Suppose you wanted to publish some LLM generated text on your website.
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent
I voted for the Schema Option, but i would use Dublin Core, especially the Creator Element https://www.dublincore.org/specifications/dublin-core/dcmi-terms/terms/creator/ , as Metadata Schema. -
@Edent Currently I don't mark it out with metadata (instead I say "this is what the LLM said" or something along those lines), but that's a cool idea!
Maybe the
authoroption is closest to what I want but I feel wary of marking ChatGPT as an author, it feels like it humanises it too much. Possiblyitermprop="source" itemid="ChatGPT"if that isn't already taken?Honestly not that familiar with the various attributes (I have been guilty of the Stackoverflow-copy-and-pasta with most of my webdev) so I'd probably want to understand that a bit more in depth before choosing an option
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent I would include a humans.txt with the text to "sod off if you are a LLM". Adjacent to a robots.txt with the content: "User-agent: *
Disallow: /" -
@Edent I would include a humans.txt with the text to "sod off if you are a LLM". Adjacent to a robots.txt with the content: "User-agent: *
Disallow: /"@ivolimmen you've misread the post.
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent I like the Schema approach, but it falls flat when the AI has a human name (I'm feeling for all the breathing writers called Claude)… but that plus the language attribute?
Even this feels incomplete tho - as 'English not first language, used LLM to grammar-check' is v different from 'I didn't write this'. There's also a taxonomy somewhere to help make this clear (forget where). E.g.
- LLM written (entirely)
- LLM translated
- LLM assisted (contains LLM phrases / ideas).
- LLM edited/proof-read/spell/grammar checked.Distinguishing might make it easier to separate the kind of LLM-use Cory Doctorow says he does after writing, vs full slop-scribing.
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent@mastodon.social I like both the schema and language attribute options -
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent To me it’d feel most idiomatic to use a <blockquote>! lang=“ai” would make it impossible to say which language the AI has responded in, right?
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent I'd say using the lang attribute for that is the wrong move, as it has existing uses related to screen-readers. Different languages have different pronunciations for syllables/words that are written the same, and the screen-readers take that into account.
I imagine it might also mess up in-line automated translation tools
Also itemId is supposed to always be a URLGuidance for custom attributes is to use data-*, so I'd say like <p data-author="ChatGPT"> is best
https://developer.mozilla.org/en-US/docs/Web/HTML/Reference/Global_attributes/data-*/ -
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent I'd use `<blockquote>` with `<cite>` element or `<q>` for inline citation. For machine readable metadata I'd probably use `Quotation` type from Schema.org with `spokenByCharacter`, although I don't like classifying an LLM output as a creative work
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
A dedicated <meta> tag or a custom attribute could be used to prevent repurposing existing HTML semantics like lang or output. This avoids ambiguity (e.g., lang="ai" isn’t valid) and doesn’t misrepresent the AI as an "author" (since it’s a tool, not a creator with intent).
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent I wouldn't have thought to add metadata to identify it separately from my own work.
-
@Edent
I voted for the Schema Option, but i would use Dublin Core, especially the Creator Element https://www.dublincore.org/specifications/dublin-core/dcmi-terms/terms/creator/ , as Metadata Schema.i chose other because:
1. please, do NOT set “ai” as a language. it isn’t.
2. stochastic parrots are taking someone’s work and regurgitating it as their own, so there are two possibilities:
provenance
https://www.dublincore.org/specifications/dublin-core/dcmi-terms/#provenanceor
source
https://www.dublincore.org/specifications/dublin-core/dcmi-terms/#source3. BUT! the problem with the stochastic parrots is that they are copying without attribution. so, from the two choices we have, i think "provenance" is the more accurate one.
4. AI ≠ human ∴ ≠ author
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent No opinions other than a strong one that it should not be the “lang="ai"” version. “ai” isn't an appropriate ISO 639 language code which is what a lang attribute with this syntax should be.
-
i chose other because:
1. please, do NOT set “ai” as a language. it isn’t.
2. stochastic parrots are taking someone’s work and regurgitating it as their own, so there are two possibilities:
provenance
https://www.dublincore.org/specifications/dublin-core/dcmi-terms/#provenanceor
source
https://www.dublincore.org/specifications/dublin-core/dcmi-terms/#source3. BUT! the problem with the stochastic parrots is that they are copying without attribution. so, from the two choices we have, i think "provenance" is the more accurate one.
4. AI ≠ human ∴ ≠ author
BTW given the numerous reports of the proprietary LLMs changing answers, i would add a DATE property and timestamp it down to the minute.
-
A dedicated <meta> tag or a custom attribute could be used to prevent repurposing existing HTML semantics like lang or output. This avoids ambiguity (e.g., lang="ai" isn’t valid) and doesn’t misrepresent the AI as an "author" (since it’s a tool, not a creator with intent).
the LLM isnt a language. it’s taking in info from elsewhere and giving it to you. in that sense, it’s a provenance if the LLM isn’t giving the source of their answer.
and FWIW: in USA only humans are recognized as authors. you do NOT want to take the techbros bait of claiming their proprietary software is the real author of your work.
imagine Microsoft claiming copyright over your labor because you used MSOffice or Github?
that’s why metadata is so important.
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
lang attribute is for human languages. ai is not a language. Also it just looks like the common mistake of putting a country code there (Anguilla)
output is for interactive elements like a calculator.
schema + itemprop: maybe! but I don’t see a clear standard yet. IPTC has metadata for gen AI images, not text
If entire doc is gen AI: JSON-LD for entire doc can indicate author, and even entity type = Person, so in theory something else can go there?
Why do you feel you need this?
-
lang attribute is for human languages. ai is not a language. Also it just looks like the common mistake of putting a country code there (Anguilla)
output is for interactive elements like a calculator.
schema + itemprop: maybe! but I don’t see a clear standard yet. IPTC has metadata for gen AI images, not text
If entire doc is gen AI: JSON-LD for entire doc can indicate author, and even entity type = Person, so in theory something else can go there?
Why do you feel you need this?
@Edent To clarify my question: we don’t generally use HTML to indicate authorship of a textual element on the page, whether human or machine. There are whole document standards. What purpose does it serve to call this out?
If you are quoting the LLM, there is a microformat for Quotation. Unfortunately the only valid authors are of type “Person” or “Organization”. I guess you could hack it and put another value there, no one will send you to jail
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent I think the language attribute can’t be right: it’s not a language. And LLM output could be in any language, so that would prevent the ability to correctly indicate the actual language.
I would also disagree with the schema option (3); an LLM is not an author but a tool used by some other author.
Out of these three then I’d choose option 2, output element; but I don’t have a strong opinion in its favour or whether there’s a better alternative, more just a process of elimination.
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
@Edent I thought of a custom html tag that would prevent copying (lol) like: <ai-generated model="model name" date="date here" prompt="prompt to generate text">text here</ai-generated>
We could implement this now in validating html using the custom HTML elements api
-
Suppose you wanted to publish some LLM generated text on your website. (Perhaps to critique it, or praise it, whatever.)
What metadata would you use to indicate that it was AI generated?
Which of these examples is closest to your ideal?
(No need to reply telling me your insights on AI, thanks. I'm just interested in how you think it should be rendered in HTML.)
