What do Welsh focus groups want from healthcare AI?
Across four focus groups involving 21 people, participants repeatedly asked for human contact, human oversight and secure governance. The peer-reviewed study offers useful reasons and contrasts between groups, but its small, NHS-connected sample cannot estimate what Wales as a whole believes.
Editorial responsibility: The Impact of AI Editorial Desk · Report a factual concern
At a glance
- 1The researchers held four 90-to-120-minute focus groups with 21 people: six healthcare professionals, six members of the public, five sixth-form students and four retirees.
- 2All groups raised information-security concerns and preferred human contact and human oversight, while attitudes differed over black-box systems and trust in clinicians or private developers.
- 3Many non-clinician participants had NHS, healthcare or digital-technology connections, no migrant or refugee group was recruited, and the sample cannot provide a Welsh population percentage.
Research topic
How healthcare professionals, members of the public, sixth-form students and retirees in Wales describe barriers and facilitators to healthcare AI adoption
The answer: retain a person, a clear responsibility and a secure route for data
Participants in four Welsh focus groups repeatedly described three conditions for acceptable healthcare AI: human contact during care, meaningful human oversight and credible protection of health information. They could see advantages in faster routine work, wider service capacity, consistent analysis and reduced human bias. Yet they did not treat technical accuracy as sufficient. Their questions focused on who remains responsible, whether a person can intervene, which organisation holds the data and whether efficiency will displace relationships that matter to patients.
The study is useful because it records the reasoning behind those views and compares four social groups. It is not a survey estimate. Twenty-one participants cannot establish that a stated percentage of people in Wales holds a view, and the recruitment pattern gives healthcare experience unusual weight. The result should guide questions for service design and larger studies rather than be quoted as a national mandate.[1]
The Impact Brief · Free
Follow the evidence in health & life sciences.
Get a five-minute weekday briefing on what changed, why it matters and where the evidence comes from. Choose the topics you care about.
Who took part and how the discussions were run
The sample comprised six healthcare professionals, six people labelled as members of the public, five sixth-form students and four retirees. The four sessions ran from 25 June to 11 September 2025 and lasted 90 to 120 minutes. Professional and public groups met through Microsoft Teams; student and retiree groups met in person and were recorded on two devices. Participants were not paid.
Some participants had responded to an earlier Wales-wide online survey and volunteered contact details. The team also approached a sixth-form college, a church and Oasis Cardiff in an effort to reach groups underrepresented in the survey. Recruitment through Oasis did not produce a migrant or refugee focus group. Most of the public group worked in non-clinical NHS roles, two worked in digital technology, three retirees had healthcare experience and one regularly used AI in study. Those connections informed the discussion but narrow the claim to a general public sample.[1]
The analysis identified seven linked thematic networks
Researchers used a semi-structured guide covering hopes, worries, trade-offs and proxy consent. They applied inductive thematic analysis, building themes from the transcripts rather than imposing a fixed survey scale, and added contiguity-based analysis to examine how concerns related to one another. Coding was completed in Microsoft Word and iteratively reviewed across the research team. The paper reports seven thematic networks, with five broad themes recurring across the groups.
Those recurring themes concerned confidentiality, security and governance; social consequences such as work and inequality; trust in people and organisations; expected benefits; and consent for a person's own care or care decided by a proxy. The study also records worries about errors, automation bias, clinician deskilling and poor digital infrastructure. Qualitative themes show what was discussed and how ideas connect. They do not rank each concern or show how often it would determine an actual treatment choice.[1]
Human oversight meant more than placing a clinician near the system
Participants wanted a human in the loop, but their concerns suggest that nominal supervision is not enough. A clinician needs time, authority and enough understanding to challenge an output. Several discussions connected black-box reasoning with liability: if neither patient nor professional can understand why a recommendation appeared, responsibility can become blurred. Human contact was valued not only as a safety check but as part of care itself, including explanation, empathy and the ability to notice context not captured in structured data.
