New blog post: Signaling AI Preferences on ATProto. Introducing community.lexicon.preference.ai, a lexicon for granular AI data usage preferences with scoped overrides for specific entities and collections. #AI #atproto

April 04, 2026 - 19:06 UTC

3
3
48
PR is live at github.com/lexicon-community/lexicon/pull/72
0
0
2
I think this is a well-designed proposal

however, I feel like there are pretty big issues with the 'inference' and 'embedding' categories

e.g. why would Bluesky PBC choose to honour an 'inference' 'deny' pref when that would interfere with well-established moderation workflows that use LLMs?
1
0
1
and if independent moderation services such as @labeler.hailey.at or @skywatch.blue chose to honour an 'inference' pref set to 'deny', wouldn't that reduce the effectiveness of those community labellers?

(AIUI both of those use LLMs in some situations as part of Osprey)
1
0
1
Gotta say, I'm not very familiar with ATProto's innards, but that PR made a lot of sense to me nonetheless. May I ask a quick question about process, if you don't mind? No worries if not.
0
0
0