Griffin-Lite safety hold: what to prepare now for AI video calls

Tavus says the reason Griffin-Lite is limited to select testers is deception risk: its own page states that the same properties that make a Human Interaction Model a natural interface 'allow them to deceive a human into believing it is not AI' (read 2026-10-03). If you will put any realistic AI presenter in front of customers, prepare three things now: documented consent for any real face, a plain-language AI disclosure, and a recorded fallback you can switch to.

None of these depends on one vendor's timeline. They are the same whether the presenter is a live video model or a pre-rendered clip, and teams that have them can adopt new tools quickly while teams that do not will be stuck in review.

Consent for any real face

Sume lets you create an avatar three ways: from a prompt, from structured traits, or from a reference image. A prompt or traits avatar is invented. An image avatar is based on the photo you supply, so if that photo is a real person, you need their agreement. The avatar docs require the image to be a fetchable public HTTPS URL, which is a technical rule and not a consent check.

Keep a record per avatar: whose likeness, what they agreed to, for which channels, and until when. A short template saved with the avatar handle is enough to start. Sume does not provide legal advice, and the rules vary by country, so involve your own counsel for anything beyond internal use.

Disclosure that a viewer actually sees

Tavus's company-run test found that 48% of 54 participants believed a one-minute call was with a person. Whatever you think of that study design, it is a good reason not to rely on viewers noticing. Say it in the video or its caption: a spoken line in the script, such as 'I'm an AI presenter', is the most robust approach because it cannot be cropped away. Sume's avatar scripts are text you author, so the disclosure is yours to write, and inline captions will burn it into the final MP4 when enabled.

Platform rules for synthetic media differ and change quickly, so check each channel's current page before publishing.

A recorded fallback

Real-time systems fail in ways files do not. If you ever put a live avatar on a page, keep a pre-rendered clip of the same presenter ready, so that a timeout or a safety block degrades to a video instead of a blank panel. Rendering a 30-second fallback on Plus costs $7.35 at the listed rate (read 2026-10-03), which is cheap insurance.

Store the finished job's result URL and the job id with the clip. The documented job envelope gives you the status and result URLs, and failed jobs expose public error metadata such as retryability and next action.

Who should own the decision

These three items fall between teams. Legal cares about consent, marketing about disclosure wording, and engineering about the fallback. Assign one owner for the whole checklist, and make the avatar handle the unit of record: every avatar in your workspace should have a consent note, a disclosure line approved for its use, and a fallback clip. A new presenter cannot go live until all three exist.

This is also how you keep up with fast-moving vendor news without churn. When a preview such as Griffin-Lite eventually reaches general availability, the question for your team becomes whether the vendor meets your checklist, not whether to reinvent it.

What Sume does and does not do here

Sume produces script-driven clips from avatars you create. It does not run live video conversations, and it does not decide whether your use is lawful or sufficiently disclosed. Those obligations stay with you. What it gives you is a clear, auditable trail: a job id for each render, the script text you authored, the avatar handle used, and the result artifact URL on media.sume.com.

Store that trail with each published clip. If a platform or a regulator asks how a video was made, you can answer in one record instead of reconstructing it.