The values of a robotic system do not live only in an ethics statement. They appear in small gestures: whether it waits for an answer, warns before approaching, accepts refusal, refers to a person, and can be stopped by staff without penalty. Those details reveal which incentives govern when care, speed, and appearance compete. Observing them with curiosity, assigning responsibility, and redesigning with care allows the robot to be useful without taking the place of human relationship.

Values become visible through behaviour#

A policy may promise autonomy while an interface selects the most intrusive option by default. A project may speak about dignity while a robot uses an infantilising tone with adults. The contradiction is rarely contained in one major decision; it is distributed across words, timing, movement, and settings. Evaluating principles therefore requires observation of real scenes, not documents alone. What repeatedly happens shows which value has operational priority.

A policy may promise autonomy while an interface selects the most intrusive option by default.

Small gestures are not merely aesthetic. A three-second pause may determine whether someone can respond. A light that is difficult to see can conceal an active camera. A button beyond reach turns the right to stop into an empty promise. The sum of micro-decisions shapes freedom and safety. When a team calls them details, it excludes precisely the point where policy meets the body.

Human behaviours reveal incentives too. If a professional apologises for stopping the robot, alerts are closed without investigation, or praise focuses only on task counts, the system teaches what is expected. Informal rules may contradict the protocol. Honest review examines language, rewards, silences, and shortcuts. It seeks conditions that make one action reasonable and another difficult rather than individual blame.

Curiosity about ordinary use#

Professional curiosity asks about specific actions: what a person does as the robot approaches, which signal they watch, how long they wait, where they request help, and what follows a refusal. Observation requires consent and must not turn care settings into permanent laboratories. Interviews, journeys, incidents, and limited logs can be combined. A successful demonstration cannot replace use during noise, fatigue, shift changes, or irregular connectivity.

Differences deserve attention. A gesture that reassures one person may disturb another. Voice, distance, speed, and appearance are interpreted through experience, culture, disability, and context. There is no universally friendly personality. Systems should offer settings and retain alternatives. Testing includes diverse people, paid when they contribute design work, without expecting anyone to represent a whole group.

The most useful question is often what would have to change for another choice to become possible. If nobody uses the refusal button, that may not indicate acceptance; it could be hidden or consequential. If everyone requests help, instructions may be confusing. Curiosity resists attributing behaviour to character before reviewing the environment. A small observation becomes a testable hypothesis.

Responsibility written into every transition#

Every transition needs an owner: activation, approach, contact, delivery, error, referral, stopping, and restart. If a robot cannot understand a command, the alert recipient and response time must be clear. If a person stops it, automatic restart may be unsafe. After software changes, someone verifies that signals remain accessible. Responsibility does not reside in general intent but in decisions supported by authority, resources, and evidence.

Suppliers control design and updates; an organisation controls purpose, setting, staffing, and oversight; professionals decide within their competence; users retain rights and preferences. Contracts should not confuse these functions or shift all risk to the last link. An accessible responsibility matrix clarifies response. An affected person should not have to locate the technical owner themselves.

Metrics assign responsibility too. A dashboard rewarding speed without displaying human interruption encourages shorter pauses. If interaction time represents wellbeing, insistence may be incentivised. Indicators should be reviewed through their effects: which behaviour they reward, which work they conceal, and who bears the cost. Including choice, understanding, referral, and remedy changes what the team treats as success.

Care without deceptive emotional simulation#

A warm voice may facilitate a task, but it does not make a robot understand or feel. The system should not claim affection, promise secrecy, or encourage emotional dependency where those capabilities do not exist. Transparency can be expressed without coldness: the device identifies itself as a tool, explains its function, and offers human contact. Appearance should not be used to obtain obedience or data a person would otherwise withhold.

Care appears in respect for pace and boundaries. The robot warns before moving, keeps a configurable distance, avoids touch without confirmation, and accepts no. Reminders do not escalate into endless pressure. When uncertainty appears, it refers rather than improvising authority. Design prioritises safe states and redundant help. These modest gestures preserve agency without pretending to a relationship the machine cannot sustain.

A human alternative must be real. A message inviting contact with a professional has little value when nobody is available. Staffing, schedules, and procedures belong to the care product. Automating a task may release relational time, but it may also justify cuts that remove supervision. The contact that will not be reduced and the means of measuring its continuity should be agreed before deployment.

Case 1: an approach that was too efficient#

A hospital robot delivers trays and approaches beds using the shortest route. Speed meets specifications, but a patient with low vision is startled because the audible warning occurs at the same time as movement. Staff begin placing the tray on behalf of the robot to prevent distress, work missing from evaluation. The project retains a high delivery rate and appears successful.

Observation shows that waiting matters more than a few seconds of throughput. The robot stops at a distance, announces its function through visual and audible channels, waits for confirmation, and refers if none arrives. Volume and speed are configurable, and diverse users test operation. Metrics add unassisted delivery, startle, and refusal. Efficiency declines slightly, while workload and risk fall. Care moves from statement to movement.

