2.1 Participant Recruitment and Selection

We recruited 47 participants across eight groups, via an independent professional recruitment company (Taverner Research), using both social media advertising and random digit dialling. All participants were Australian residents over the age of 18 years.

Potential participants were excluded if they did not have a device with a webcam and internet connection that allowed them to participate via Zoom or if they had worked in health care in the previous 5 years, to avoid knowledge disparities preventing those with less healthcare knowledge from contributing to the groups. Participants who were not capable of participating in the group without the aid of an English language interpreter were also excluded, as we were not able to resource interpreters for multiple languages in a way that would allow natural conversation flow for all participants.

Three separate participant profiles were recruited for the groups (Table 1). Profiles were based on results from a previous analysis conducted by the authors investigating which Australian subgroups were more or less likely to be supportive of the development of HCAI [4]. Profiles were designed to ensure that the project engaged with individuals with a diversity of opinions on HCAI.

For Groups 1, 4 and 5, we recruited Australians over the age of 45, ensuring a substantial proportion identified as having a chronic health condition or disability (N = 11; 65%). For Groups 2, 3 and 6 we recruited Australians aged 45 or under. For Groups 7 and 8 we recruited 12 Australians with computer science or programming experience. Participants with computer science or programming experience were excluded from participating in Groups 1–6 to prevent knowledge disparities within groups.

Potential participants were contacted by Taverner to ascertain that they met the inclusion criteria and were available to attend one of the groups. EF emailed a copy of the participant information statement to all potential participants, before arranging a phone call to answer any questions and to request participants' verbal consent. After providing consent, participants who were not confident using Zoom were supported to download and test the programme.

2.2 Dialogue Group Design

All eight dialogue groups were conducted via Zoom, with EF moderating, and a two-person co-moderation team comprised of a research assistant and either YSJA or SMC. Groups took at least 1 h, and up to 2 h. All data were collected in June 2022. Upon joining the Zoom call, participants and researchers participated in a brief icebreaker before the discussion commenced. The discussion was separated into two parts: (1) an initial discussion about AI in general and then about AI in health care and (2) a scenario-based discussion about a symptom checker tool that uses AI.

In Part 1 of the discussion, Zoom's poll feature was used to present questions to the participants, which were then discussed by the group. Between poll questions, participants were presented with some examples of AI currently in use or in development. These examples were to ensure that participants had a baseline understanding of a range of AI applications, and the extent to which they are used in everyday life (Table 2). We intentionally avoided providing detailed case studies or scenarios in the first part of the groups to encourage discussion about the promises and potential harms of HCAI in general, rather than judgements on individual technologies.

In Part 2 of the discussion, we asked participants to reflect on a time where they had concerning and unfamiliar symptoms. Two participants were invited to share their stories if they were comfortable doing so. Participants who shared their stories were asked to reflect on what worked well about the current healthcare system, and what did not work so well. This interlude encouraged participants consider their own personal circumstances and whether they felt that AI would benefit them. Following a brief discussion, participants were shown a video about an AI symptom checker (Supporting Information: File 1).

After the video, participants were asked about their initial thoughts and reactions. Throughout the discussion, the moderator used additional prompts to encourage participants to think about different aspects of utilising a symptom checker app. Example prompts are in Box 1 and the full script is in Supporting Information: File 1; prompts were used to redirect the conversation when required, and therefore not all prompts were used in all groups. Following the groups, participants received an AU$100 gift card to thank them for their time.

A professional transcription company transcribed the audio-recorded data verbatim. One audio-recording was lost in a file transfer error. Seven final transcripts and the researcher's comprehensive notes from the lost recording were used in the final analysis.

2.3 Analysis

We used Braun and Clarke's [23] reflexive thematic analysis (RTA) to analyse the qualitative data. RTA is a qualitative analysis method which emphasises the researcher's role in the process of theme development [23]. RTA requires the researcher to move iteratively between analysis stages, and remain reflexive about how their methodological and analytical decisions could affect the final results [24].

The first author coded and analysed the data. Analysis followed the process outlined by Braun and Clarke [24]. EF familiarised themselves with the data and took down observations in an analysis journal. Then, EF initially systematically coded the data in Obsidian (https://obsidian.md/).

Instead of codebook or inter-rater processes, Braun and Clarke [24] recommend that one researcher codes and themes the data, engaging with co-authors to reflect on their positionality and assumptions. In regular supervision meetings, EF and SMC discussed EF's positionality, process, interpretive decisions, and theme ideas. Whilst SMC was not involved in the coding process, they co-moderated several of the groups, and conducted several other dialogue group studies on public views on AI, so were well-informed enough to challenge EF's decisions and discuss themes. Through these discussions, EF refined the themes and drafted theme structures for discussion with all co-authors. This process resulted in further refinements to generate the final themes. An example of how quotes were coded and eventually formed into themes is provided in Figure 1. EF analysed the categorical data generated from the polls descriptively, using MS Excel.

