Experiment tracking: Posting with AI assistance #685

Documenting here for posterity that we're experimenting with how to present AI-written and AI-cowritten content in H2.

Content increasingly falls into three overlapping buckets:

1. Written solely by AI (eg automatic meeting notes)
2. Written mostly by AI, with human review and publishing
3. Written mostly by a human, with AI used to assist

(This partially overlaps with the EU's labelling, which distinguishes 1 as "AI generated" and 2+3 as "AI modified")

With this being a relatively new change to communication, we've considered different ways to make this possible and present it:

• Creating a user directly for each agent to publish as (for 1)
• Allowing agents to post as users (for 1 and 2)
• Implementing co-authoring for posts to indicate a collaboration with AI (for 2 and 3) - Display ordered post authors #649

While we're iterating on this, we've decided to go with a lightweight implementation currently: agents can access H2 sites via the same access users would have, and no coauthoring notes are provided. However, we're not implementing fully-autonomous posting (ie 1 is not possible), and we're encouraging (but not enforcing) human review (for 2).

Part of this is a bet on AI assistance being mostly normalised, such that 2 + 3 "merge" effectively into the same process.

This also means that in cases where 1 might be useful (such as automatic posting of call notes), we're instead encouraging tools that provide a human-in-the-loop review - that could be things like an agent only creating drafts rather than automatically publishing, eg.

We'll continue tracking this to see if the impact of "slop grenades" is too high, and whether further indicators or speed humps to AI posting end up necessary.