Google Is Renting Its Most-Cited AI Meeting Stats From a Different Product
Google Beam's most-cited demo number is not about the AI.
The company has spent the past week promoting Sophie, its new enterprise AI agent, using a statistic that describes something else. According to internal surveys cited across coverage, 90 percent of employees felt genuine physical co-presence during Beam tests. The figure appears in AI CERTs roundups, TechEBlog reports, and The Verge's hands-on writeup. It is the most repeated number in the Sophie launch narrative.
According to Google's own research summary, published May 20 on the Google Research blog, the company ran its hardware — the HP Dimension unit, a $25,000 display with six depth cameras, custom chips, and light-field compression — between human participants in different office locations. The 90 percent figure appears in the results section describing how those humans felt talking to each other through the hardware. The blog post's methodology section does not describe any sessions involving Sophie, the AI agent, and Google did not provide the original research paper on request.
Sophie is a separate product. She is a conversational AI rendered on a Beam display, wearing a dark turtleneck, built to appear lifesize. She runs on cloud servers — the Beam unit captures the local environment and streams it to Google's infrastructure for AI processing, rather than running the model locally. She can read documents, answer multilingual questions, pull up maps, and check the weather. In the current samples, Sophie appears in two dimensions on the screen, not the volumetric 3D that human Beam participants see of each other. Latency during the Sophie demo averaged under 300 milliseconds, which is fast enough to feel conversational. The product is currently running only at Google Labs in Mountain View for extended durability testing — there is no broader deployment yet.
Google has not published equivalent survey data for Sophie sessions. No internal number comparable to the 90 percent co-presence figure exists in any document cited by the company's researchers or covered by outlets that have tested the agent. The Verge's Sean Hollister documented Sophie's capabilities in detail — multilingual fluency, document reading, maps and weather — without citing a co-presence survey for AI-to-human interaction. AI CERTs, in a technical breakdown, lists Sophie's latency and specs without a corresponding user-feel metric.
The distinction matters beyond the surface similarity. Physical co-presence — the sense that someone is actually in the room with you — is partly a function of visual fidelity and latency. A 2D rendered agent on a screen operates in a different perceptual category than a volumetric 3D reconstruction of a human being, even at sub-300ms latency. The research describes one thing. The product is another.
A separate question lingers over the product: what happens to the data. Sophie runs cloud-side, which means every interaction — the documents shown to the camera, the conversations overheard, the room itself — is processed on Google's servers. The company's Beam group meetings research notes a 50 percent stronger sense of social connection and a 21 percent increase in reported ability to contribute to conversations for human-to-human Beam sessions. Those numbers are real. They are also not Sophie's numbers.
What to watch next: Google has not said when Sophie exits the Mountain View lab or what a broader rollout looks like. The company also has not published any user-comfort or co-presence metrics for the AI agent version. If Beam succeeds as an enterprise product, that question — do people feel the same sense of presence with an AI as they do with a human on the same hardware? — will be the difference between a meaningful interaction tool and a video call that happens to have a face on it.