Case 2: a reminder that would not accept no#

In residential care, a robot provides activity reminders. Without a response, it repeats the message at increasing volume. The function is intended to support participation, but a resident with fatigue feels required to explain each refusal. Her silence is recorded as lack of interaction, so the system insists. The team values response numbers and does not notice that some answers are given simply to stop the machine. The metric turns pressure into engagement.

Redesign adds a straightforward refusal, preferences by activity, and a repetition limit. The robot confirms choice without expressing disappointment and offers human contact only upon request or a defined risk. Residents can review settings with support. The dashboard distinguishes voluntary response, refusal, and referral without penalising no. The small gesture of accepting a refusal returns control and supplies better knowledge about which proposals have real value.

Case 3: a stop button that placed staff at risk#

A transfer robot has an accessible stop control, but every activation generates a report linked to the professional and lowers the shift's productivity indicator. Nobody prohibits stopping, yet the incentive creates doubt. A worker waits for several seconds after hearing an unusual sound to avoid being singled out. No harm occurs, but the near miss shows that human control exists physically while being weakened organisationally.

Management separates justified stops from performance and treats them as safety evidence. The button is tested, the sound investigated, and staff receive explicit authority. Reports analyse conditions without automatic blame. Reporting risk receives recognition without turning safety into competition. Stops initially rise and later fall after technical improvements. Pressing the control ceases to threaten the worker and again protects the person.

Values hidden in default settings#

Initial configuration is an ethical decision because many people never change it. Active cameras, long retention, high volume, or data sharing may favour a supplier over privacy. Defaults should minimise collection, movement, and exposure while allowing understandable adjustment. Higher-risk functions need explicit selection. An update should not restore intrusive options or erase preferences.

Accessibility belongs in the initial state too. Captions, contrast, plain language, compatibility with assistive tools, and adjustable speed should not be hidden in menus. A person can configure with support without permanently giving control away. A profile, where used, is protected and erasable. Visitors and workers receive information where sensors affect them as well. Technical convenience does not automatically prevail over rights.

Default procedures matter as much as software. After failure, a robot may stop and call, continue in reduced mode, or transfer the task. The option should reflect severity and context rather than maximum continuity. Teams rehearse it. If human referral becomes overloaded, deployment is limited. A safe state involves more than stillness; it must prevent abandonment of an essential need.

Participation that changes product and service#

Participation begins before purchase while purpose, tasks, and alternatives can change. Users, families, and workers receive accessible information about limits and data. Sessions provide appropriate formats, schedules, support, and payment. Ceremonial approval is not the goal. Proposals, decisions, and disputes are recorded. A rejected observation receives explanation and remains open to review.

During a pilot, channels should accept uncomfortable gestures that do not meet the threshold of an incident. Questions about which moment caused tension, what explanation was missing, or which action felt intrusive generate early learning. Responses are not used to profile individual emotions without consent. Patterns are analysed and results returned. A person can leave the pilot without losing care.

After deployment, a plural body reviews changes, alerts, refusals, human interventions, workload, and relationship. It has authority to limit functions. Suppliers participate without controlling conclusions. Rotating representation avoids dependence on the same voices. Cooperation becomes infrastructure rather than a one-time consultation.

From an observed gesture to repair#

When a detail causes harm, response begins by attending to the person, stopping risk, and hearing their account. Records are corrected and privacy protected. An apology names the event without reducing it to personal sensitivity. Expenses, lost opportunity, or additional work receive compensation. The person chooses whether to join investigation and receives information about outcomes.

System improvement connects the gesture to its causes. It may require a changed interface, metric, contract, training plan, staffing model, or purpose. Every action has an owner and deadline. Testing occurs under realistic conditions and checks for new barriers. Closing an administrative item does not demonstrate remedy. Evidence comes from later use and is shared accessibly.

Near misses deserve proportionate attention. They should not be dramatised, but neither should they be archived because no injury occurred. Analysis may reveal an early signal and prevent harm. The organisation protects reporters and publishes aggregate learning. Trust grows when a small observation receives a response before becoming a large problem.

Institutional sources for assessing values and risk#

The World Health Organization provides guidance on ethics and governance of artificial intelligence for health, supporting analysis of autonomy, transparency, and accountability. The UNESCO Recommendation on the Ethics of Artificial Intelligence offers a rights- and diversity-centred framework. Both references help turn general principles into questions about practice.

The NIST AI Risk Management Framework organises governance, identification, measurement, and continuous management. The European Commission maintains institutional information about artificial-intelligence regulation. These sources require complementary applicable law, technical tests, and lived experience. No framework replaces observation of everyday system behaviour.

A modest and useful place for the robot#

The smallest gesture matters because it turns abstract values into a bodily experience. Waiting, informing, accepting no, stopping, and referring indicate that relationship is not subordinate to throughput. Approaching without warning, insisting, or hiding sensors shows the opposite. Curiosity detects these signals, responsibility changes incentives, and care verifies the effect.

A useful robot need not look human or perfect. It needs to perform a bounded task, communicate limits, and remain under meaningful control. The organisation retains staff, alternatives, and capacity for repair. When everyday details align with written principles, technology can expand options without appropriating care. Values then leave the document and begin to govern where they truly matter: in each interaction.