Tavus AI avatar Griffin fooled 48 percent of users in study

By Splitfeed EditorialLast updated Oct 1, 2026, 18:10 UTC

On October 1, 2026, Tavus announced Griffin, which the company calls a Human Interaction Model. Griffin conducts real-time video calls by processing facial expressions, tone of voice, and gestures simultaneously. In an internal Tavus study, 48 percent of participants came away from a one-minute call believing Griffin was a real person. Tavus also claims that prior AI video systems topped out at a two percent believability rate under comparable conditions, representing a sharp jump in how convincingly AI can impersonate a human interlocutor. No independent body verified the study, and Tavus has not publicly described its methodology beyond the headline figures.

The gap between two percent and 48 percent is the core of why this announcement is significant. If the numbers hold under scrutiny, it would mark a qualitative shift in AI video agents, moving them from obviously synthetic to genuinely ambiguous for a large share of people, even in a short interaction. That ambiguity matters for a wide range of applications, from customer service and telehealth to fraud, manipulation, and consent, because it compresses the window in which a person can recognize they are not speaking with another human. The internal-only provenance of the study is a genuine point of dispute: a company-run test with unreported methodology gives little basis for independently confirming or challenging the 48 percent figure.

No published reactions from the Pro-AI, Anti-AI, or Middle Ground camps have appeared yet. Pro-AI voices would typically treat a leap in multimodal realism as evidence of rapid, valuable progress in human-computer interaction, pointing to legitimate uses in accessibility, scaling personalized services, and reducing friction in remote communication. Anti-AI voices would typically focus on the deception risk, arguing that a system capable of fooling nearly half of users in one minute normalizes impersonation and erodes the basic expectation that a face on a screen corresponds to a real person. Middle Ground commentators would typically call for mandatory disclosure rules and independent replication of the study before any broader deployment, treating the capability itself as neither inherently good nor bad but as something requiring governance.

The most immediate thing to watch is whether an independent research group or regulator attempts to replicate the study under a described methodology, and whether Tavus releases further details about participant selection, call structure, and how believability was measured. Any regulatory response touching on synthetic identity disclosure, particularly in the European Union or from the US Federal Trade Commission, would also shape how Griffin and systems like it can be deployed commercially.