A figure about robotics changes meaning when its context is restored. A ninety-eight per cent completion rate may conceal human rescues, excluded users, concentrated risk, or uncounted work. Arithmetic accuracy does not guarantee understanding. Knowledge requires knowing how data was produced, humility acknowledges what it cannot represent, and cooperation introduces the perspectives of people who use, maintain, and receive the system. Only then can a robot find its place without dehumanising care.

A correct figure can tell the wrong story#

Robotic projects are frequently summarised through attractive indicators: tasks per hour, accuracy, minutes saved, interactions, or availability. These measures may be correct within their definition and still provide a misleading image. Without denominator, exclusions, setting, and consequences, the number answers a question that is too narrow. Responsible interpretation begins by making visible what was counted and what was left out.

Robotic projects are frequently summarised through attractive indicators: tasks per hour, accuracy, minutes saved, interactions, or availability.

A robot may complete a journey after a person clears the corridor, restarts the system, or finally delivers the object. When the report assigns the whole result to the machine, it erases genuine cooperation. Similarly, a recorded interaction does not demonstrate usefulness or consent. The person may have answered simply to stop repeated prompts. Data needs connection to autonomy, safety, burden, and alternatives.

Purpose determines the metric. A technical measure may be sufficient to compare energy use; a care decision needs human and organisational outcomes. No universal success indicator exists. Before collecting data, teams should define the decision it will inform, the risk it must identify, and who might be harmed by an incomplete reading. Measurement by habit produces full dashboards and little knowledge.

Denominators, exclusions, and withdrawal#

A rate always depends upon its denominator. Reporting that a robot succeeded in ninety of one hundred attempts requires knowledge of who completed the other ten, whether attempts were discarded, and how success was defined. Testing can exclude difficult people, spaces, or situations. If the system is used only under favourable conditions, performance does not represent the whole service. Exclusions must be documented and justified.

Withdrawals are particularly easy to lose. Someone may stop using a robot, choose another channel, or avoid an activity. If the system records only initiated and completed sessions, absence appears to show a lack of demand. Voluntary interviews, observation, and accessible reporting channels can identify reasons. Refusal is not presumed to be inability or irrational fear. It may be a valid evaluation of costs and preferences.

The test population matters. Age, disability, language, body, environment, and technological familiarity can affect sensors and interfaces. A strong average may conceal poor performance for a small group. Disaggregation helps but creates identification risk. Minimum group sizes, protection, consent, and limits are needed. When evidence for a population is insufficient, the gap should be declared and a human alternative retained.

Technical, physical, and relational context#

The same robot behaves differently with changing light, noise, connectivity, furniture, maintenance, and load. Laboratory data cannot automatically transfer to a home, hospital, or residential service. Tests should use realistic scenarios and record conditions. Updates alter performance and require renewed measurement. An old figure may not describe the version interacting with people today.

Organisational context also determines outcomes. Staffing, training, stopping authority, and response time turn an alert into protection or noise. A robot can detect an incident accurately without improving safety if nobody responds. The technical metric must be connected to the human chain. Measuring an isolated component creates a false sense of control.

Relationship is not an immeasurable residue. People can be asked whether they understood the function, could refuse, retained human contact, and felt respected. These experiences should not be reduced to one score. Accounts reveal mechanisms and contradictions. Qualitative evidence does not replace rates; it interprets them. When the two disagree, investigation should follow rather than dismissing the uncomfortable source.

The human work behind efficiency#

Time saved should subtract preparation, cleaning, configuration, training, monitoring, rescue, documentation, and emotional repair. Part of that work may fall on other roles or families. If it does not appear in the budget, efficiency rests on a transfer. A task map identifies who acts before, during, and after each robotic function.

The availability a system requires must also be counted. A professional can remain on call for incidents despite rarely intervening. This capacity has value and cost. Reducing staff because average alerts are low may leave serious events unanswered. Peak periods, nights, and simultaneous failure scenarios complement the average. Resilience cannot be calculated only on quiet days.

Cooperation receives recognition through authorship and decision-making. Workers and users who contribute knowledge need time, payment, and feedback. Their observations should not simply improve someone else's product. They participate in success criteria and limits. The organisation retains accountability even after purchasing technology. A supplier cannot define acceptable human outcomes alone.

Case 1: a ninety-seven per cent delivery rate#

A hospital reports that logistics robots complete ninety-seven per cent of deliveries. The figure includes journeys where a nurse moves trolleys, opens a door, or guides the device through crowds. Attempts cancelled before departure are absent from the denominator. The indicator appears to demonstrate autonomy and savings while ward staff describe interruptions and avoided routes.

Review redefines autonomous delivery and records intervention, cancellation, net time, physical load, and clinical effect. Results are separated by schedule and route. Several stable routes provide benefit, while others consume more work than they save. Use is limited, doors redesigned, and support assigned. The rate declines but becomes useful: it shows where the robot deserves a place and where it does not.