3.1 RQ1: How do People Describe the Reasons for Their Responses to 5-Point Semantic Scales About Their Support for AI in Health Care?

During and after the groups, participants were asked three poll questions. Figure 2 shows participant responses, aggregated across the eight groups. Both during and after the groups, most participants reported that they were either ‘somewhat’ or ‘strongly’ supportive of AI and its introduction in health care. However, there was a tendency for participants to move toward the ‘somewhat support’ category by the conclusion of the groups, both from participants initially more supportive and more oppositional to AI.

3.2 RQ2: Under What Conditions do People Support the Use of AI in Health Care?

In the second part of the groups, we asked participants to consider a time when they or a loved one had symptoms and were unsure about the cause or diagnosis. We showed the participants a symptom checker tool as a case study and asked them to consider the circumstances under which they would support this use of HCAI. In our thematic analysis of the transcripts, we identified three main conditions under which the participants supported the use of HCAI.

3.3 RQ3: Under What Conditions do People Find it Unacceptable to Automate a Healthcare Task With AI?

The participants also strongly expressed two conditions under which they felt that it would be unacceptable to automate healthcare tasks.

4.1 Participants Were Conditionally Supportive of HCAI, Particularly After Discussing Their Views With Their Peers

By the conclusion of the groups, most participants said that they ‘somewhat support’ the use of AI in health care. In being somewhat supportive of the use of AI in health care, participants recognised that AI could lead to meaningful benefits in health care, but that there were certain use cases that they would not support. Qualitative studies with publics typically have similar findings, that publics are mostly open to the idea of AI being used in health care, but are cautious and uncertain about its potential risks [11, 12]. We found that very few participants in these dialogue groups diverted from this conditionally optimistic view. A minority of participants were less supportive than the majority because they were more cautious about whether HCAI would be implemented with the necessary controls to prevent patient harms and a loss in quality of care. A separate minority were more supportive of AI than the majority because they felt strongly that AI would be implemented in ways that were, on balance, beneficial to patients.

When asked to compare whether they supported the use of HCAI more or less than they supported the use of AI in general, participants were divided. Overall, the participants saw health care as an opportunity to have a significant positive impact, but they also identified that risks associated with bias, over-automation, and privacy were more significant in health care than in other fields where AI might be used.

In regard to their level of support for HCAI, several participants moved toward the ‘somewhat support’ response category throughout the groups, from initial positions that were more supportive or more oppositional. The change in participants' views throughout the group indicates that, when exposed to information and alternative viewpoints on AI, people's views on HCAI are likely to shift. In these groups, we intentionally gave participants an opportunity to learn about HCAI and discuss their views with their fellow group participants. We expected that many participants may not have had an opportunity to learn about HCAI before participating in our study, and we wanted to ensure as much as possible that participants' views were based on correct and realistic information [25]. Given that the process of participating in the group did shift many participants' views, it may be useful for researchers conducting future studies on public views on HCAI to consider incorporating information and discussion into the research process to develop a common understanding of HCAI amongst participants.

4.2 Participants Felt That HCAI Innovations Should Benefit Healthcare Consumers, be Controlled to Prevent Harms, and Enable the Delivery of Patient-Centred Care

In our thematic analysis, we identified the conditions under which participants typically found HCAI acceptable. First, the participants wanted to see HCAI implemented in ways that delivered benefits to healthcare consumers, particularly through targeting structural problems in Australian healthcare systems. They were concerned about whose interests were reflected in innovation agendas, and they preferred to see HCAI targeted toward benefiting healthcare consumers. Engaging publics through democratic consultation and codesign processes can encourage alignment between innovations and the healthcare issues that people want addressed [26, 27]. Public engagement processes on HCAI can be challenging, as publics may not have previously considered their potential benefits and risks [3]. However, processes which are designed to engage participants in these challenging problems, and which integrate learning and capacity-building exercises into the research methodology, have generated nuanced recommendations on HCAI related issues (e.g., [28]). Similar processes could be used for involving publics in agenda-setting to identify areas where HCAI implementation could be beneficial.

Second, we found that participants wanted to know that people were protected from any HCAI-related harms via regulatory or other control mechanisms. When asked, the participants were uncertain about exactly what these controls should be. However, they often expressed a sense of vulnerability and a need for assurance that they would be protected from harms. During the limited discussion group time, we did not incorporate any information or learning with the participants about the potential harms of HCAI or about potential regulatory solutions. A community jury with Australians on AI in health care, which incorporated more extensive learning about HCAI's benefits and harms, found that participants recommended a series of mechanisms to protect patients from harms [28]. These recommendations included mandatory peer-assessed evaluations of the safety of HCAI tools, independent bodies to oversee these evaluations, and mandatory transparency reporting of the HCAI tools' efficacy. Incorporating recommendations such as these into regulatory systems may both prevent harms to patients and increase public confidence in Australian regulatory systems.