The groups were not identical. Healthcare professionals showed greater willingness than some other participants to trust black-box tools, while trust in clinicians, the NHS, academia and private developers varied. Private, profit-seeking companies attracted particular concern. These contrasts caution against a single trust score. Acceptance can depend on the task, organisation, data flow, clinician relationship and consequences of an error.[1]
Consent questions did not create a separate rule for AI
When considering consent, participants generally applied familiar healthcare principles. They expected valid consent for their own care and best-interest decision-making when another person acts as a proxy. The analysis did not identify a distinct additional barrier or facilitator unique to proxy consent. That does not mean consent is solved: the authors note the practical difficulty of explaining an AI system when staff themselves may not know that it is embedded in a workflow or how it produced an output.
For patients, meaningful information should therefore identify the role of the system, what data it uses, whether it changes a decision, who checks it and how to question an error. A generic notice that a service uses AI cannot carry all of that work. Some uses may be too remote from an individual decision to require transaction-by-transaction consent, while high-consequence recommendations may require fuller disclosure. The focus groups raise the issue but do not determine a legal standard.[1]
What NHS teams can take from a small qualitative study
Clinicians and managers can use the findings as a design checklist. Before introducing a tool, they can map where human contact is preserved, who owns the decision, how staff can override the output, which data leave the health service, and how patients obtain correction. They should also measure whether promised time savings create more review work, whether paper and digital records are complete enough for safe use, and whether staff without specialist technical knowledge can recognise failure.
Confidence would rise with purposeful recruitment of Welsh speakers, migrant and refugee communities, people with disabilities, carers and patients who have experienced digital exclusion or diagnostic error. Larger surveys could then estimate prevalence, while observed-choice studies and real service pilots could test whether stated preferences predict behaviour. The work was supported by Health Education and Improvement Wales and NHS Wales Performance and Improvement; the authors declared no commercial or financial conflicts. It is a grounded starting point, not a referendum on healthcare AI.[1]
What this means for people
- Patients want the option to receive explanation and care from a person, not merely a technically supervised automated process.
- Clinicians need time, training and authority to challenge outputs if human oversight is to be meaningful.
- NHS teams can use the themes to design engagement and evaluation, but should not present 21 participants as the voice of Wales.
Global context
The study is rooted in Wales and its NHS institutions. Trust in public healthcare, digital infrastructure, consent law, language and the role of private suppliers differ internationally. The themes are relevant elsewhere, but numerical or policy conclusions should not be exported without local participation and service-level evidence.
What the evidence does not yet show
- The qualitative sample contained 21 participants across four groups and was not designed to represent the Welsh population.
- Many public and retiree participants had NHS, healthcare or digital-technology experience.
- Recruitment did not produce the intended migrant or refugee group, and Welsh-language and other linguistic-minority views were not specifically examined.
- Participants often discussed AI broadly rather than consistently distinguishing machine-learning modalities.
- The study records stated views, not observed choices or outcomes from a deployed healthcare system.
What to watch next
- Research involving Welsh speakers, underserved communities, carers and patients with direct AI-care experience
- Larger representative surveys that estimate how common each concern is
- Service pilots measuring patient understanding, override behaviour, workload and actual safety outcomes
- Clear disclosure and accountability rules when clinicians cannot inspect an embedded AI system
Living evidence record
Impact record IAI-1LEJGXB
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
13 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Frontiers in Digital Health published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 13 October 2026
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Continue the story
Related reporting
Health & Life Sciences
What does it take to scale healthcare AI safely?
An OECD conference report turns input from 142 experts across 34 countries into ten proposed actions for health systems. It is a practical consensus framework for trust, infrastructure and financing—not evidence that scaled AI has improved patient outcomes.
10 min · 2 sources
Health & Life Sciences
Can brain MRI reveal more than BMI?
A peer-reviewed study trained a deep-learning model on 45,702 MRI scans from six cohorts. Brain-derived features tracked BMI and separated several disease groups better than BMI alone—but the disease analysis stayed inside UK Biobank and cannot establish cause, diagnosis or clinical benefit.
9 min · 2 sources
Health & Life Sciences
Can AI make tumour mitosis counts more consistent?
A preprint study paired 13 pathologists' unaided and AI-assisted reviews of 385 tumour slides from three European centres. Agreement rose and counting time fell, but the unreviewed study did not establish which counts were correct, whether diagnoses improved or whether patients benefited.
11 min · 1 source
Reader commentary
Add evidence, experience or a question
No account is required. Reader notes are published after a brief civility, relevance and safety check; disagreement is welcome.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.