How and why we use AI in these docs

Highlighting AI-written text

Knowing how content was made helps you assess what you’re reading. The Highlight AI edits toggle reveals the characters that were recorded as AI-written. Turning it on changes how the content is displayed, but does not change the content itself.

What the highlights mean

Highlighted text was recorded as being AI-generated text. This happens during the development/writing process and is integrated directly into the dev tools we use. It's at an early stage and is not a perfect process. Content can still be highlighted after it's been reviewed by a person, so it doesn't mean that content is bad or low quality; just that those characters originated from AI output.
Text that isn't highlighted doesn't mean it was written by a person. Older content was not labelled retroactively, so a page with no highlights can still contain earlier AI work.
The marks do not indicate review or accuracy. They record how text was written, not whether someone has checked it or whether it is correct.

Important: these highlights show text labelled as AI-written during writing and editing. They do not automatically detect who or what wrote it, or where a thought or idea originated.

Similar efforts in the industry

Making the creation process more transparent is also a goal of industry efforts such as C2PA Content Credentials, which uses digitally signed records to document content’s origin and editing history.

Our highlighting comes from TextProv, an approach we created to help developers and writers keep track of their own AI-assisted content. It embeds markers directly in the text, without maintaining any kind of history.

Your display preference

The toggle remembers your display preference in this browser when local storage is available. You can leave highlighting off if it is not useful to you.

Why I use AI tools and label its writing

I use AI tools to help publish documentation more often. I'm natrally a slow writer and without the help the content gets outdated quickly. In some cases new functionality was going undocumented for months (in some cases not at all). I would like to improve my writing speed so that's something I'm concious of as well but in the meantime I'm experimenting with this approach. I'm figuring it out as I go along.

Making AI content visible

I consider publishing AI-written prose without disclosing it to be dishonest, in a way that's not unlike serving soup made with beef stock to a vegetarian or farting in the elevetor and denying it. No one is getting hurt, but it creates an unpleasant situation for everyone involved.

To be clear: it is not a judgement on how others cohose to work; it's a personal belief that helps me sleep at night. Keeping track of how content was produced is a way to be honest with readers and myself.

Reviewing after publication

I draft with AI, review what I can at the time, and publish with its origin tracked. As a separate pass, I review more closely after the fact. That means some documentation is available before I have checked properly. Review means checking the content and deciding whether to stand behind it, even when I keep the wording unchanged. The origin marks remain after that review; they are not a record of which pages I have checked.

Reports of ~errors, confusing explanations, or poor writing~ any and all crap are appreciated, even more so when it's something I've written or considered properly reviewed. That feedback does not replace intentional review but it fits in with the ethos of opensource software in general. We're all in this together.

Radical transparency

This transcript but is included here to demostrate how it's not so straightforward to determine what is AI-written and what is not. Things I write can be influenced or direct copies of AI output (e.g. when I prefer to suggested phrasing to my own). Likewise, AI output can be influenced by my own writing style and phrasing.

NOTE: This conversation is not highlighted like the site content. It's just a raw transcript created from Zed Agent's "Open Thread as Markdown", parsed by bin/zed-thread-clean.py, and saved as an .mdoc file so it can be included as a partial in this page.

Read one of the conversation behind this page

User

@text-provenance.md (34:39) How to improve the visual display using existing Astro configuration and functionality (i.e. without adding a new component).

Image

Assistant

I’ll check the existing Starlight callout syntax and styles for a way to render this note without a new component.

Use Starlight’s built-in note aside, rather than GitHub’s > [!NOTE] syntax. The screenshot shows the GitHub marker rendered as ordinary blockquote text.

Replace lines 36–38 in src/content/docs/en/contribute/text-provenance.md with:

:::note
I'm switching to first-person "I" instead of "we" here to be direct about
whose thoughts and opinions we're talking about.
:::

This provides a distinct callout with a note icon, title, tinted background, and light/dark theme styling. It separates the author’s aside from the main argument without additional CSS, imports, components, or configuration changes.

The site already enables Starlight in config/integrations.mjs, and other English pages use this syntax. No files changed; rendering was not tested.

User

This header needs to be updated

### Origin and review

TextProv records declared origin; it does not detect authorship.

