Trust develops over the course of the conversation—not on the landing page

The model originates from collaboration

Bao, Cheng, and de Vreede examine AI in the context of virtual teamwork. There, it can increase efficiency and support collaboration. At the same time, trust becomes a prerequisite for people to appropriately use AI contributions.

The authors propose a holistic theoretical model. Cognitive and emotional perceptions in the interaction process are thus associated with trust in the AI. Certain implementation characteristics can moderate these relationships.

The abstract does not report a completed experimental test of the entire model. Its contribution lies in structuring possible relationships that can be examined in subsequent experiments.

Cognitive perception concerns the ability to work

Cognitively, people assess whether a system appears understandable, competent, consistent, and suitable for the task. In a team context, the question is whether the AI’s contributions genuinely improve collaboration.

In a conversational chat, this level manifests in small moments. Does the response pick up on the right statement? Does the system distinguish between “sleeping little” and “sleeping poorly”? Does it maintain the chosen direction of the conversation across multiple messages?

A friendly interface can announce such capabilities but cannot replace them. Cognitive trust grows when the system repeatedly acts appropriately and corrects errors in a comprehensible way.

Another component is predictability. Users do not need to know every response, but they should understand the basic rules by which the chat operates. If the same conversational request is respected today and ignored tomorrow, uncertainty arises, even if both individual responses look linguistically convincing.

Emotional perception also arises from behavior

People react not only to functional quality. Tone, respect, and the feeling of being taken seriously influence collaboration. The theoretical model therefore also accounts for emotional perceptions.

For mental health conversations, this level is obvious. A response can be factually correct and still come across as condescending or cold. Conversely, it can sound warm and completely misunderstand the content.

Trust requires both: substantive fit and a form that matches the desired conversation. Emotional quality must not mask cognitive errors, and technical precision does not excuse disrespectful conversation management.

Implementation Is More Than the Model

Bao and colleagues emphasize that concrete characteristics of the AI implementation can influence the trust relationship. This prevents the usual reduction to a model name.

A conversational product consists of system instructions, trajectory selection, memories, moderation, interface, data path, and the underlying model. The ability to rephrase a response or delete a conversation also changes the experience.

Two services using the same language model can therefore appear entirely different in terms of trustworthiness. Responsibility lies at the product level, not with an abstract “GPT” or “Claude.”

This also applies to outages and model changes. If a provider switches to a cheaper model in the background, tone and accuracy can shift even though the interface remains the same. Such changes should be documented and retested, especially when users have already built personal conversational routines.

Everyday Experiences Provide Concrete Moments of Trust

Ta and colleagues found in Replika reviews and open-ended responses companionship, encouraging messages, available information, and a space experienced as non-judgmental. Such experiences show how trust can arise in practice.

The sense of safety does not rest solely on an explanation before the first chat. It emerges when users write something personal and do not experience the feared judgment. Every fitting response confirms the expectation; every inappropriate one can damage it.

Because the data come from ratings and open-ended responses, they do not prove general efficacy. However, they do make visible which interaction qualities people themselves describe as relevant.

These responses should not simply end up as praise quotes on a homepage. They can be translated into concrete evaluation criteria: Did an answer feel judgmental? Was the information available at the right moment? Was the user able to control the course of the conversation? In this way, subjective experience becomes verifiable product work.

Research recognizes more than one definition of trust

Bach and colleagues reviewed 23 empirical studies on trust in AI. They found multiple definitions and recommend selecting the appropriate meaning for the respective context rather than pitting abstract terms against one another.

Three areas of influence emerged: socio-ethical conditions, technical and design characteristics, and user characteristics. User characteristics were particularly often the focus, which speaks for participation from development to monitoring.

A mental health chat should therefore separately assess whether people appropriately trust the data protection, the facts, the continuity of the conversation, and the social presentation. A single value cannot capture these levels.

Corrections are a test of trust

What happens after an error is particularly telling. If the chat misunderstands a statement and is corrected, it must incorporate the correction. An apology without a change in behavior is merely linguistic repair.

A changed request for the conversation also tests collaboration. If someone says they want to collect rather than hear solutions, the system should steer less. If it sticks to the old pattern, it shows that the person’s choice carries little weight in the process.

Such trajectories can be tested systematically. They are often more important for real trust than a flawless first response, because they show whether collaboration is actually mutually adjusted.

A test bench should therefore not only contain ideal responses. It needs misunderstandings, unclear statements, disagreement, topic changes, and explicit rejection. Only under these conditions does it become apparent whether the system maintains its role as a conversational partner or falls back into the general helper autopilot.

Trust is an outcome, not a marketing promise

A homepage can explain who operates the service, how data is processed, and that an AI responds. This transparency is necessary. However, it can only create preconditions.

Whether people rely on the chat is decided in the course of the conversation. Cognitive fit, emotional tone, and the concrete implementation take effect with every message. A provider cannot establish this trust by claiming “trustworthy.”

The fair approach is therefore: make verifiable promises, test behavior, document errors openly, and involve users in improvements. Trust then emerges—or it does not.