Ethics of AI and Sensing in Older-Adult Care

Technology can support independent living, medication management, fall detection, remote care, and earlier recognition of health changes. The ethical problem is not whether such systems are good or bad in general. It is whether a specific system gives a specific person enough benefit to justify its intrusion, uncertainty, and loss of control.

Older adults should not be treated as a homogeneous vulnerable population. Age alone does not imply cognitive impairment, low technical literacy, or inability to consent.

Capacity is decision-specific

Decision-making capacity is not a binary property attached permanently to a person.

It can depend on the complexity and consequences of the decision and may fluctuate over time.

A person may be fully capable of deciding whether to use a motion sensor while needing support for a complicated medical decision.

Ethical systems should therefore assess the relevant decision rather than infer incapacity from age.

Consent must describe the actual system

Consent should cover more than device installation.

A person should understand, in accessible form:

• what is measured
• whether audio or video is captured
• how often data are transmitted
• who can see the data
• what algorithms infer
• what actions alerts can trigger
• how long records are kept
• whether data are reused for research or model training

A broad checkbox for "AI monitoring" is not meaningful consent.

Ongoing consent

Monitoring systems can change through software updates, new integrations, or new inferred features.

Consent should therefore be reviewable and revocable.

A person who accepted fall detection did not necessarily agree to behavioral profiling, sleep inference, or sharing data with insurers.

Privacy and proportionality

Continuous sensing creates information about daily routines, visitors, sleep, bathroom use, mobility, and time spent alone.

Even if no camera is present, combinations of sensors can reveal highly intimate patterns.

Data collection should be proportionate to the care objective.

If a door sensor and accelerometer answer the safety question, always-on video may be unjustified.

Data minimization

Collect the least information needed for the intended function.

Where possible, process signals locally and transmit derived events rather than raw streams.

Retention periods should also be justified. Data useful for detecting today's fall do not automatically need indefinite storage.

False alarms and missed events

A monitoring model has both false positives and false negatives.

False alarms can produce anxiety, unnecessary caregiver intervention, or reduced trust.

Missed events can create a false sense of safety.

Evaluation should therefore be tied to operational consequences rather than one aggregate accuracy metric.

Autonomy and safety can conflict

Care systems often optimize safety, while the person may value privacy, independence, or freedom to accept risk.

Those preferences are legitimate.

An ethical design should not silently maximize surveillance because relatives or providers prefer lower perceived risk.

Where capacity is present, the person's own values should remain central.

Family access is not automatically harmless

Giving relatives a dashboard can support care, but it can also create surveillance and conflict.

Access should be role-based and consented to.

A family relationship is not blanket authorization to inspect health and behavioral data.

Human oversight

Human involvement matters when model outputs are uncertain or consequential.

But a nominal human-in-the-loop is insufficient if staff are expected to accept every alert.

Oversight should include authority to question the system, access to contextual information, and a documented escalation path.

Bias and representativeness

Models trained on healthier, wealthier, more digitally connected populations may perform poorly for people with different mobility patterns, disabilities, languages, housing types, or devices.

Subgroup performance and calibration should be evaluated.

Accessibility should also be tested with users who have visual, auditory, motor, or cognitive impairments.

Dignity and language

Terms such as "the elderly" can flatten a diverse population into one category.

Older adults differ in health, preferences, capacity, culture, and desired independence.

Ethical design begins by treating people as participants in system design rather than passive recipients of care technology.

When a representative is needed

If a person lacks capacity for a specific decision, legal and clinical frameworks may permit a representative or substitute decision maker.

Even then, the person's known wishes, values, assent, and objections should be respected as far as possible.

Substitute consent should not become a shortcut around disagreement.

A practical governance checklist

Before deployment, ask:

1. What care objective does the system serve?
2. What is the least intrusive sensing needed?
3. Can the person understand and control the monitoring?
4. Who sees raw data and inferred data?
5. What are the false-alarm and missed-event consequences?
6. How can the person pause or revoke monitoring?
7. How is model performance monitored across relevant subgroups?
8. Who is accountable when the system fails?

Conclusion

Ethical AI in older-adult care is not achieved by adding a consent form to a surveillance system.

It requires proportional sensing, real choice, accessible design, calibrated uncertainty, human accountability, and respect for the person's own tolerance for risk.

Technology should expand agency where possible, not quietly replace it.

References

• World Health Organization. Ethics and Governance of Artificial Intelligence for Health.
• Nuffield Council on Bioethics. Work on assistive technologies, consent, and care.
• European Data Protection Board. Guidance on health data and data protection principles.