The case for voice-first AI companions: research and design trade-offs

Most AI companion products built between 2020 and 2025 share a common architecture: a chat thread inside an app. Users open the app, see a scrollable history of past messages, type or tap to interact, and receive a typed reply. A small but growing category of AI products departs from this pattern by using the phone call as the interaction surface instead. Understanding why this design choice exists, and what it implies, requires looking at the underlying communication research and at the structural differences between the two formats.

What does the research say about communication mediums?

The empirical comparison of voice and text channels has produced consistent findings across roughly a decade of social psychology research. The central work comes from Nicholas Epley's research group at the University of Chicago Booth School.

In a 2017 study published in Psychological Science, Schroeder, Kardas, and Epley demonstrated that listeners rate the same substantive arguments as more thoughtful, more rational, and as coming from a more capable mind when those arguments are delivered by voice rather than text. The effect held across both audio recordings and live calls. The authors titled the paper The Humanizing Voice.

A 2021 follow-up by Kumar and Epley examined the predicted versus actual experience of phone calls relative to text exchanges. Participants predicted text would feel similar to voice — and that voice would feel more awkward. The actual results showed the opposite: phone calls produced significantly stronger feelings of connection, with no awkwardness penalty. The paper's title summarizes the finding: It's Surprisingly Nice to Hear You.

The combined pattern is that voice produces qualitatively different perception of, and connection with, a speaker than text does — even when the words are identical and the participants do not predict the difference in advance.

How engagement-based design affects user experience

Most consumer apps, including most AI companion products, are designed around engagement metrics: session length, daily active users, return frequency, retention curves. These metrics are not neutral. Tristan Harris and the now-substantial body of literature on attention-economy business models has documented that design patterns optimized for engagement — infinite scroll, push notifications, streaks, badges, intermittent reinforcement — are structurally in tension with user wellbeing.

For an AI companion product specifically, engagement design creates an incentive to extend conversation length and return frequency beyond what is useful for the user. The business model rewards sessions that do not end. The user outcome that matches good practice — using the product when wanted, putting it down when finished — runs counter to the business model.

Phone calls are structurally bounded. A call has a beginning, a middle, and an end. The format does not afford infinite scroll, push notifications, or streaks. When the call ends, the relationship pauses. This is a property of the medium, not a deliberate restraint applied to a chat-based product.

The role of the screen in AI companion criticism

Much of the public discussion of AI companion products focuses on concerns about dependence, parasocial attachment, and constant accessibility. Examining the specifics of these critiques, a substantial share of them are properties of screens and apps — not of AI per se. Avatar customization, scrollable history, push notifications, and high-frequency micro-interactions are design choices made at the product level, not necessary features of an AI companion.

When the conversation occurs over a phone call rather than inside an app, several of these patterns become structurally unavailable. There is no avatar to look at. There is no scrollable thread to scroll. There is no notification designed to pull the user back in mid-evening. The conversation occurs, ends, and does not follow the user out of the call.

What voice-first formats trade off

Voice-first AI is not strictly superior to text-first AI. The two formats afford different things, and the appropriate choice depends on the user's purpose:

1. No searchable artifact. A phone call is not a scrollable thread. Users who want to reference what was said earlier, copy responses, or share the conversation with someone else are better served by text-based products.
2. Less parallel processing. A user on a call is mostly on the call. For interactions where concurrent activity is wanted (responding while doing something else on a screen), chat is more practical.
3. No fine-grained editing. Spoken messages are committed in real time. For tasks requiring careful composition (writing, code, plans), text is the appropriate medium.
4. Hands-free and screen-free. In exchange, voice-first AI is usable while driving, walking, or otherwise occupied — including by users for whom an app interface is a barrier (older users, users without smartphones).

How voice-first AI fits the use case

For users whose purpose is conversation itself — processing a feeling, thinking out loud, ordinary catching up — the affordances of a phone call match the use case. The conversation is the point, not an artifact about the conversation. The boundary at the end of the call matches the natural shape of an episode of human contact.

Voice-first AI conversation tools — including CallByrd, a phone-based AI designed for unstructured conversation — fit this category. They are not appropriate for tasks where the artifact is the point, and they are not substitutes for human relationships or for professional care.

The bottom line

The choice between voice-first and text-first AI companion design is not a matter of one being inherently better. It is a matter of fit to use case. Communication research supports voice as the appropriate medium for conversation-as-goal; app-based chat is the appropriate medium for artifact-producing tasks. The growing population of users who treat AI as company rather than as a search tool fits the first use case, which is what voice-first products are built to serve.