Chapter 5
Patient autonomy—the right to make informed decisions about one's own healthcare—is a cornerstone of modern medical ethics. When AI systems influence clinical decisions, informed consent takes on new dimensions. Patients must understand not only the proposed treatment but also the role algorithms play in their care, what data is used, and what the implications of AI involvement mean for their health outcomes and personal information.
Informed consent emerged from recognition that patients have the right to determine what happens to their own bodies. This principle, enshrined in law and medical ethics, requires that before any medical intervention, patients receive adequate information to make a voluntary and intelligent choice. The elements of informed consent—disclosure, comprehension, voluntariness, competence, and consent—were developed for human-delivered care. AI challenges each of these elements in new ways.
Traditional informed consent focuses on disclosing the nature of the proposed treatment, its risks and benefits, alternatives, and the consequences of refusing treatment. When AI is involved, the scope of disclosure expands: What role does AI play? How was it trained? What are its limitations? Can the patient opt out? How is their data used? The challenge is providing sufficient information without overwhelming patients or undermining their trust in the care they receive.
When AI participates in clinical care, patients may need information beyond traditional consent disclosures. The challenge is determining what must be disclosed, how to present complex technical concepts accessibly, and how to integrate AI disclosure into clinical workflows without creating consent fatigue or undermining trust.
| Disclosure Element | Traditional Consent | AI-Enhanced Consent | Rationale |
|---|---|---|---|
| Nature of Intervention | What procedure will be performed | AI is involved in diagnosis/treatment decision | Patient right to know who/what influences care |
| Risks | Medical complications, side effects | AI error rates, performance limitations | AI-specific risks may affect patient decisions |
| Benefits | Expected health improvements | AI accuracy advantages, consistency | Patients deserve to know potential AI benefits |
| Alternatives | Other treatment options | Human-only assessment available | Autonomy requires option to decline AI |
| Data Use | Limited data disclosure | How patient data is processed by AI; storage; sharing | Privacy rights; secondary use implications |
There is no consensus on exactly what AI-related information must be disclosed for consent to be valid. Different stakeholders propose varying standards, from minimal disclosure (AI is used) to extensive disclosure (algorithm specifics, training data sources, performance metrics). The appropriate level likely depends on the AI's role, the stakes of the decision, and the patient's preferences.
| Standard | Required Disclosure | Proponents |
|---|---|---|
| Minimal | AI is used in care; general purpose | Efficiency advocates; some institutions |
| Material Risk | AI involvement when it materially affects outcomes | Traditional malpractice standard applied to AI |
| Reasonable Patient | What a reasonable patient would want to know about AI | Patient advocacy groups |
| Full Transparency | Comprehensive AI details including training, performance, limitations | AI ethics scholars; some regulators |
A fundamental question in AI ethics is whether patients have a right to know when AI influences their care—even if that knowledge might cause anxiety or change their behavior in ways that affect outcomes. This tension between transparency and potential psychological impact mirrors debates in other disclosure contexts, such as informing patients of incidental findings.
Healthcare systems must find approaches that respect patient autonomy while avoiding consent overload. Potential strategies include:
If AI disclosure is required, should patients have the right to refuse AI-assisted care? The right to opt out raises practical and ethical questions: Is human-only assessment always available? What if AI-assisted care is demonstrably better? How do opt-out requests affect the patient-physician relationship?
| Opt-Out Consideration | Patient Perspective | Institution Perspective | Ethical Analysis |
|---|---|---|---|
| Availability | Want meaningful alternative | May be impractical for integrated AI | Autonomy requires genuine choice |
| Quality Differential | Concern about inferior non-AI care | Liability for worse outcomes without AI | Informed choice includes quality information |
| Cost Implications | May face higher costs for human-only | Human-only assessment more expensive | Equity concerns if opt-out is costly |
| Documentation | Want opt-out respected and recorded | Need clear documentation for liability | Process must ensure genuine voluntariness |
| Workflow Impact | Expect opt-out to be honored smoothly | Disrupts standardized AI workflows | Systems must accommodate patient choice |
There may be situations where opt-out rights are constrained. If AI is deeply integrated into standard diagnostic equipment, opting out may be impractical. In emergencies, obtaining specific AI consent may be impossible. Resource constraints may limit availability of human-only alternatives. These limitations should be disclosed and justified.
Beyond consent for AI-assisted care, patients may need to consent to the use of their data for AI training and development. This secondary use of health data raises distinct ethical considerations about privacy, benefit-sharing, and control over personal information.
Traditional one-time consent may be inadequate for ongoing AI development. Dynamic consent models allow patients to make ongoing choices about how their data is used, updated as new AI applications emerge. While more respectful of autonomy, these models are more complex to implement and may create data availability challenges for researchers.
| Consent Model | Approach | Advantages | Challenges |
|---|---|---|---|
| Broad Consent | General permission for future research | Simple; enables research | Limited patient control; may not cover novel uses |
| Specific Consent | Separate consent for each AI project | Precise patient control | Impractical for large datasets; slows research |
| Tiered Consent | Categories of uses with patient choices | Balances control and practicality | Categories may not match actual uses |
| Dynamic Consent | Ongoing patient engagement and choices | Continuous autonomy; updated preferences | Technology requirements; engagement burden |
| Meta-Consent | Consent to governance process rather than specific uses | Scalable; trusts oversight bodies | Indirect patient control; governance quality critical |
Special considerations apply to obtaining consent from vulnerable populations who may face barriers to understanding AI or expressing preferences. Children, elderly patients, those with cognitive impairment, and individuals with limited health literacy require adapted consent approaches.
Healthcare institutions must develop practical approaches to obtaining AI-related consent that respect patient autonomy while remaining feasible within clinical operations. This requires balancing thoroughness with efficiency.
Clear documentation of AI-related consent protects both patients and institutions. Documentation should record what information was provided, patient questions and concerns, the consent decision, and any opt-out requests.