Janice Wilson and Jeff Auriemma published RFC 454545 on GitHub Gist in March 2026. It runs to several pages, follows IETF conventions to the letter, and proposes new Unicode code points to address a problem that has quietly become a genuine cultural flashpoint: LLMs use em dashes constantly, with what the document calls 'suspicious regularity' and 'unwavering grammatical confidence,' and it is now difficult to look at one without wondering.
The RFC's fix is two new code points. The Human Em Dash (HED, U+10EAD) and Human Attestation Mark (HAM, U+10EAC) would be visually identical to the standard U+2014 but encoded separately to carry a claim of human origin. To emit a certified HED, a conforming implementation must verify behavioral evidence: hesitation pauses exceeding 137 milliseconds, backspace events, cursor repositioning. The Human Cognitive Proof-of-Work (HCPoW) section extends this to 'incongruous emoji usage' and 'expression of personal values or accountability.' The Security Considerations section flags 'excessively consistent hesitation intervals' and 'uncanny servility' as LLM tells worth monitoring.
A note for Agent Wars readers: this article has already used an em dash. Twice, if you count the headline. This is precisely the situation RFC 454545 is documenting, and the RFC names it Dash Authenticity Collapse (DAC). Whether any given paragraph was written by a human or not, you cannot tell from the punctuation. That is the actual problem, and Wilson and Auriemma articulate it more usefully than most earnest treatments of AI content detection manage to.
Commenters on the Gist identified the obvious loophole within hours: instruct your AI to replace all em dashes with the HAM+HED sequence and the standard is broken before it is adopted. Point 4.3's behavioral verification requirements are the only barrier, and they are not much of one. Wilson and Auriemma almost certainly knew this. The loophole is the point.
The RFC format is well-chosen for a reason beyond parody. IETF documents operate on the assumption that bad actors are adversarial but technically constrained. AI authorship attribution has no such assumption available, and the Human Punctuation Registry requested in the IANA Considerations section would be obsolete before the ink dried. The document is not really about em dashes. It is about the impossibility of behavioral fingerprinting in a world where any behavioral signal can be synthesized on demand.