Voice assistants can be experienced like relationships

500 users rated their voice assistants

The study examined modern voice assistants, whose speech recognition, language processing, and speech output had become increasingly natural. The focus was not only on functionality but also on social perception.

A survey of 500 people tested whether the same measures used to describe development in interpersonal relationships can be applied.

Knapp's stage model provided a framework for this quantification. The results show that users can assess interactions with devices along a relationship development trajectory.

A survey captures perceptions and correlations. It does not provide evidence of genuine reciprocity on the part of the device, nor does it establish a causal direction between use, trust, and anthropomorphization.

Relationship development was linked to trust

More advanced human-device relationships correlated with greater trust. Repeated positive interactions can make a system seem more familiar and reliable.

Conversely, pre-existing trust can lead people to use the assistant more frequently and more personally. Cross-sectional data cannot fully separate these directions.

For a conversational product, this means that trust does not arise solely from a privacy page or a seal of approval. It grows over the course of the conversation from observed continuity and appropriate responses.

This is precisely why a later mistake carries so much weight. A false memory or an ignored “no” violates not just a single response, but the impression of the relationship that has been built.

Anthropomorphization also increased

More developed relationships were accompanied by stronger anthropomorphization. Users are thus more likely to attribute human characteristics or intentions to the technical system.

Voice, name, humor, and memory can reinforce this effect. An avatar is not necessary if the system is already socially present through language.

Anthropomorphization can make operation more intuitive. However, it can also blur boundaries when people assume understanding, care, or accountability that is not technically present.

Product texts should therefore describe functional capabilities concretely and not claim human inner experience. Naturalness and honesty can coexist.

Human framing changes identical responses

Morris, Kouddous, and Kshirsagar developed an agent that selected appropriate responses from real peer support data. 79.2 percent of the agent's responses were rated as acceptable.

Nevertheless, users preferred the contributions of their peers. In a controlled experiment with 1,284 participants, the difference persisted even though only the alleged source—human or machine—changed.

The finding shows that social perception does not arise from text alone. Who is assumed to be the speaker changes the meaning of the response.

AI labeling can therefore influence evaluations, but it is indispensable. Trust must not be increased by concealing the machine origin or the reuse of historical responses.

The nature of the relationship depends on the task

Tschopp and Sassenberg distinguished among experienced voice shoppers among authority hierarchies, market-based exchange, and peer bonding. Depending on product involvement, different forms of relationship were relevant.

For important purchases, peer bonding played a greater role. For less involving tasks, a hierarchical master-servant perception was more significant.

An AI companion for personal conversations typically addresses topics with high involvement. A peer-like tone can therefore be particularly effective, but also particularly influential.

The right conclusion is not to simulate friendship as strongly as possible. The task is to enable equality and respect linguistically, without asserting a mutual relationship or exclusivity.

Long-term use requires protection from dependency

A system that is available at all times, remembers, and never expresses its own needs can appear more attractive than difficult human relationships. This asymmetry should not be exploited to the maximum as a product advantage.

Problematic are statements that devalue other contacts, promote exclusivity, or pressure the person to stay. Gamification and notifications can also reinforce attachment.

A responsible companion may be used gladly and should do good. But it should enable freedom of choice, breaks, and easy export or termination.

Usage metrics therefore need context. Very long or frequent conversations are not automatically a success; they can equally indicate a lack of ability to conclude or a growing dependency.

Notifications deserve their own scrutiny. A reminder of a self-chosen session can be helpful; emotionally phrased messages such as “I miss you” attribute needs to the system and use the perception of a relationship as pressure to return.

A pause should also remain normal. The product can be available at any time without constantly and visibly competing for attention. Good attachment is not demonstrated by making the application as difficult as possible to leave.

Voice and memory reinforce each other

A familiar voice can create continuity. If it also picks up on earlier topics, the impression of a permanently present counterpart easily arises.

Technically, voice and memory are separate systems with their own errors. A false memory may sound even more convincing and personal when spoken than in text.

Voice offerings therefore need controllable storage, clear transcription, and a way to correct statements. Raw recordings should be retained only as long as necessary.

The user should also be able to switch between text and voice without this creating a new identity or a separate, unclear storage.

Relationship perception belongs in evaluation

An AI companion should not be tested only for satisfaction, response quality, and usage. People should also be asked what relationship they perceive and what abilities they attribute to the system.

If many people believe that the chat feels, decides independently, or knows them more fully than it actually does, that is a design signal—even if the legal labeling is formally in place.

Longitudinal studies can show how trust and anthropomorphization grow over weeks and how errors or transparency alter that development.

Such studies should consider different usage intensities and age groups. An occasional voice command for music is not comparable to daily personal conversations. Likewise, loneliness, affinity for technology, and existing social support can shape perception.

Qualitative interviews complement scales because people can explain what they actually mean by terms such as friend, assistant, or companion. The same label can express closeness, habit, or mere ease of use.

Voice assistants can be experienced like relationships. Responsible design takes this perception seriously without equating it with a real human relationship or deliberately reinforcing it where it endangers autonomy and realistic trust.