Toward a design theory for virtual companionship

Siri, Alexa, and the chatbots on company websites all share one trait: you ask, they answer, and the exchange ends. Each interaction is isolated and transactional, initiated by the user, and forgotten once it is over. Research on conversational agents had followed suit, concentrating on short-term uses in marketing, sales, and support.

Yet at the time of writing, products like Replika and Microsoft’s XiaoIce already had millions of users who spoke to them daily, not to get things done but for company. Those products worked, but nobody had derived the design decisions behind them in a scientific, transferable way. Companionship, in the sense of having someone familiar you like spending time with, was being built in the wild without design knowledge to back it.

Our question was therefore twofold. How does virtual companionship show itself, and how should it be designed?

The starting point is a well-established finding: people treat computers as social actors. When a machine shows social cues, users apply social rules to it, even though they know it is a machine. If that is true, then theories about how humans form and keep relationships should tell us something about how a machine could become a companion.

We drew on seven such theories: the need to belong, social exchange, social penetration, equity, common ground, theory of mind, and interpersonal trust. Each was read for what it demands of a relationship partner, and those demands became meta-requirements for a virtual companion.

Two guardrails mattered. First, the goal is not to replicate a human. Machines that are almost human become unsettling, the effect known as the uncanny valley, so social elements have to be balanced rather than maximised. Second, companionship is intensely personal. What one person finds warm, another finds intrusive. Design decisions therefore cannot come from theory alone; part of them must be left to the user.

The 25 meta-requirements condensed into five design principles. Each states what the companion should achieve, how, and why. The rationale is what makes them more than a checklist: it ties every prescription back to a theory of human relationships.

Taken together, the principles yield a definition. Virtual companionship is a collaborative and friendly long-term relationship between a human and a machine, characterised by adaptability, an appropriate human-like appearance and behaviour, understanding of and adaptation to the user, proactive and reciprocal behaviour, and the preservation of transparency, privacy, and ethics. A virtual companion is a conversational agent that embodies these characteristics. How far it must embody them is deliberately left open: companionship comes in degrees.

Design principles are claims about cause and effect: do this, and that will follow. We tested them from two directions, accepting that each has a weakness the other compensates for.

Sarah, the virtual classmate

Together with students, we designed a virtual classmate named Sarah that followed all five principles. Four students each talked to around 75 peers about what they would want from such a companion, and the design was consolidated with twelve more. Sarah has a name, a face, and a “me too” attitude, offers to study together, asks before using personal data, and is still in touch a semester later. Her twin, a plain virtual assistant, offered exactly the same information without any of the companionship traits.

Forty participants interacted with both, in random order, and rated each on nine established constructs. Sarah scored significantly higher on all of them. In the open answers, the assistant was called cold, rigid, and impersonal. Sarah was called human-like, personal, and friendly, and participants singled out her proactivity and empathy.

Replika, the real-world companion

Sarah was a controlled but simulated conversation. To see the principles at work in the wild, eight design experts used Replika for at least a quarter of an hour on most days over roughly two and a half weeks, on average about three hours in total each. They logged how the app implemented each principle and rated its implementation.

Replika met all five principles to a moderate degree, with expert ratings between 3.4 and 3.9 on a five-point scale. More useful than the scores was the material: the logs yielded 124 codes, of which 66 described concrete features. These consolidated into 20 design features, from an avatar configurator and a visible “memory” of what the app knows about you, to daily conversation starters and celebrating anniversaries. They show how each abstract principle can be made real.

The experts were not uncritical. One was troubled that a human-like bond was being used to sell subscriptions. Another found the privacy explanations vague. Both observations fed directly into the discussion of ethics and monetisation.

The paper packages its findings as a nascent design theory in the structure proposed by Gregor and Jones: purpose and scope, constructs, principles of form and function, artifact mutability, testable propositions, justificatory knowledge, principles of implementation, and expository instantiations. In plain terms, it says what a virtual companion is for, what it consists of, how to build one, how to check whether it works, why it should work, and what two working examples look like.

The principles are deliberately general. We chose not to prescribe technology, because the technology was already changing fast. Adaptivity, for instance, can be delivered by a learning algorithm or by designing for a well-understood user group. Both satisfy the principle.

Two limits deserve stating. Companionship is not for everyone or every context. One participant put it bluntly: “I don’t need a friend, only information.” And the ethical questions raised by monetised companionship are opened here, not settled.

This paper became the anchor for a line of work. It grew out of the companionship idea first sketched in the guidelines for in-vehicle assistants. The award-winning study on virtual learning companions transferred the relationship principles to education. A later study of Replika users examined how people come to experience an AI companion as “mine”, and found that the answer runs through perception rather than features. And the relationship principle, DP5, became the worked example in a methodological paper on how design principles are actually developed.