Case 2: accuracy in emotion recognition#

A residential service trials a robot that classifies expressions to adapt conversation. The supplier reports high accuracy from an image dataset. In use, people with atypical movement, low expressiveness, or different cultural backgrounds receive incorrect interpretations. The robot insists on cheering them or alerts staff. The average does not represent these groups, while an emotion label receives authority that its original measurement cannot justify.

The service suspends decisions based on classification and examines the provenance, purpose, and limitations of data. People can disable the function and correct records. Human observation remains, while clinical inference is prohibited. Metrics are examined by group where safe, with realistic tests and paid participation. Restored context shows that the function provides insufficient benefit and it is removed, although other robotic capabilities continue.

Case 3: supposedly released care hours#

A home service claims that an assistant robot releases ten care hours every week. The calculation multiplies tasks by estimated human duration. It omits daily configuration, maintenance, support, and a daughter's work repositioning objects. It also fails to distinguish time the user wished to spend independently from relational time she valued. The number turns different activities into interchangeable units and hides who bears the new burden.

Participatory evaluation records usage diaries, support, reliability, choice, and time quality. The robot increases autonomy in some tasks while creating technical dependency in others. Professional support is introduced and chosen personal assistance retained. The indicator moves to outcomes by task and distributed workload. Technology remains valuable, although the broad promise of ten hours disappears. Contextual information supports honest budgeting.

Designing responsible measurement#

Measurement begins with a theory of change: what problem exists, how the robotic function may improve it, and which adverse effects are plausible. Every step receives evidence and indicators. A baseline supports comparison with a real alternative rather than an ideal scenario. Criteria are agreed before a pilot so that only favourable results cannot be selected afterwards.

Data is minimised. Continuous recording is not required for every objective. Purpose, access, security, retention, and deletion are defined. People receive understandable information and can exercise rights. Consent to service does not automatically include research or model training. Aggregate data protects privacy, but small groups need review for reidentification.

Quality is documented through provenance, dates, versions, variables, missingness, errors, changes, and limitations. Dashboards show notes rather than colour alone. Visualisations are accessible and provide tables or descriptions. A report about care robotics should be understandable to people who bear its consequences. Technical language can be explained without losing precision.

Humility in communicating uncertainty#

Humility does not weaken a decision; it prevents promises that evidence cannot sustain. Ranges are used where appropriate, correlation is distinguished from causation, and small samples are declared. Commercial claims receive independent verification. If a metric cannot capture dignity or relationship, that limit is acknowledged. Uncertainty leads to restricted scope, stronger oversight, or another test.

Negative results are reported internally and, where appropriate, in aggregate. Concealment prevents other teams from learning and biases future decisions. A pilot can conclude that a function should not be deployed and remain valuable. Learning success is separated from product success. This distinction reduces pressure to interpret every figure in favour of investment.

Communication avoids blame. A group using the robot less may not lack interest; it may encounter a different interface or risk. Frequent human intervention does not show staff incompetence; it may reveal device limits. Plausible explanations are compared with evidence. A person should not become an anomaly merely to preserve the average.

Governance and continuous review#

A plural body brings together care, engineering, privacy, accessibility, safety, workers, and users. It reviews purpose, indicators, data, incidents, and updates. It has authority to limit or stop functions. Conflicts of interest are declared, and a supplier provides information without controlling conclusions. Decisions record reasons and dates.

Indicators include tasks, failures, near misses, interventions, response time, workload, refusal, alternatives, privacy, contact, and repair. No one measure summarises the whole system. Trends and cases are examined. More reports may indicate a safer channel. Absence of complaints is interpreted alongside confidence and ease of reporting.

When data reveals harm, repair attends to the person and changes the system. Risk stops, records are corrected, alternatives offered, and costs addressed. Systemic actions have owners and evidence of closure. Measurement then repeats under actual conditions. Closing a document does not demonstrate that the problem has disappeared.

Institutional sources for rigorous interpretation#

The World Health Organization provides guidance on ethics and governance of artificial intelligence for health. The UNESCO Recommendation on the Ethics of Artificial Intelligence supplies principles concerning rights, transparency, diversity, and oversight. Both help connect technical performance with human consequences.

The NIST AI Risk Management Framework structures governance, identification, measurement, and management. The International Labour Organization offers resources concerning the care economy and decent work. These sources need applicable regulation, technical testing, and local experience. No figure replaces responsible deliberation.

Data in the service of care#

Restoring context turns a number into knowledge. It requires questions about who appears, which task was measured, what human work sustained the result, and which consequence remained absent. Humility permits abandonment of an attractive conclusion; cooperation contributes perspectives; knowledge connects evidence to decision. Accuracy retains value but stops pretending to be complete.

A useful robot knows its place when performance is interpreted within the human system that enables it. It does not claim autonomy dependent on invisible rescue or turn averages into moral authority. Data helps bound functions, preserve alternatives, and repair harm. Measurement does not automatically legitimise technology; it supports a decision about where it truly contributes to care and where it should stop.