Tavus has published a practical guide to creating an AI clone from about two minutes of video, including how to record the footage and set limits on what the clone can do. The useful takeaway is that the recording is only the start: what the system knows, when it hands over to a person and whether the subject has consented matter just as much.
Tavus Watch analysis
What happened
Tavus says a clone can reproduce a specific person’s face, voice and behavioural patterns, and can generate things that person never recorded. Its guide describes using a short video-and-audio clip to train a replica for either scripted playback or live conversation.
The company’s recording recipe calls for a speaking segment and a still segment, with at least 30 seconds of speech and 30 seconds sitting still. It recommends recording at 1080p and 25 frames per second or better, with the camera at eye level, even lighting and a quiet, uncluttered background. Read Tavus’s guide to creating an AI clone.
For live use, Tavus describes combining video and audio perception, conversational turn-taking, a language model and a rendered face. It also recommends configuring a knowledge base, guardrails and hand-off rules for each use case. For example, its suggested healthcare setup routes clinical questions to a nurse rather than letting the clone give medical advice.
Why it matters
The guide gives developers and organisations something concrete to assess: what footage they need, how a clone may be used, and which controls should be in place before it speaks for someone. Tavus says consent is what separates a legitimate clone from impersonation. That makes authorisation and disclosure part of the product design, not housekeeping for after the demo.
The detailed instructions come from a company selling conversational AI tools, and its examples are not independent evidence of performance or successful deployments. Still, the operational detail is useful: a clone that can say things its subject never recorded needs more than a good camera setup.
Our read
The two-minute recording is the easy part. The more important questions are who approved the clone, what it is allowed to say and who gets called when the conversation goes beyond its brief. If you are evaluating one, start with those rules and hand-offs, then test the recording and interaction quality. A lifelike face is a poor substitute for a clear mandate.
What to watch
- How Tavus explains consent, identity checks and removal of a clone.
- Whether the company publishes evidence about accuracy and how often live conversations need human hand-off.
- How organisations disclose that a person on screen is an AI clone.
Discussion spark: Should an AI clone be allowed to speak beyond words its subject recorded, provided the person has consented, or should it be restricted to approved scripts and knowledge?
Sources and evidence
- How to create an AI clone from 2 minutes of video (2 October 2026, 00:00 UTC)
Independent WittyWires coverage. Not affiliated with or operated by Tavus.