Griffin-Lite 48% is 26 of 54 callers: how wide is the range?

The number and its sample

Tavus reports that 48% of participants believed they were talking to a real person in a one-minute video call with Griffin-Lite, against 2.4% for its earlier Phoenix-4.5 model. The Tavus page gives the group sizes: 54 participants for Griffin and 41 for Phoenix-4.5. At 54 people, 48% is 26 people (26 / 54 = 48.1%), and 2.4% is 1 person (1 / 41 = 2.4%).

How wide is the uncertainty

With a sample this small, the true rate could sit well away from 48%. A standard 95% Wilson score interval for 26 successes in 54 trials runs from about 35.4% to 61.1%. That is our own arithmetic, not a figure Tavus published. For Phoenix-4.5, one success in 41 trials is also a very small count, so the exact earlier rate is loose too.

• Center: 26 / 54 = 48.1%.
• 95% Wilson interval: roughly 35% to 61%.
• The two ranges do not overlap, so the jump from the older model is large even with this sample.
• The page does not describe a standard test protocol, so treat the figure as a vendor result.

What it does not tell you

The result is about a live, one-minute call, run by the vendor. It says nothing about a rendered clip, a longer conversation, or viewers who know an avatar may be involved. Tavus also notes the dual-use risk on its page: the properties that make these models useful interfaces also let them deceive a person into thinking they are not AI. Read the 48% as a signal about realism, not as proof of anything about your use case.

It also helps to separate two questions. The first is whether a model can look human in a short call. The second is whether your audience would feel misled by a presenter that looks human. The Tavus study speaks to the first only. The second depends on your channel, your claims and the rules of the platform where you publish.

What to do with it as a buyer

If you publish talking-head video, the useful takeaway is that viewers cannot be relied on to notice synthetic presenters. That points toward saying so. With Sume Avatar 1.0, the script drives the video, so a disclosure sentence can be the opening line of the script or the first caption cue, and you can approve the first frame before you pay for a full render.

Whatever you build, decide the disclosure wording before the avatar goes live, and keep it in the same place on every clip.

A small habit pays off: write the disclosure sentence once, store it with your brand notes, and paste it into every script. That way the wording does not drift between creators, and a later change to the wording is one edit instead of a hunt through old clips.