidance selection bias.