Viewers suspect AI within 20 seconds: disclose in the first frame

What Tavus reports

Tavus's Griffin page, read 2026-10-03, says 26 of 54 participants (48%) believed a one-minute call with Griffin-Lite was a real person, against 1 of 41 (2.4%) for Phoenix-4.5. It says those who doubted usually suspected within the first 20 seconds, and that participants were not told during the call they were speaking to an AI; every participant was told afterward.

Tavus itself writes that the properties that make such models useful can let them deceive a person into believing they are not AI, and says it is working on safe disclosure features before wider release.

What that means for a recorded avatar

A recorded avatar clip does not need to fool anyone, and a clip that hides its nature is a liability. If most doubts arrive within 20 seconds, a disclosure that appears at second 15 reaches viewers after they have already decided. Put it in the first frame.

Sume's avatar video can burn text in with the captions option, but captions are built from the spoken script, so a separate label line is easiest to place as the first words spoken or as an overlay added afterward. Check the avatar video docs for the caption styles, and keep the statement short.

{
  "avatar_handle": "brand_host",
  "aspect_ratio": "9:16",
  "script": "I'm an AI presenter for this brand. Here is what's new this week.",
  "captions": {"enabled": true, "style": "punch", "language": "auto"}
}

Practical rules

• Say it in the opening line and show it as text from frame one.
• Keep the clip within the 4 to 60 second range per job; short clips make the opening count more.
• Store the job id and the script with the published clip, so you can show what was said.
• Rules differ by country and platform; check the one you publish on.

What this does not settle

The Tavus study is about a live call with a model that talks back, run on 54 people for one minute each, and Tavus reports it itself. It does not measure a recorded clip, and it does not say how viewers of a feed react. Use it for one idea only: if the suspicion window is short, do not delay the label.

Rules on AI labels vary by country, platform and ad type. Read the rules of the place you publish, and keep the label in the video itself rather than only in a caption field the viewer may never open.