Proposal: <podcast:aiTransparency>

[nathangathright]

Background

There have been several discussions about disclosing AI-generated content in podcasts (#531, #663, #669, #731, #735), but none have reached consensus. Meanwhile, Apple Music has adopted an AI transparency spec. (hat tip to: @kookster) Their <ai_transparencies> tag takes a category-based approach that balances meaningful disclosure with practical adoptability. This proposal adapts that model for podcasting.

Design Principles

1. Disclose the what, not the how. We declare which content categories involved material AI generation, not which tools were used or whether it was "assistance" vs. "generation."
2. Material portion threshold. Following Apple's language, disclosure is triggered when AI is used to generate "a material portion" of the content. Incidental use (spell-check, an ML-powered noise gate) does not trigger disclosure. Material generation of listener-facing content does.
3. Silence is silence, not denial. The absence of any aiTransparency tag means the podcaster has not made a declaration either way. This is distinct from actively declaring that no AI was used. Feeds without the tag are simply undeclared.
4. Let humans declare themselves human. Podcasters who use no AI should be able to say so positively, not just by omission. This is the tag's primary adoption driver: it gives every podcaster a reason to opt in.
5. Simple enough to actually adopt. A small, fixed vocabulary that a hosting platform can expose as checkboxes.

Three States, Not Two

This proposal departs from Apple's "if omitted, none is assumed" convention. In podcasting, we think a three-state model is more accurate:

1. Declared human: The podcaster actively asserts no material AI use. Apps can surface this as a badge or filter.
2. Declared AI use: The podcaster discloses specific categories of AI use.
3. Undeclared: No tag present is the default state for all existing feeds and should not be treated as equivalent to either of the above.

This three-state model is also the primary adoption driver. Podcasters who use no AI have a positive reason to add the tag. Apps that want to surface "certified human" badges have a signal to rely on. And the ecosystem avoids willfully conflating absence with nothing to disclose.

Why This Approach

Previous proposals in this repo have explored a range of designs. A simple boolean flag (#531) can't distinguish between "AI read the whole episode" and "AI made the cover art." Disclosing specific tools (#663) creates maintenance burden and ambiguity about what counts. Bundling AI with affiliate and sponsorship disclosures (#669) conflates different concerns. Extending podcast:person with a synthetic attribute (#731) only covers voice. And enumerating 15+ granular aspects with a four-state value system (#735) creates combinatorial complexity that hosting UIs won't expose.

Specification

Tag: <podcast:aiTransparency>

Discloses that AI was used to generate a material portion of a specific content category.

Parent: <channel> or <item>

Count: Multiple allowed (one per disclosed category)

Attributes:

Attribute	Required	Description
category	yes	The content category where AI was materially used.

Category Values

Value	Meaning
none	No AI was used to generate a material portion of any content. This is an active declaration, distinct from simply omitting the tag.
voice	AI generated a material portion of the spoken audio (text-to-speech, voice cloning, synthetic narration).
script	AI generated a material portion of the written content (scripts, show notes, outlines that directly became spoken content).
audio	AI generated a material portion of non-voice audio (music, sound design, mixing/mastering that materially altered the sound).
artwork	AI generated a material portion of visual elements (cover art, episode images, video).
transcript	AI generated a material portion of the transcript.
translation	AI generated a material portion of translated content (audio, transcripts, or metadata).

When category="none" is present, no other aiTransparency tags should be present on the same element. If both none and a specific category appear, the specific category takes precedence and none should be ignored by parsers.

Scope

When used at <channel> level, the tag declares the default AI transparency for the podcast. When used at <item> level, it replaces the channel-level declaration entirely for that episode. If an episode specifies its own aiTransparency tags, the channel-level tags do not carry over. An episode that wants to inherit most of the channel-level disclosures but add one more must redeclare all of them. The tradeoff gives podcasters precise per-episode control without ambiguity about how channel and item tags combine.

