Skip to main content

When 48% can't tell: Tavus Griffin and why disclosure is part of the product

By Alan Magliocca · November 14, 2026

Quick answer: On October 1, 2026, Tavus unveiled Griffin and said 26 of 54 people in its own one-minute test thought its avatar was a real person; the model is a research preview, not for sale. If you run an AI presenter, put the disclosure on the registration page, in the reminders and in the first 20 seconds of the event.

KEY TAKEAWAYS
  • Tavus says 48% of people in its company-run test (26 of 54) thought Griffin-Lite was human after a one-minute call.
  • Participants were told to expect a person and were asked about it only at the end; Decrypt notes the results aren't independently verified.
  • Griffin-Lite is a research preview for select testers, and Tavus says it is working on disclosure features before a wider release.
  • If attendees can't tell an AI host by sight, disclosure has to be built into the event: label, opening line, chat name and replays.

What happened

On October 1, 2026, Tavus unveiled Griffin, which it calls a "Human Interaction Model," in an announcement on Business Wire. Griffin is a full-duplex video-to-video model. In plain terms, it holds a face-to-face conversation in real time instead of waiting for its turn like a chatbot with a face.

The number that made headlines comes from a test Tavus ran itself. According to Decrypt's report, Tavus says 26 of 54 participants, or 48%, believed they were talking with a real person after a one-minute video call with Griffin-Lite, the version it tested. Its previous system convinced 1 of 41 people.

How the test was run matters as much as the result:

  • Participants were told they'd be matched with another person to talk about what they were looking forward to this year.
  • Nobody was told a bot might be on the call. Only at the end were they asked whether it had crossed their mind that their partner wasn't real.
  • Tavus says the people who did grow suspicious usually did so within 20 seconds.
  • Decrypt notes a community note on X saying the results aren't independently verified and don't follow a standard protocol.

Tavus also points to NVIDIA's VideoFDB benchmark, which it says NVIDIA ran independently. On the generation track, Griffin-Lite scored 3.83, against 2.80 for the next-best system and 3.92 for a human reference.

You can't buy Griffin yet. Griffin-Lite is a research preview for select trusted testers, and The Decoder reports that the more capable version is held back until "safety concerns are addressed." Per Decrypt, Tavus says it is working on disclosure features and with AI safety organizations.

What it means for people selling with webinars

Read the 48% for what it is: a company's own claim, from one-minute chats where people were told to expect a human. A 45-minute sales webinar with a pitch and a live chat is a very different setting. Still, the direction is plain. AI presenters are getting harder to spot on sight.

That moves disclosure from the footer into the product. When avatars looked obviously synthetic, the look did the disclosing for you. Once people can't tell, the only disclosure your audience gets is the one you build in. Tavus itself treats disclosure features as work to finish before a wider release. Webinar hosts can apply the same logic: the on-screen label, the spoken line and the named chat assistant are parts of the event, like the join link.

There's a sales reason as well as a legal one. A sales webinar ends by asking people to trust you with money. Someone who buys and later learns the friendly host was software may feel misled, and that's a poor start for a customer relationship. Someone who knew from the first minute has nothing to discover later.

The rules point the same way. For EU attendees, the AI Act already expects disclosure from the start of the first interaction; see EU attendees must see the AI disclosure first. Whether a disclosure duty applies to your own events is a question for a lawyer.

What does a disclosed version of the Tavus test look like?

The Tavus test removed every cue on purpose. A disclosed webinar puts the cues back. Here is a worked example for an evergreen webinar run by an AI version of a real host, Dana:

  1. Registration page. "This session is presented by Dana's AI presenter. Dana wrote the material, and a person on her team reads the questions." Use only claims that are true for your event.
  2. Confirmation and reminder emails. The same sentence, word for word, so nobody arrives surprised.
  3. The first frame. A visible "AI presenter" label on screen, kept in place for the whole session.
  4. The first 20 seconds. Tavus says suspicious participants usually noticed within 20 seconds. Put the spoken disclosure before that point, for example: "Hi, I'm Dana's AI presenter. Everything you'll hear today comes from Dana's own playbook."
  5. The chat. If an assistant answers questions, give it a name that says it's AI.
  6. The offer. Say which real person or company stands behind the offer and how to reach a human before buying.
  7. Replays and clips. Keep the label and the opening line in every replay and every clip cut from the event.

For ready-to-use wording, see a plain-English disclosure line for AI-hosted webinars.

Which disclosure questions should you ask an AI presenter vendor?

If disclosure is part of the product, it belongs in your vendor checklist next to price and voice quality. Ask:

  • Can the AI label stay on screen for the whole session, including replays and exported clips?
  • What does the presenter say if an attendee asks, "Are you real?" For live, interactive avatars, the answer should always be yes, it's AI.
  • Does the output carry machine-readable marks, and can you keep them through editing? The EU's December 2 deadline for those marks is covered in December 2 deadline: machine-readable marks on AI video.
  • Whose likeness is it, and is written consent on file for the face and the voice?
  • What do the vendor's terms say about undisclosed use? Read them before you build a funnel on the tool.
  • Can disclosure be switched off? If it can, set your own rule that it stays on.

Keep the answers in writing. They're useful later if an attendee, a platform or a regulator asks how your AI host was presented.

What to do this week

  1. List every AI element in your events: presenter, voice, chat assistant, replays and clips.
  2. Write one disclosure sentence that covers all of them, and put it on your registration page and in your reminder emails before your next scheduled event.
  3. Rewrite the opening of your presenter script so the disclosure is spoken within the first 20 seconds.
  4. Rename your chat assistant so its name says it's AI.
  5. Add one line near the offer that tells buyers how to reach a person.
  6. Send the vendor questions above to your avatar provider and file the answers.

How BrightStage AI helps

BrightStage AI's video creator builds a script and slides with AI presenters, so the opening disclosure line goes straight into the script you build there. During the event, a chat assistant answers attendees in the event chat, and replays run as Encore, so check that the label and the spoken line carry through both. You can split-test registration pages (A/B/C), which is a good way to compare how you phrase the disclosure, never whether it appears. Agencies making AI video for clients face the same question from the other side; AIVX Labs covers it in why disclosing first beats getting caught.

Sources

AIVX LABS ACCOUNT VERIFICATION
Get AIVX Labs verified for $1: RICO, your AI operator, and 35+ AI tools, with 7 days of full access.
Get AIVX Labs Verified for $1 →

Frequently asked questions

Did 48% of people really mistake Griffin for a human?

That is Tavus's own claim: 26 of 54 participants on one-minute calls, who had been told they'd meet another person. Decrypt notes the results haven't been independently verified.

Can I use Griffin to host my webinars?

Not now. Griffin-Lite is a research preview for select trusted testers, and the more capable version is held back until Tavus addresses safety concerns.

Where should the AI disclosure go in a webinar?

Before anyone could wonder: on the registration page, in the reminders, on screen from the first frame and in the opening spoken line. Keep it in replays and clips too.

Related reading

Written by Alan Magliocca, founder of BrightStage AI and AIVX Labs (Podavinci LLC). Written with AI assistance and checked against the linked sources. Follow along: LinkedIn · YouTube · TikTok · X · Facebook · Instagram · Skool community