AI Preferences (preference signals)
Machine-readable signals — an evolution of robots.txt — by which a website declares whether its
content may be used for AI training or deployment. The work of an IETF AI Preferences
working group, documented here from eff-web-under-attack-ietf.
What it is
A standard vocabulary and transport for “do / don’t use this for AI” expressed at the page or site
level. As a convention it is voluntary, the way robots.txt has always been. The governance
question is what happens once a venue blesses it: in some jurisdictions a standardized
preference signal could become legally enforceable, converting a request that crawlers were
free to ignore into an access right backed by law.
Why it sits in this spoke
It is governance of AI by controlling its inputs rather than its conduct — a lever aimed at the training-data supply chain, distinct from the conduct rules of the eu-ai-act or the content-output rules of China’s labeling regime. EFF reads it as a gate that, scoped too broadly, would also block non-AI crawling (archiving, research, accessibility); defenders read it as creators reclaiming consent over reuse. Both are recorded; the draft is unsettled.
Connections
- developed at ietf
- analyzed in eff-web-under-attack-ietf
- companion mechanism: web-bot-auth