Enhancement Proposal: Disclosing synthetic voices with podcast:person

[nathangathright]

After recent news, there’s renewed interest in a tag for AI disclosure. Incorporating the feedback from my previous exploration of the topic, I think the approach most likely to see significant adoption would be a new attribute on the podcast:person tag disclosing usage of a synthetic voice.

Goals

Since adoption and accuracy are voluntary, this effort is more aligned with enabling transparency than enforcement. This attribute cannot be accurate enough for advertisers to make informed decisions without perversely incentivizing less accuracy. Solving that problem is not within the scope of this proposal.

Examples

AI-Cloned Voices

Evan Ratliff uses an AI-cloned voice to host Shell Game.

  <podcast:person
      role="host"
      img="…"
+     synthetic="cloned"
  >Evan Ratliff</podcast:person>

AI-Generated Personas

The host of Bloom and Banter is an AI-generated character.

  <podcast:person
      role="host"
      img="…"
+     synthetic="generated"
  >Nigel Thistledown</podcast:person>

Suggested Usage

Clients may choose to disclose this information as they see fit. I recommend a tooltip alongside the other information from the podcast:person tag as shown above. Others might add a general disclosure statement à la YouTube.

Anticipated Concerns

Just repurpose the role attribute

The formalized spec only supports a single role per person from the list maintained by the Podcast Taxonomy Project. We could just allow multiple values and expand the list of valid values. TrueFans achieved this result a year ago without other clients following suit (to my knowledge). This proposal is attempting to find more momentum by approaching it differently.

Percentage Value

@jamescridland suggested an additional property for a percentage of AI content in an episode, enabling clients to filter "majority AI" content. I’m presuming this would be determined by estimating percentage of total speaking time. I believe this adds complexity without a clear benefit to the podcaster, thus harmful to achieving widespread adoption.

If a podcaster wouldn’t otherwise create a podcast:person tag for a voice in the podcast (like an announcer intro purchased from Fiverr), then that usage of AI probably falls below the threshold that most would consider it "majority AI" content. If podcaster would like more fine-grain disclosures beyond inclusion as a podcast:person, that would be beyond the scope of this proposal.

[picard102]

We could just allow multiple values and expand the list of valid values.

Fully support expanding the person tag to allow multiple roles. It's such low hanging fruit at this point that is clearly needed.

[nathangathright]

Yeah, I’m unclear if it was just too low-impact for hosts/apps to prioritize before now or if they have other concerns that have prevented them from copying TrueFans’ approach.

[picard102]

I have at least two clients asking for it right now.

[theDanielJLewis]

While I like the idea of putting a disclosure on <podcast:person>, that would cover only the voice, which I think is the least important disclosure.

Please take another look at my proposal for <podcast:disclosure>, which would be more inclusive as well as provide ways to disclose legally required things, such as sponsorships, affiliates, paid appearances, and such.

[nathangathright]

I think you meant to link to #669.

[theDanielJLewis]

Updated. Thanks.

[samsethi]

As Nathan said, TrueFans addressed this issue over a year ago by allowing for a secondary role in the person tag. We also extended the person taxonomy to include AI. This is useful when you have a human host(s) and you have an additional AI host.

However if every host is AI generated then you can simply tag the whole episode as AI.

So there are three things that need to happen.

1. The podcast taxonomy needs to come under the control of either the podcast 2.0 github or the PSP. It is currently languishing here and has not be updated for years.
2. The person tag list then needs to be extended to add AI as an role option.
3. The person tag needs to have a secondary role e.g announcer, AI. The secondary role can also be used for non-AI reasons. For Example podcasters who are hosts and editors. Or it might be the primary role is music artist and secondary is songwriter.

[samsethi]

We need both. Truefans has added support for both <Podcast:txt purpose=ai-content>truepodcast:txt for AI disclosure and Person Host= AI

In the image below you will see a toggle to self disclose AI. As an app we read the podcast:txt purpose=ai-content and dispaly a robot icon in the UI and as a host we add this to the podcaster's RSS feed.

We also support multiple roles. James is both the host and editor and I am the host and producer.

[samsethi]

@redimongo is suggesting we use AI to measure how much of a transcript is AI and then give it a score. Then in the app the user can select the percentage of AI they are willing to accept.

[jamescridland]

a) this isn't about AI voices. This is about use of AI. An entirely AI-generated script that I read out? That's still majority AI. That doesn't live in a person tag.

b) Use of podcast:txt is a temporary kludge. It's fine, but I hope we treat it as temporary, pending resolution into a detailed tag we want to use. We can literally push anything into podcast:txt but that shouldn't be our long-term goal.

c) What is absent from all of this technical discussion is the why. What is the use-case? What are we marking this for? All of the feedback to Alberto's shouldidisclose.ai that I've had all asks - yes, this is very cool, but WHY?

I'd really like to understand the c) bit first.

[picard102]

For my use case, as part of the person tag, it would be to easily filter who in the cast list is AI or not. We check new user registrations against the podcasts listed team, so it would be one less team member we'd have to check against.

As part of a episode or channel level tag, it would allow users to avoid AI content potentially.

[Marzal]

As a listener I would like to toggle off AI podcast when searching.
And also, know if the voice I'm listening it's human or not.