RESEARCH
AI + HUMAN INTERACTION
Designing Healthcare AI That People Can Understand and Control
How can we design healthcare AI experiences in which people understand what the technology is doing and retain meaningful control over their information and next steps?
Sandra Osaretin · Lynk · Independent nine-source evidence synthesis
Research synthesis complete · Design exploration in progress
This is a focused, independent synthesis of nine existing sources, not an exhaustive systematic literature review or original participant study. Proposed design responses have not been validated through testing.
Lynk is an independent design concept, not a clinically validated system or medical advice service. Findings come from different countries, populations and healthcare contexts and may not generalize to consumer AI health chat.
01 / Overview
Understanding and control in healthcare AI
As conversational AI becomes more prominent in healthcare, an important design question is emerging: how can people understand and control their interactions with AI-assisted health services?
This investigation explores three connected challenges: understanding consent to health-data sharing, recognizing the involvement and limitations of AI, and knowing when and how to transition to human care. It informs Lynk, an independent healthcare AI product concept I am developing to explore more understandable and human-centred interactions.
A 2026 Santis–Nanos survey found that 71% of Canadian respondents were open or somewhat open to AI reviewing medical tests or scans if doing so reduced specialist wait times. Across proposed applications, respondents generally preferred AI assisting professionals over operating independently [1]. These are reported attitudes, not demonstrated clinical outcomes or understanding of specific AI interfaces.
Canadian privacy guidance emphasizes understandable explanations about the collection, use and disclosure of information [2]. Are users simply being offered controls, or are we designing interactions that help them understand those controls?
02 / Research approach
A focused evidence-to-design synthesis
I reviewed nine sources spanning Canadian public attitudes, meaningful-consent guidance, medical consent comprehension, AI ethics, accessibility, health-information legislation and empirical studies of healthcare AI.
The sources included a national survey, peer-reviewed experimental studies, a systematic review, regulatory guidance and technical standards. Each source was examined for its contribution, limitations, possible design implications and unanswered questions.
03 / What the evidence suggests
CONSENT / Glaser et al. · 2020
Research on medical informed-consent interventions suggests that interactive information and feedback can support comprehension in certain clinical contexts [9]. For digital health products, this creates an opportunity to investigate permission explanations that are specific, timely and understandable.
Evidence from medical-procedure consent cannot establish the effectiveness of AI health-data consent interfaces. The review found heterogeneous measures and substantial risk of bias.
APPROPRIATE RELIANCE / Ayre et al. · 2025
A randomized study of educational materials about ChatGPT health questions found that brief interventions improved knowledge of limitations and reduced reported intentions to use ChatGPT for higher-risk questions [5]. Helping people use AI more appropriately may be a better design objective than simply increasing confidence in it.
Reported intentions are not observed behavior or clinical outcomes. These findings motivate investigation, rather than validate a Lynk feature.
TRANSPARENCY / Zhu et al. · 2026
Research using simulated healthcare AI labels suggests that professional oversight and information about AI performance can influence reported trust and acceptance [4]. But trust alone is not evidence of understanding. An interface should help users identify the status and limitations of AI-generated information without implying review or clinical validation that has not occurred.
The survey experiments involved 340 US patients and simulated cardiovascular AI labels. Findings do not establish how people will use Lynk.
04 / Human support and responsibility
Appropriate reliance, human support and responsibility
More trust is not always the right outcome. A randomized study of educational materials about ChatGPT health questions found that brief interventions improved knowledge of limitations and reduced reported intentions to use ChatGPT for higher-risk questions [5]. Helping people use AI more appropriately may be a better design objective than simply increasing confidence. Reported intentions are not observed behavior or clinical outcomes; these findings motivate investigation rather than validate a Lynk feature.
A study comparing chatbot-assisted and physician-led pre-CT information found that physicians were more effective at alleviating patients’ concerns [6]. An AI interface should not treat the transition to human care as an afterthought. This specific clinical context does not establish the effectiveness of a general health-chat handoff.
Accessibility and privacy require more than attractive interface design. WCAG 2.2 provides testable accessibility criteria [7], while Canadian privacy guidance and Manitoba’s health-information framework inform information handling and individual rights [2,8]. WHO guidance emphasizes autonomy, transparency and accountability [3]. These sources guide design questions; they do not demonstrate Lynk’s accessibility, legal compliance or clinical safety.
05 / The opportunity: Lynk
From evidence to proposed interactions
Make consent understandable
Explore optional health-record connections with clear permission explanations: what information is shared, why it is requested, and how permissions can change. Distinguish revoking access from requesting data deletion.
Make AI involvement explicit
Explore clear distinctions between AI-generated and professionally reviewed information. Communicate limits at the point of use without implying review that has not occurred.
Keep human support accessible
Explore accessible pathways to appropriate human support and patient-controlled summaries for potential handoffs. These are proposed design responses informed by existing evidence, not validated features.
06 / What comes next
A focused interaction prototype
The next phase involves translating these principles into a focused interaction prototype, reviewing accessibility and usability, and documenting design decisions and limitations.
No original participant data has been collected for this synthesis. Findings come from different countries, populations and healthcare contexts. Not all are directly generalizable to consumer AI health chat. Lynk remains an independent design concept.
Consent comprehension: can people explain what they share, why it is requested, and how permissions can change?
AI transparency: can people recognize AI-generated information, professional review and system limits?
Human-care transitions: can people identify appropriate support and control what information follows them?
07 / References
[2] Office of the Privacy Commissioner of Canada. Guidelines for obtaining meaningful consent.
[3] World Health Organization (2021). Ethics and governance of artificial intelligence for health.
[7] World Wide Web Consortium. Web Content Accessibility Guidelines (WCAG) 2.2.
[8] Government of Manitoba. The Personal Health Information Act.