- **Highlighted text was recorded as originating from AI output.** This happens during the
  development process and is integrated directly into the dev tools we use. It's at an early
  experimental stage and is not a perfect process. Content can still be highlighted after
  it's been reviewed by a person, so it doesn't mean that content is bad or low quality;
  just that those characters originated from AI output.
- **Text that isn't highlighted doesn't mean it was written by a person.** Older content
  was not labelled retroactively, so a page with no highlights can still contain earlier
  AI work.
- **The marks do not indicate review or accuracy.** They record how text was
  written, not whether someone has checked it or whether it is correct.

Assistant

I’ll change the heading to “What the highlights mean” and leave the explanation unchanged.

The page has existing edits, which I’ll preserve. I’ll use the repository’s TextProv tool to change only the heading, then run the required checks.

The heading is updated. Tests, frontmatter, navigation, and the build passed. The combined TextProv check reported mark-preservation errors; I’ll check whether those come from the pre-existing edits.

Updated src/content/docs/en/contribute/text-provenance.md:14:

### What the highlights mean

Body text and existing edits are unchanged.

Validation: 312 Vitest tests, 66 Python tests, frontmatter, navigation, and production build passed. The heading-only TextProv check passed; the combined diff retains 26 pre-existing mark-preservation errors.

User

With that change, how to differentiatew from the section header itself?

## What the highlighting tells you

The **Highlight AI edits** toggle reveals the characters that were recorded as AI-written.
Turning it on changes how the content is displayed, but does not change the content itself.

### What the highlights mean

Assistant

Make the parent heading name the feature, and the subsection explain how to interpret it:

## AI highlighting

[Existing introduction about the toggle.]

### What the highlights mean

[Existing explanation and bullets.]

### Your display preference

This separates the feature, its meaning and limitations, and the reader’s settings. Only the H2 needs changing.

User

How to rephrase without using the technical term TextProv which isnt relevant yet? "TextProv records declared origin; it does not detect authorship."

Assistant

“The highlights show text labelled as AI-written during writing and editing. They do not automatically detect who or what wrote it.”

User

Great. How to followe that with a brief sentence about the industry CPST or whatever it's called.

Assistant

You’re likely thinking of C2PA, the standard behind Content Credentials. Follow with:

This follows the same broad principle as Content Credentials: recording how content was created and edited, rather than guessing its origin.

This draws a comparison without implying the site implements C2PA.

User

Is Content CRedentials a portion of C2PA?

Assistant

Not exactly a portion—Content Credentials is the public-facing name for provenance records based on the C2PA standard.

C2PA is the Coalition for Content Provenance and Authenticity, which develops the technical standard.
Content Credentials are digitally signed records describing content’s origin and editing history, using that standard.

For your sentence, “the C2PA standard behind Content Credentials” is accurate. Your highlighting shares the transparency goal but should not be described as Content Credentials unless it implements that standard.

User

Ah, thanks for the clarification. TextProv is not trying to replace C2PA or be an alternative. It's more like a developer's tool for keeping tracking themsweves. How to reframe that one sentence?

Assistant

“Content Credentials provide a broader standard for content history; our highlighting is a developer tool for tracking and sharing which text came from AI during editing.”

User

We can;t assume the reader of this page knows anything about C2PA or even why anyone cares

Assistant

“Knowing how content was made helps readers assess what they’re reading—a goal behind industry efforts such as Content Credentials, which records content’s origin and editing history. Our tool focuses on helping developers track and disclose their own AI-assisted edits.”

User

IS it along the lines of DMCA?

Assistant

No. They address different concerns:

DMCA is US copyright law, including rules for takedown notices and circumventing digital access controls.
C2PA / Content Credentials records content’s origin and editing history. It does not determine copyright ownership or grant permission to reuse content.
TextProv helps developers track and disclose AI-written text during editing.

The closer analogy is a source label, not a copyright notice.

User

Thanks. How's this?

## Highlighting AI

Knowing how content was made helps you assess what you’re reading. The **Highlight AI edits** toggle
reveals the characters that were recorded as AI-written. Turning it on changes how the content is
displayed, but does not change the content itself.

