On October 10, 2026, the New York Post reported that Bay Area startup Tavus has unveiled "Griffin," a new AI avatar model the company says approaches Turing-Test-level performance in a job-interview setting. In Tavus's test, Griffin fooled 48% of 54 participants — that is, nearly half believed they were speaking with a human interviewer. Among participants who thought the interviewer was human, average confidence was 79%.
The company's CEO and co-founder is Hassaan Raza. Tavus has been building real-time digital humans for enterprise use cases, and Griffin is its latest — and most ambitious — model.
What Griffin Does
Griffin generates every pixel of every frame in real time, starting from a single reference image. That includes the face, arms, fingers, chair movement, shadows, and the background. The avatars breathe and blink naturally, react to conversation in real time, and — critically — can interrupt and be interrupted, the conversational back-and-forth that has long been the hardest part of making digital humans feel real.
The 48% figure deserves context. This was not the full Turing Test — Alan Turing's original 1950 proposal of a sustained, open-ended conversation where an interrogator tries to distinguish machine from human. It was a job-interview scenario, a setting with strong social scripts: the interviewer asks structured questions, the candidate expects a certain cadence, and neither party expects to be tricked. Structured social contexts are easier for AI to navigate than open-ended ones. Still, 48% of 54 participants is a striking result, and the 79% average confidence among those who were fooled suggests the illusion, when it worked, was convincing.
How Real-Time Video Diffusion Avatars Work
To understand what Tavus has built, it helps to understand the underlying technology at a high level.
Diffusion models are the same class of AI that powers modern image generators: they start from random noise and iteratively refine it into a coherent image, guided by learned patterns from training data. Video diffusion extends this across time, generating sequences of frames that remain consistent from one frame to the next.
Real-time generation is the hard part. Generating a single high-quality frame with a diffusion model traditionally takes seconds; video requires 24 to 30 frames per second. Achieving that in real time requires enormous optimization — distilled models, specialized hardware, and architectures designed to reuse computation across frames rather than generating each one from scratch.
From a single reference image means the model learns a person's appearance from one photo and then synthesizes novel poses, expressions, and movements it was never shown. The model must understand 3D structure — how a face turns, how fingers bend, how fabric moves — from 2D training data, and apply it consistently across frames.
Interruption handling is a systems problem as much as a modeling one. A natural conversation requires the avatar to stop mid-sentence when interrupted, process the interruption, and respond — all within hundreds of milliseconds. That demands tight integration between speech recognition, the language model driving the dialogue, and the video generation pipeline.
The Fraud and Deepfake Implications
The same technology that powers compelling sales avatars creates serious risks. Remote hiring is the most immediate concern: if an AI avatar can conduct a convincing job interview, it can also take one. Fraudulent candidates — or the organized operations behind them — could deploy photorealistic avatars to pass video interviews for remote roles, particularly in technical positions where the interview is the primary screening gate.
This puts pressure on identity verification. Companies hiring remotely will need to verify not just that a candidate exists, but that the person on the video call is the actual candidate, live, in real time. Expect growing demand for liveness detection, document verification tied to live sessions, and possibly new standards for high-trust remote interviews.
Beyond hiring, the technology raises familiar deepfake concerns: impersonation in video calls, social engineering attacks with a human face, and the erosion of "seeing is believing" as evidence. The 48% figure is a milestone worth marking — not because machines have passed the Turing Test, but because a specific, high-value social context (the professional interview) is now substantially automatable by a synthetic human.
The Product Use Cases
Tavus is not building this for fraud — the legitimate use cases are substantial:
- Sales avatars that conduct personalized product demos at scale, available 24/7, in any language.
- Training and coaching, where consistent, patient, endlessly repeatable role-play partners accelerate learning.
- Customer support, where a human face improves trust and comprehension compared to text chatbots.
The through-line is that video adds trust and engagement to automated interactions. That is exactly what makes the technology valuable — and exactly what makes its misuse dangerous.
Why It Matters for Founders and Investors
For founders, Griffin is a signal that the synthetic-human layer is arriving faster than expected. If you're building in hiring tech, identity verification, or remote-work tooling, the deepfake-interview problem is now a product requirement, not a hypothetical. If you're building in sales, training, or support, photorealistic avatars are becoming a viable interface — but plan for the trust backlash and build verification into your product from day one. For investors, two categories look increasingly essential: the avatar infrastructure itself (generation, real-time pipelines) and the counter-technology (liveness detection, provenance, verification). Every powerful synthetic media advance creates a matching market for authenticity.
Angeled.com connects founders with investors.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



