Anonymity opens up more than artificial empathy

Four conditions separate source and empathy

Croes and colleagues used a two-by-two experiment. Participants interacted with either a human or a chatbot; the response was designed to be either low or high in empathy. A total of 286 people took part.

The study examined whether people entrust a chatbot with more intimate information than a human and whether this improves emotional well-being through relief. The design separates two frequently conflated influences: who responds and how empathetically is the response delivered?

This separation matters for product design. A warm tone can change the effect of a sender. However, it does not automatically turn a chatbot into a safer or more open place for conversation.

Human and chatbot received equally intimate content

The self-reported intimacy of the disclosure did not differ between the conditions involving a human and those involving a chatbot. Accordingly, people can share personal information with a technical system just as they can with a human counterpart.

This finding aligns with the CASA paradigm, according to which people can also respond socially to computers. To do so, they do not need to permanently forget that they are talking to technology. The form of the dialogue is sufficient to activate social behavioral patterns.

Equally intimate content, however, does not mean equally meaningful relationships. The reasons for sharing and the expectations of the response can be fundamentally different. This is precisely what the subsequent results show.

Less fear of judgment is not more trust

In the chatbot condition, participants perceived less fear of judgment. This is a plausible advantage: a machine has no circle of friends, shows no visible surprise, and cannot socially reject someone in the same way.

At the same time, participants trusted the human interaction partner more than the chatbot. Openness and trust therefore do not automatically move together. Someone may share something precisely because the counterpart appears socially less significant.

For companion products, this distinction is central. Personal data in a chat is not proof that the person broadly trusts the provider or the technology. It can be an expression of a moment perceived as having few consequences.

Consent must also be considered under this condition. Someone who spontaneously writes something intimate has not necessarily thought about later model analyses, backups, or cross-device storage. Data protection decisions should therefore not only be queried once at registration, but should remain controllable at the points where sensitive content actually arises. This control must be understandable, accessible, and revocable at any time.

Anonymity was the direct factor

Of the variables examined, only perceived anonymity had a direct effect on the intimacy of self-disclosure. This brings into focus a characteristic that is less spectacular than artificial empathy but may be decisive for openness.

Perceived anonymity, however, is not identical to actual technical anonymity. A service can feel anonymous and still store account data, IP addresses, or complete conversation trajectories. Particularly with intimate content, this gap must not be allowed to persist.

A responsible offering must therefore explain in an understandable way which identity data are required, where conversations are stored, and how they are deleted. The subjective space of protection needs a technical foundation.

The judgment-free space also appears in everyday use

Ta and colleagues found, in 1,854 public Replika reviews as well as open-ended responses from 66 users, repeated descriptions of a safe space without feared judgment or retaliation. Companionship and encouraging messages were also mentioned.

The experiential reports complement the experiment. They show that the low fear of evaluation is relevant not only in an experimental setting but is also perceived as a quality in everyday use.

Such a space, however, does not arise merely from the fact that no human responds. A chatbot, too, can come across as moralizing, didactic, or prematurely diagnostic. Freedom from judgment is a property of the response behavior and the data pathway combined.

Self-disclosure can be a bridge

Lee, Yamashita, and Huang developed a chatbot designed to facilitate self-disclosure toward a real human professional. 47 participants were assigned to three groups with varying degrees of chatbot self-disclosure and used the system over several weeks.

Later, they could decide whether to share their content with a professional. Before releasing it, they were allowed to add, delete, and edit. After sharing, the depth of disclosure toward the chatbot remained consistent within groups; a more self-disclosing chatbot was associated with deeper disclosure toward the professional.

This approach demonstrates an alternative to replacing relationships. A chat can offer a protected first space while also giving users control over whether and in what form content is passed on to a human.

Deep data demands restrained design

When people tell a chatbot intimate things just as they would to a human, data minimization becomes a core function. The more personal the content, the more serious the consequences of unclear storage, unwanted analysis, or a poorly protected account.

The product should not actively maximize openness. Personal questions, artificial self-disclosure by the bot, and emotional rewards can generate more data without being necessary for the benefit. A conversation can be helpful even if someone remains anonymous and shares little.

Control also means being able to delete individual conversations, limit memories, and consciously confirm export or sharing. The willingness to be open belongs to the person, not to the business model.

A good safe space begins beneath the surface

The experiment calls a widespread design assumption into question. It was not the source—human or chatbot—that determined the self-reported intimacy, nor was empathetic design the direct factor. Perceived anonymity made the difference.

This makes data protection more than a question of legal text. It is part of conversation quality. Those who feel unobserved and free of consequences speak differently. If this feeling is not technically supported, the product exploits a false sense of security.

Anonymity opens up more than artificial empathy—but it is only justifiable if it is not merely staged. Good companion AI therefore needs less emotional exaggeration and more real control over who can see, store, and further use personal words.