Since kernel 7.0 the rule has been that patches developed with LLM help carry an Assisted-by tag in the changelog. Kernel 7.3 is the record release — 18,533 non-merge changesets through -rc7 — and 1,884 of them, about ten percent, wear the tag. At the Maintainers Summit in Prague last week, Dave Airlie put actual LLM use among contributors somewhere near ninety percent. The rule ran the whole cycle. Those two numbers don’t even share units — one counts changesets, the other counts people — but the room read the gap as a confession problem: Torvalds figured many submitters were simply ashamed to admit using the tools, and Airlie suspected fear of an anti-LLM witch hunt.

Jonathan Corbet’s writeup — the only public account of an invitation-only session, and his interpretation rather than a communique — records the room agreeing the tag had outlived any purpose it may have had and that the requirement could go. Corbet had framed the choice as enforce-or-drop; nobody in the room wanted the enforcement half. Contributors skip the tag out of ignorance of the rule or out of active hiding, maintainers mostly don’t ask, and so, in his words, the tag “provides no useful signal.” Torvalds said he would be happy to drop it, figuring the requirement itself scared people off the tools; Christian Brauner called it pointless; Shuah Khan alone still found it useful. But the same room did not stop wanting to know how patches are made. Which models, where in the process, whether a human is involved — that information, they agreed, belongs in the cover letter, where a maintainer already reads. And separately, the general agreement stands that even when the code is written by a model, the changelog should be written by the human behind the work.

The sharpest line out of the session was a joke. Thomas Gleixner suggested, “not entirely seriously,” a Not-assisted-by tag instead, and Arnd Bergmann answered that he would find that a useful indicator of developers he should spend time helping. The joke is the honest part: when ninety percent of your contributors use LLMs, “I used an LLM” carries no information and “a human did this” does — but only to Bergmann, as a hint about where to spend scarce mentoring attention, not as a fact anyone can check. A flipped tag is as enforceable as the original, and now there is a reason to lie on it. What transfers is the other half: stop asking for a mark on the work and ask for prose from the person. Last week QGIS banned the content it could not afford to review; yesterday openai/math locked the conversation on pull requests it still merges. The summit’s answer, in the only account there is, was neither: drop the count, ask for the story, where maintainers already read. That ask is voluntary too; the bet is that prose gives more away than a tag does, and the tells are already there: a changelog suspiciously verbose, a reply to one comment that comes back LLM-shaped.