The slides
Presentation — 9 slides
Self-contained Reveal.js deck that runs in any browser, also offline. Navigate with the arrow keys
or a presenter remote; press S for speaker notes, F for fullscreen.
The talk
A 5,155-line AI-generated patchset (six SRv6 Mobile User Plane behaviors, RFC 9433) reached the netdev list with green CI, textbook commit messages and 2,400 lines of passing selftests. One month of expert review later, nothing was mergeable. We analyzed the full thread — and re-verified, against kernel source, every finding of the AI reviewer that looked at the same series.
- The asymmetry, measured. ~1 day per generated iteration vs one month of expert review for 5 of 7 patches — ending in "start over".
- Plausible-but-wrong. Idiomatic code, precise RFC citations — implementing neither of the two RFC sections it cites; tests that pass by arithmetic coincidence.
- Human vs AI review. The AI reviewer reproduced ~22% of the human review, missed every outcome-deciding class, and its unique Critical was a false positive built on an invented kernel mechanism.
- Two proposals. Feedback loops + maintainer-owned per-subsystem prompts for AI review; and a desk-rejection threshold based on quality, not provenance.
Materials & references
- Case-study repository. Full thread analysis, findings catalog and verification notes: netgroup/srv6-ai-patchset-case-study.
- The patchset thread. [PATCH v2 0/7] seg6: add SRv6 Mobile User Plane (RFC 9433) behaviors on lore.kernel.org.
- AI review of the same series. Sashiko dashboard.
- BoF session. New Age Tooling BoF @ Netdev 0x1A.
- RFC 9433 (SRv6 Mobile User Plane) · RFC 8986 (SRv6 Network Programming) · RFC 8754 (SRH)
Notes
We are not here to blame anyone. We are here to learn and to improve the kernel development process, taking into account both the opportunities and the risks of AI.
Special thanks to Andrea Mayer, SRv6 maintainer, who performed the manual review at the heart of this case study.
Contact: stefano.salsano@uniroma2.it