Tavus unveils Griffin: 48% of participants mistook its AI video avatar for a real person
lmoroney · x · 2026-10-03
Tavus has introduced Griffin, which it calls a Human Interaction Model: a full-duplex video-to-video system that listens and watches continuously, then responds with speech, facial expressions, gestures, and body movement in real time.
Key numbers: in a face-to-face study, 48% of participants (26 of 54) who talked with Griffin-Lite for one minute believed their partner was human. On NVIDIA's independent VideoFDB benchmark, Griffin-Lite led published baselines on both the generation and perception tracks. It's currently a research preview for select testers while the company works on disclosure and safety features before wider release.
The author also offers a practical checklist for educators: disclosure, consent, and review — students should know when a partner is AI, and someone should own what the model sees on camera.
Related event: Tavus Launches Griffin, Claims First to Pass the Video Turing Test(16 posts)→
More from Models
- Cloudflare's Clef decision models (27B and 9B) now available on Ollama — ollama · 2026-10-03
- Clef Flash: 9B multimodal decision model fine-tuned from Qwen3.5, 256K context — ollama · 2026-10-03
- "Random statistical predictions" misreads LLMs: models learn internal algorithms — burny_tech · 2026-10-03
- TypeSafe's Jev undercuts frontier labs: $0.042 per million input tokens, now eyeing $10B valuation — mattturck · 2026-10-03
- Google introduces tiered Gemini access: model choice and limits now vary by plan — QH96 · 2026-10-03
- Meta Open-Sources Muse Code to Put AI in Your TV and Toaster — TechCrunch AI · 2026-10-03