Examples

A human-made podcast with AI transcripts:

<channel>
  <title>My Human Podcast</title>
  <podcast:aiTransparency category="transcript" />
</channel>

A proudly human-made podcast, no AI:

<channel>
  <title>Handcrafted Audio</title>
  <podcast:aiTransparency category="none" />
</channel>

A fully AI-generated podcast:

<item>
  <title>Episode 42: The Future of Synthetic Media</title>
  <podcast:aiTransparency category="voice" />
  <podcast:aiTransparency category="script" />
  <podcast:aiTransparency category="audio" />
  <podcast:aiTransparency category="artwork" />
  <podcast:aiTransparency category="transcript" />
</item>

A podcast that normally uses AI transcripts, but one episode also uses AI artwork (note: channel-level transcript must be redeclared):

<channel>
  <title>My Human Podcast</title>
  <podcast:aiTransparency category="transcript" />
</channel>

<item>
  <title>Episode 7: Special Edition</title>
  <podcast:aiTransparency category="transcript" />
  <podcast:aiTransparency category="artwork" />
</item>

An episode that overrides a channel-level AI declaration to declare itself fully human:

<channel>
  <title>Usually AI-Assisted Show</title>
  <podcast:aiTransparency category="script" />
  <podcast:aiTransparency category="transcript" />
</channel>

<item>
  <title>Episode 12: Live Recording</title>
  <podcast:aiTransparency category="none" />
</item>

[Marzal]

a material portion it's a bit diffuse/fuzzy, but no more than the actual txt implementation. 👍

I like the Declared human option.

I don't get this part Item-level tags are additive to channel-level tags.
Shouldn't it work as the person tag?. Channel = default but item overwrites.

One episode with AI artwork, everything else human example seems to showcase that, but maybe I'm not familiar enough with the term additive (Spanish native speaker 👋🇪🇸)

PD: I would mind using this schema on the #669 proposal.
Or implement this one and also the new person attribute #731

[nathangathright]

Good point about additive. I was optimizing for the wrong use-case (podcasters wanting to add a single category without re-declaring everything), but it's clearer for the apps if it matches podcast:person. Updated!

[albertobeta]

@nathangathright @Marzal I think we are overcomplicating this.
I just launched shouldidisclose.ai. The rationale and reference framework are here: shouldidisclose.ai/about.html.
It might not be perfect, but I think it is a solid start. I call it the "Substance Test" because it focuses on substance, not performative disclosure. It is based on research into the disclosure requirements and guidance from major platforms and regulatory bodies. It is designed to be actionable and to cover most real-world cases, including the difference between cloning your own voice with AI versus using a third-party AI voice.

[Marzal]

• I will love to have this info Generative AI Disclosures #663 / Proposal: <podcast:ai> (AI Generation and Assistance Disclosure) #735 -> But I agree that it's too much and probably will never be implemented.

• I think that in the future this proposal could come in handy Proposal: <podcast:disclosure> for affiliates, AI, sponsorships, etc. #669 -> But no hosting seems OK to support it.

• Proposal: `<podcast:aiTransparency>` #753 Seems like a good compromise -> But if hosting prefer to start easy and slow with the txt tag, I'm ok with that (in fact I have already implement it)

  • If txt route gets adoption and all other proposal are discarded, I would push again for at least exentending the person tag Enhancement Proposal: Disclosing synthetic voices with podcast:person #731 . It's easy and it would complement with relevant info (now it's hard to tell some times, in the future could it be impossible).

Thanks @albertobeta for igniting the flame. Because what I think it's more important it's to have something even if imperfect rather that a great solution without implementation.

I would just show this @nathangathright's proposal to the PSP group, and if they don't see it as the way forward (meaning immediate implementation), "a tope" with <podcast:txt purpose="ai-content">

[matthewruzzi]

I think that it would be better to have a separate yes/no for each category.

[samsethi]

@albertobeta suggestion IMHO is a good first step to test the waters.

<podcast:txt purpose=ai-content>truepodcast:txt

RSS.com (host), TrueFans (host &app), Podnews (webpage) and David Marzal have already implemented this.

Would a good compromise be

<podcast:txt purpose=ai-transparency>truepodcast:txt

And if/when Apple podcasts adds AI transparency to their namespace we can adopt it too.

What I like about the use of TXT tag - Apple has already adopted this tag.