Manifold recently added a feature where AI can edit my market description without my consent. Naturally these edits are terrible, frequently containing incorrect information and not understanding what the market is about at all.
This resolves YES if there's a way for me to turn this off, or it's removed entirely, or the AI gets smart enough that it stops making detrimental edits, or it otherwise stops annoying me.
Update 2025-02-19 (PST) (AI summary of creator comment): Arbitration Costs:
Paying for arbitration costs will not count as a valid solution to the issue.
1,000
An example of Manifold's AI making a blatantly wrong addition to my market description. (No, simply paying for arbitration costs would not count as a success.)
https://manifold.markets/IsaacKing/if-i-try-to-fight-paypal-for-my-350
I don't think we'll add a toggle to turn off the clarification, but I do think o1-o3, etc. will make them better. I also support adding a notification that the AI made an edit
@ian I was personally thrown off by this feature, and would prefer if it was a “suggested clarification” that the market creator can accept/reject. Even just a bot response like “@market_creator Do you want me to edit the description to add [edited text]? Thumbs up to accept.” would be great.
COI: I hold YES on this market.
@ian I suspect the inference costs on o3, even on low settings, could be prohibitive for this for the rest of 2025 - about an OOM more expensive than highest settings o1 on lowest settings o3 from what they've shown
@ian I work for OpenAI so I am obviously among these models’ biggest fans, but my two cents — I think it’s better to empower users, and it seems like this feature is maybe disempowering them at present.
Based on the example, I don’t think reasoning models will help here — the problem sounds like that the model both has incomplete information and has been shoehorned into believing its summary will be Relevant, when it was not.
And if it’s a summary of the market creator’s own comment, I think the market creator can be trusted to make the call. Empowering the human user to override the AI’s suggestion is almost always better for achieving good outcomes, even in non-existential situations like this one!
@No_uh The "comment summaries" are what I'm referring to, it will sometimes "summarize" my comment to say something totally different from what I actually said.