### What the highlights mean

- **Highlighted text was recorded as being AI-generated output.** This happens during the
  development/writing process and is integrated directly into the dev tools we use. It's at an early
  stage and is not a perfect process. Content can still be highlighted after it's been
  reviewed by a person, so it doesn't mean that content is bad or low quality; just that
  those characters originated from AI output.
- **Text that isn't highlighted doesn't mean it was written by a person.** Older content
  was not labelled retroactively, so a page with no highlights can still contain earlier
  AI work.
- **The marks do not indicate review or accuracy.** They record how text was
  written, not whether someone has checked it or whether it is correct.

Important: these highlights show text labelled as AI-written during the writing and editing process. They
do not automatically detect who or what wrote it or where the thought or idea originated. That is the
goal behind industry efforts such as [C2PA Content Credentials](https://c2pa.org/), which records
content’s origin and complete editing history.

What we're doing is just a fraction of that as a way to help developers and writers track their own
AI-assisted content.

Assistant

The opening and heading hierarchy work. The final paragraphs need one factual correction: Content Credentials does not automatically detect authorship or where ideas originated, and it does not guarantee a complete editing history. “That is the goal” currently implies otherwise.

Replace the closing paragraphs with:

These highlights reflect labels added during writing and editing. They do not automatically detect who or what wrote the text, or where an idea originated.
Making content’s history visible is also the aim of industry efforts such as Content Credentials, which uses digitally signed records to document its origin and edits. Our highlighting is a developer tool for tracking AI-assisted writing and sharing that information with readers—not an implementation or replacement of that standard.

Two smaller wording changes:

“Highlighting AI-written text” is more specific than “Highlighting AI.”
“AI-generated output” can be shortened to “AI-generated text.”

User

I don't think that differentiation with Content Credentials is clear enough. You make it sound like TextProv isn't that different when it's at least an order of magnitude less complex and only focused on the editing and presenting stage, without any notion of history.

Assistant

The distinction is labels on the current text versus a record of its history, not just a narrower version of the same system.

Industry efforts such as Content Credentials use digitally signed records to document content’s origin and editing history. Our tool does something much simpler: it marks AI-written text during editing and lets readers display those marks. It does not track previous versions, who changed what, or the sequence of edits.

User

"our tool" is weird here. We're not talking about our implementation here.

it marks AI-written text during editing and lets readers display those marks.

It's not about the readers. It's about the person doing the writing/editing and giving visibility to their AI tool augmented process. Choosing to display to readers is a separate concern that TextProv makes possible. it's asynchronous to even make a statement about that since we don't do that for C2PA either.

Assistant

“Content Credentials uses digitally signed records to document content’s origin and editing history. TextProv is a much simpler editing aid: it helps writers and editors track their own use of AI through labels embedded in the text, without maintaining an editing history.”

User

And how to incorporate that into our content which again, readers need to be introduced to.

Important: these highlights show text labelled as AI-written during the writing and editing process. They do not automatically detect who or what wrote it or where the thought or idea originated. That is the goal behind industry efforts such as C2PA Content Credentials, which records content’s origin and complete editing history.

What we're doing is just a fraction of that as a way to help developers and writers track their own AI-assisted content.

Assistant

Important: these highlights show text labelled as AI-written during writing and editing. They do not automatically detect who or what wrote it, or where a thought or idea originated.

Making the creation process more transparent is also a goal of industry efforts such as Content Credentials, which uses digitally signed records to document content’s origin and editing history.

The labels here come from TextProv, a much simpler tool that helps writers and editors keep track of their own AI-assisted writing. It embeds labels in the text being edited; it does not maintain a history of previous versions, contributors, or edits.

User

Ansewr only. Why do you keep emphasizing "Content Credentials" without mentioning C2PA?

Assistant

I prioritized the public-facing name to avoid introducing another acronym, but omitted the relationship you wanted to explain. “Content Credentials, based on the C2PA standard” would name both clearly.

User

So then why not "C2PA Content Credentials"

Assistant

“C2PA Content Credentials” works here. I was overcorrecting for the distinction between the standard and its public-facing name; it didn’t warrant avoiding your wording.

User

Nice. Also how to be clear that we made TextProv?

Our highlighting comes from TextProv, a more basic approach that helps developers and writers keep track of their own AI-assisted content. It embeds markers in the text directly, without maintaining history of previous versions, contributors, or edits.

Assistant

Our highlighting comes from TextProv, a tool we created to help developers and writers keep track of their own AI-assisted content. It embeds markers directly in the text, without maintaining a history of previous versions, contributors, or edits.