Thirdly, the participants wanted HCAI to give people access to more epistemic resources about their health. The participants described the process of getting diagnoses and information about their health as drawn-out and challenging. They saw AI as a potential opportunity to make personalised information more accessible to patients. In suggesting that AI should provide patient access to more epistemic resources, the participants envisaged a healthcare system based on a model of patient-centred care, where patient autonomy is prioritised by equipping patients with the necessary knowledge about their health to make more informed decisions [29]. While the potential for AI to support patient-centred care has been discussed in the AI ethics literature [30], these discussions rarely focus on AI's potential to enhance patients' autonomy via enabling access to more epistemic resources. Our results indicate that this access to additional information may be one of people's main conditions for supporting the use of AI in health care.

Although the case study presented to the participants showed a consumer-facing symptom checker app, the participants typically envisaged AI being more useful in GP clinics or hospitals. They wanted these tools to provide patient access to more information within the clinical setting, but without requiring appointments with multiple different clinical specialists. We speculate that the participants' preference for HCAI to remain within the clinical setting, rather than to be delivered direct-to-consumer via apps, is likely a reflection of their preference for patient-physician interaction and overarching physician authority to remain unaffected by the introduction of AI.

4.3 Participants Valued Human Contributions to Healthcare Systems Beyond What HCAI can Provide

The participants identified two main instances where they did not want AI to automate healthcare tasks. Participants did not want their interactions with healthcare workers to be automated by AI, and they did not want any decision-making about their health to be automated by AI. Similar findings are common in the literature on public views on HCAI [10, 15, 31-33], where studies have found that publics often express concerns that HCAI systems may not be as effective or as empathetic as human physicians.

Our study findings indicate that participants saw human healthcare workers as valuable for more than just their efficiency and their ability to be empathetic. Participants described instances where they wanted a person physically present to examine a loved one's symptoms, or where a healthcare worker acted in a patient's best interests to make a difficult decision about their health. Their descriptions of why healthcare workers are valuable appealed to very fundamental traits of being human, including the presence and embodiment of a human person, and the capacity to make decisions with practical wisdom [34].

A number of studies have recommended that HCAI tools should be developed to appear more empathetic to more closely resemble care delivered by humans [31-33]. Our findings, however, suggest that if AI were to emulate human-like empathy, this would not be sufficient for people to support the replacement of human healthcare workers in their decision-making capacity or in physician–patient interactions. Instead, the participants felt that HCAI should be implemented as an epistemic tool to give patients and their physicians access to more information about their health.

4.4 How Can HCAI be Developed and Implemented in Australia to Meet Public Needs?

A significant effort is currently underway in Australia to develop more effective mechanisms for regulating AI. A whitepaper released in 2023 by the Australian Government Department of Industry, Science and Resources described ‘low levels of public trust’ in AI as a factor impeding the adoption of AI technologies in Australia [35].

Our findings suggest that Australians are not necessarily distrustful of AI. The participants were mostly open to the use of HCAI, but they were often unsure whether enough controls were in place to prevent unintended harms. There is an opportunity for more public engagement to align Australia's AI policy with public interests, which may help people feel more confident in HCAI tools. These engagement approaches may be more effective when they provide participants with opportunities to learn about and discuss AI's benefits and risks, as they may not have had an opportunity to learn about AI before participating.

Overwhelmingly, we found that Australians prefer a healthcare system where important decisions are made by human healthcare workers, and where people have an opportunity to interact with their physicians. It is unlikely that publics will support the use of HCAI tools that displace physicians. Instead, our findings indicate that people hope HCAI tools will increase their access to diagnoses and information about their health without taking away human healthcare workers' decision-making authority or opportunities for patient-physician interaction. To implement and use HCAI in ways that publics support, these key conditions should be considered.

4.5 Limitations

Our study has some limitations. First, whilst we were successful in recruiting participants diversely across a range of indicators such as age, working status, and country of birth, our final sample lacked sufficient representation from some groups. We excluded those who could not participate without the aid of an English language interpreter. Future research should aim to examine the views of Australian residents who do not speak English, as their concerns about AI may differ to the English-speaking population. The sample was disproportionately comprised of Australians with a bachelor degree or higher (89%; N = 42), compared to the Australian population aged 15–74 (36%) [36]. Given evidence that education levels are often associated with differing views on AI [3], this may limit the generalisability of our findings to the broader population. In addition, our recruitment company did not allow prospective participants to identify with an alternative gender category, which may have inadvertently excluded gender diverse Australians from participating. Future research on HCAI should aim for better representation from underrepresented groups to address the views of gender diverse Australians and Australians without higher education degrees (e.g., [28]).

The symptom checker case study used in the latter part of these groups is one example of many ways that AI could be used in health care. Since this symptom checker tool was a conversational agent marketed to the public, this likely prompted participants to consider the benefits and risks of this type of AI tool. Using multiple case studies may have elicited discussion of different conditions; however we made the decision to focus on one case study to facilitate meaningful and in-depth discussion.