'Vouch subject' #9

zesterer commented 2026年03月08日 23:15:39 +01:00

I love the idea of this project! A thought I had while reading this is that having 'human-authored' be a hard binary makes the trust model open to attack via reinterpretation.

To one person, 'human-authored' may mean 'each character typed by hand', but to another it might mean 'the original idea is human' or 'an LLM has been used for touch-up'. This opens the door to a game of telephone where multi-hop vouching does not necessarily imply that all links in the chain are strictly human-authored, even if the end user is looking for such a guarantee.

Instead, I'd propose the ability to vouch for different subjects, or along a gradient of humanness, then allowing users of the browser extension to select which level of humanness they're willing to place faith in.

I love the idea of this project! A thought I had while reading this is that having 'human-authored' be a hard binary makes the trust model open to attack via reinterpretation. To one person, 'human-authored' may mean 'each character typed by hand', but to another it might mean 'the original idea is human' or 'an LLM has been used for touch-up'. This opens the door to a game of telephone where multi-hop vouching does not necessarily imply that all links in the chain are strictly human-authored, even if the end user is looking for such a guarantee. Instead, I'd propose the ability to vouch for different subjects, or along a gradient of humanness, then allowing users of the browser extension to select which level of humanness they're willing to place faith in.

robida commented 2026年03月09日 02:35:57 +01:00

Thanks for the feedback and suggestion, @zesterer! I like the idea, I just wonder if it's possible to define a spectrum between "fully AI" and "fully human". Something like:

• Written by a human
• Used spellchecker
• Used grammar checker
• Used an AI proofreader
• ...
• Fully AI generated

The problem is that it's hard to define a 1D axis: if the idea was suggested by AI, where does that fit? If the images are generated by AI but not the text? Maybe it would be easier to just have 3 categories (just 2, really, since there's no point in declaring the 3rd):

• No AI
• AI-assisted
• Full AI

We probably want to have an array for self URLs then, something like where people can declare different content with different levels:

{
 "version": "0.1.2",
 "content": [
 { "url": "https://robida.net/posts", "level": "AI assisted" },
 { "url": "https://robida.net/photos", "level": "No AI" }
 },
 "vouches": [
 {
 "url": "https://rauh.net/",
 "vouched_at": "2026年03月08日",
 "note": "Personal friend"
 }
 ]
}

As you can see in the example above, I was also thinking about adding an optional note field, to explain why you're vouching for someone.

Thanks for the feedback and suggestion, @zesterer! I like the idea, I just wonder if it's possible to define a spectrum between "fully AI" and "fully human". Something like: - Written by a human - Used spellchecker - Used grammar checker - Used an AI proofreader - ... - Fully AI generated The problem is that it's hard to define a 1D axis: if the idea was suggested by AI, where does that fit? If the images are generated by AI but not the text? Maybe it would be easier to just have 3 categories (just 2, really, since there's no point in declaring the 3rd): - No AI - AI-assisted - Full AI We probably want to have an array for self URLs then, something like where people can declare different content with different levels: ```json { "version": "0.1.2", "content": [ { "url": "https://robida.net/posts", "level": "AI assisted" }, { "url": "https://robida.net/photos", "level": "No AI" } }, "vouches": [ { "url": "https://rauh.net/", "vouched_at": "2026年03月08日", "note": "Personal friend" } ] } ``` As you can see in the example above, I was also thinking about adding an optional `note` field, to explain why you're vouching for someone.

jentza commented 2026年03月10日 17:18:24 +01:00

I love the idea of this whole spec, as well as being able to tag the vouches. With the tags as Full AI, it might be worth adding a 4th tag - AI Generated, Reviewed/Edited by Human. That might fit into the AI-assisted category, but as you said above the spectrum would get massive.

Another move might be to split the level into 2 separate entities:
Content Generation=Full AI, Partial AI, No AI
Tooling, Editing=Full AI, Partial AI, No AI

May not be the case for everyone, I care a lot less about items that are edited/fixed by AI than I do about content that's generated by AI.

I love the idea of this whole spec, as well as being able to tag the vouches. With the tags as Full AI, it might be worth adding a 4th tag - AI Generated, Reviewed/Edited by Human. That might fit into the AI-assisted category, but as you said above the spectrum would get massive. Another move might be to split the level into 2 separate entities: Content Generation=Full AI, Partial AI, No AI Tooling, Editing=Full AI, Partial AI, No AI May not be the case for everyone, I care a lot less about items that are edited/fixed by AI than I do about content that's generated by AI.