AI avatar disclosure line as the first caption cue on a Sume clip

To label an avatar clip as AI on Sume, burn an authored text cue into the video with standalone Video captions: send cues (or segments) with text, start and end, and no speech-to-text runs. The Tavus Griffin page says its model can deceive a person into believing it is not AI and that the company is working on safe disclosure features, so the label is something you should add yourself on any realistic avatar.

How the cue route works

The docs describe cues (or segments) as authored overlay text, with text, start and end, burned without speech-to-text. That is exactly a fixed label. Use a Latin style for English text. If the label is Korean, select a Hangul style such as black-outline, because Sume rejects Korean text on slam, punch or tiktok-green with 400 caption_hangul_text_latin_style.

The standalone job takes a public HTTPS video_url, so run it on the finished avatar MP4. This is a second step after generation.

What the label should say

Keep it plain and early. Show it for the whole clip or at least the first three seconds, long enough to read at phone size.

• Say what it is: AI-generated presenter.
• Say who stands behind it if viewers could confuse it with a real person.
• Repeat it in the post text, where the platform allows.

What this does not settle

A caption cue is a visible label. It is not a legal opinion on the disclosure rules of any country or platform, and Sume's docs make no such claim. Check the rules for where the clip will run, and keep the label in the master file so every cut inherits it.

Order of operations

Generate the avatar clip first, with or without inline captions for speech. If you want both speech captions and a label, caption the speech inline, then run the standalone job on the resulting clean video with only the label cue, or send all cues together in one standalone job. Keep one master file that has the label so that every later edit inherits it.

Put the label's start at 0 and its end at the end of the clip for a short video. For a long one, a repeating label every 15 seconds costs nothing extra in the cue list, though it does take screen space.

Test the burned label on a small phone screen before publishing. Make sure it does not collide with platform interface elements at the bottom or top of the frame; the placement design fields can move it for styles that support design overrides.