Kernel maintainers require Assisted-by tags for AI rubber-duck reviews

A hardening series for the xillybus character driver has forced a clear answer from Linux kernel maintainers: if AI tools helped find the issues, the patches need Assisted-by tags, even when a human wrote every line of the fix.

Eli Billauer, the driver's author, posted seven patches adding mutex protection, defensive sanity checks on hardware- and user-derived values, unsigned jiffies arithmetic, and logging of unsolicited interrupts. Under normal conditions the driver should behave as before, aside from a possible small bandwidth gain from a corrected checkpoint comparison that had been promoting a signed leap value to unsigned and sending flow-control messages too often.

Billauer described the work as the product of a three-week "reverse rubber duck" session with Deepseek, Kimi Thinking K2.6, ChatGPT, and Claude Sonnet 4.6. The models proposed bugs; he spent most of the time disproving them. When they were right, their suggested patches were unusable, so he wrote the corrections himself. He had no per-patch record of which tool contributed what, and questioned whether Assisted-by was meant for this kind of use.

Jonathan Corbet asked whether the tags were required. Greg Kroah-Hartman answered directly: yes, add something. The exchange settles, at least for this tree, that the kernel's emerging AI-credit rule tracks assistance in the review process, not only machine-generated code.