Analysis, not certainty: Dialogatlas separates sources, observations, and editorial conclusions. New evidence may change this assessment.

Agreement and mistrust are part of the same finding

Sweeney and colleagues wanted to learn how professionals and experts in the field of mental health think about dialogue-oriented user interfaces. To this end, they conducted an online survey on awareness of and attitudes toward such systems. According to the reported results, 65 percent of participants saw advantages, 74 percent rated the systems as comparatively important, and 79 percent thought they could help clients better manage their own mental health. The publication reports p < 0.01 for each of these proportions.

At the same time, 86 percent were of the opinion that the systems do not adequately understand or show human emotions. This is the central finding of the study because it rejects a widespread but unproductive equation: anyone who acknowledges practical utility does not have to attribute either empathy or the capacity for relationships to the system. Conversely, a lack of emotional understanding does not automatically refute every useful function. The experts were apparently able to distinguish between a tool that makes something easier and a conversational partner that understands a person.

The study measures expectations, not care outcomes

The survey recorded perceptions and attitudes of professionals, not changes in patients or clients. Agreement with the statement that a system could help with self-management therefore does not imply that this actually succeeds under everyday conditions. Likewise, the assessment of a lack of emotional understanding does not prove that every concrete interaction is experienced as cold or unusable. Perceived suitability and demonstrated efficacy are different things.

The provided abstract does not state the number of respondents nor any details on recruitment, composition, or the response scales used. The reported significance values can also only be assessed to a limited extent without this information and without a precise description of the tests. Whether the respondents represented the professional groups of the field remains open on that basis. This does not weaken the documented stance of the participants, but it limits the generalizability of these attitudes. From an editorial perspective, the study is therefore primarily a precise snapshot of professional expectations in 2021, not evidence of efficacy for a product class.

Little use, predominantly satisfactory experiences

Another tension is noteworthy: the surveyed professionals’ personal experience with such systems was limited. Where they had been used, however, the experience was predominantly described as satisfactory. The study also reports a correlation between longer professional experience and a stronger belief that digital conversational systems could support clients in managing their mental health. It does not follow that professional experience causes this belief. Different roles, deployment contexts, or selection processes are conceivable; the source does not clarify this.

The finding at any rate contradicts the convenient assumption that agreement comes primarily from inexperienced technology enthusiasts, while experienced practitioners are fundamentally dismissive. The material rather shows a limited, functional openness. This positive side should not be downplayed: people with professional experience considered support for self-management plausible, and existing usage experiences were mostly satisfactory. At the same time, it remains unclear which systems, tasks, and situations the respondents each had in mind. The umbrella term probably conceals very different applications.

The task is more important than the simulated person

From the results, a role-related reading can plausibly be developed. For narrowly limited tasks, linguistic interaction can be useful without an inner experience underlying it. A system can sort content, formulate questions accessibly, or encourage regular reflection. Whether exactly these functions were meant in the study, however, is not broken down in the abstract. They are therefore an interpretation of the reported self-management finding, not its measured content.

Our editorial position is: a lack of human emotional understanding becomes the decisive problem where a product claims more than a limited auxiliary function. The more strongly it appears as a reliable, empathetic counterpart, the more significant the difference between linguistic presentation and actual understanding becomes. A polite, emotionally appropriate sentence can facilitate use. But it does not prove any perception of the person, any sympathy, or any professional judgment. Good product design should not conceal this difference, but rather make the function so precise that it convinces even without the fiction of a feeling counterpart.

The surrogate question bundles too many different applications

In 2024, Zhang and Wang explicitly ask whether AI could replace psychotherapists. However, their text describes a broad field: predictive analytics, detection, personalized planning, virtual applications, monitoring, and support for clinical work. The provided text does not document its own randomized trial with sample, intervention, and comparison group, although the source is described as a randomized study in the metadata. Reliable results from such a primary study therefore cannot be reported from it. In terms of content, the available material is a broad discussion of existing applications and literature.

The authors refer to preliminary findings on anxiety and depression symptoms, but themselves emphasize small participant groups, lack of long-term observation, and some missing longer-term effects. They call for larger randomized controlled trials and ultimately argue for using AI as a supplement rather than a replacement. In addition, they discuss limitations such as algorithmic biases, data protection issues, lack of genuine empathy, and difficulties in consistently integrating information over longer periods. Thus, this publication also does not answer the big replacement question. Rather, it shows why the question is too coarse: a profession consists of many tasks whose technical replaceability and human significance are not identical.

Linguistic emotionality is not yet emotional understanding

Zhang and Wang cite, among other things, a study in which ChatGPT’s responses to hypothetical scenarios were evaluated using the Levels of Emotional Awareness Scale. According to this, the generated responses were able to reach a level comparable to that of the general population and in part exceeded it. The authors simultaneously mark the decisive limitation: the system recognizes linguistic patterns and produces appropriate formulations, but does not itself experience emotions. What was measured was the quality of responses on a scale, not an inner understanding.

Thus, the finding from the expert survey is not refuted. Both observations can be true simultaneously: a model can produce emotionally appropriate language, while experts deny it human emotional understanding. For product development, this difference is practically relevant. The linguistic performance can be tested and possibly used in a targeted manner. The claim of genuine empathy, on the other hand, would be an attribution that is not substantiated by appropriate texts. Our assessment is therefore neither technology-skeptical nor euphoric: precisely because the linguistic performance can be real, it must be more precisely specified what it consists of.

In 2024, usage is visible, but not unambiguous

Cross and colleagues provide a closer look at actual usage. Two web-based surveys in Australia included 107 people from the general population and 86 mental health professionals. The general attitude was neutral in the population group, and more positive among the professionals. According to the abstract, 28 percent of the general population respondents and 43 percent of the professionals used AI. The reported counts are 30 and 37, respectively; for the first proportion, however, the abstract gives a denominator of 108, although the final sample is given as 107.

The purposes of use differed significantly. Among the 30 users from the general population, 18 used AI for quick support and 14 in a manner they described as a personal therapeutic role. Of the 37 professionals who used AI, 24 used it for research and 20 for reports. This distribution in particular supports a task-related perspective: professional use often focused not on replacing the conversation, but on knowledge- and text-related work. The self-description as a personal therapist in the other group, by contrast, documents an expectation of use; it confirms neither a clinical function nor its quality.

Benefits and harms are reported by the same users

The Australian survey prevents a simple narrative of success or decline. Among actual users, 77 percent of people from the general population and 92 percent of professionals rated AI as useful overall. At the same time, 47 and 51 percent respectively reported concrete harms or concerns. Positive overall experiences and problematic individual aspects are therefore not mutually exclusive. In the open-ended responses, sentiment about the future of AI in mental health care was equally positive and negative. Respondents mentioned expected benefits in accessibility, cost, personalization, and work efficiency, but also concerns about human connection, ethics, privacy, regulation, errors, misuse, and data security.

This study, too, is a survey, not evidence of improved health. It comes from Australia and, based on the abstract, does not allow any conclusion about how representative the participants were of the population or professional groups. A temporal trend compared with 2021 cannot be derived from two differently designed studies either. Nevertheless, it decisively complements Sweeney and colleagues: abstract openness has, at least among some respondents, turned into practical use. Our conclusion from this is deliberately narrow. Psychological conversational systems do not need to be declared human to be useful. They must remain assessable as concrete tools – especially when their language creates closeness that their technical and institutional capabilities do not support.

Sources & further reading