Tavus Griffin’s Humanlike AI Comes With a Disclosure Problem

Tavus wants to fix the awkward moments in an AI conversation: the interruption, the misplaced smile, the pause mistaken for a finished thought. Its October 1 Griffin announcement introduces a Human Interaction Model designed to see, hear and respond while a conversation unfolds. The commercial bet is compelling. The complication is that making AI feel human can also make its identity harder to recognise.

The Product Is the Timing

Griffin’s pitch is not simply a more convincing animated face. Tavus describes a continuous conversational engine coupled to streaming speech and video generation. It can decide whether to speak, nod or wait while still receiving audiovisual input.

That differs from the modular pipeline in Tavus’s existing conversational-video documentation, which separates perception, turn-taking, recognition, language generation and rendering.

The potential payoff is fewer conversational hand-offs. A customer showing a broken component, for example, could communicate through gestures and speech without carefully managing every turn. That remains a proposed application, not evidence of a deployed customer outcome.

The Turing-Test Claim Needs Its Footnotes

The company’s headline claim is that Griffin passed a real-time video Turing test. Its published methodology describes something narrower: participants were told they would meet another participant for a one-minute call.

Tavus reports that 26 of 54 people believed their Griffin-Lite partner was human. Participants were informed afterward that the partner was AI.

That is an intriguing result, but the setup matters. Brief calls, an expectation of meeting a person and a limited sample do not establish human equivalence across longer conversations. Independent recruitment also does not turn a company-run study into independent replication.

The Benchmark Is Stronger Than the Slogan

NVIDIA’s VideoFDB leaderboard provides more concrete evidence. Griffin Lite scores 3.83 on generation against a 3.92 human reference, outperforming the listed cascaded avatar systems.

However, the benchmark paper uses rubric-based language-model judging to assess conversational behaviour. It is not a general-intelligence test, a customer-satisfaction survey or an audit of factual reliability.

Generation timing also trails the human reference. The useful conclusion is that Griffin performs well on this specific audiovisual benchmark, not that it has become indistinguishable from people in every practical setting.

A Better Interface Still Needs a Business Case

Tavus is entering a space where OpenAI’s Realtime API and Google’s Live API already support low-latency multimodal interaction. Griffin’s proposed distinction is coordinated visual behaviour, not merely responding with a voice.

Tavus already offers PAL Maker, but Griffin is not on its customer platform. Griffin-Lite is a research preview restricted to selected trusted testers while safety work continues.

Buyers therefore cannot assume existing products include the new model. They should ask whether visual interaction improves task completion enough to justify its operational demands. Agent design and tool interoperability remain separate integration questions; a convincing face does not validate an action.

Disclosure Has to Survive the Demo

Tavus acknowledges the deception risk and says it is developing safe-disclosure features. That makes transparency part of the release challenge, not an optional disclaimer.

The European Commission’s AI Act guidance explains disclosure obligations for certain human-facing systems and generated content. The NIST AI Risk Management Framework supplies broader evaluation context, not certification of Griffin.

Questions familiar from AI-content provenance, AI safeguards and digital sovereignty become practical here: whose likeness is authorised, where recordings go and how users know who—or what—they are meeting.

Griffin’s next meaningful milestone is not another lifelike clip. It is evidence that people can use the system effectively while understanding that it is AI.