Artificial Intelligence in Chinese Elderly Care: From Prediction to Responsible Practice

An elderly-care worker notices that an older resident is walking less, sleeping irregularly and eating more slowly than usual. A traditional system depends on somebody recognising that these separate changes form a pattern. An artificial intelligence system may be able to identify the combination earlier, compare it with previous behaviour and prompt a review before deterioration becomes obvious. That possibility captures both the promise and the challenge of AI in long-term care.

Across the China Ageing, Long-Term Care & Community Support Knowledge Hub, technology has increasingly emerged as part of the response to rapid population ageing, workforce pressure and the need to support more people at home and in the community. China is now moving beyond digital records, telemedicine and connected devices towards systems that can classify, predict, recommend and increasingly interact with older people directly.

The policy environment is moving quickly. National development of smart health and elderly care has already encouraged artificial intelligence, wearables, rehabilitation technology and service robotics. In 2025, work on digital age-friendly products and intelligent elderly-care robots expanded further, including a 2025–2027 programme testing intelligent service robots in real care settings. Local initiatives are moving faster still: Beijing’s Haidian District, for example, has adopted a 2026–2028 “AI + elderly care” plan covering smart communities, intelligent elderly-care institutions, demand analysis, health monitoring and care scenarios. These developments are important, but they should not be mistaken for uniform national deployment. China is entering an implementation phase in which AI is expanding rapidly while many applications remain pilots, local innovations or emerging commercial models.

Artificial intelligence changes elderly care when it influences a decision

The term “AI elderly care” can describe very different technologies.

Some applications are comparatively simple: analysing sensor information, automating scheduling or identifying unusual patterns in service records. Others involve large language models, computer vision, predictive analytics, intelligent rehabilitation systems or robots capable of more complex interaction with people and environments.

The distinction between ordinary digitalisation and artificial intelligence matters.

A digital care record stores information. An AI-supported system may analyse that record and predict which residents are at greater risk of falling. A traditional monitoring device records vital signs. An intelligent system may combine those readings with previous history to determine which changes deserve professional attention.

The operational significance therefore begins when technology influences prioritisation, interpretation or action.

That is also where governance becomes more important.

A system that merely stores information can still create privacy and quality risks. A system that recommends which older person needs urgent attention can influence safety directly.

The broader field of artificial intelligence and automation in care therefore needs to be understood as a decision-support issue rather than only a technology-development issue.

China is developing AI elderly care through both national policy and local experimentation

China’s approach combines industrial policy, digital-health development, elderly-care reform and local innovation.

National programmes have promoted smart health and elderly-care products, intelligent rehabilitation, wearables, digital services and service robots. Work on digital accessibility has also emphasised that technological innovation needs to remain usable by older people rather than creating new barriers.

The development of intelligent elderly-care robots illustrates how policy is moving from demonstration towards practical testing.

The 2025–2027 pilot programme for intelligent elderly-care service robots focuses on real application scenarios and includes objectives such as improving older people’s quality of life, reducing family-care pressure and supplementing workforce capacity in institutions and communities.

That wording is significant.

The robots are being tested because their operational value still needs to be demonstrated. The existence of the pilot should therefore not be interpreted as evidence that robotic care is already routine across China.

Local government initiatives add another layer. Areas with strong technology industries can experiment with AI faster than many other localities. Haidian District’s 2026–2028 programme, for example, brings together universities, technology firms, elderly-care organisations and public authorities while exploring AI-supported demand analysis, intelligent institutions, community services and care technologies.

This local experimentation can generate valuable learning.

It also means that AI adoption will initially be uneven. A technologically advanced Beijing district may be able to test systems that would be difficult to operate in a rural county with weaker digital infrastructure, fewer technical staff and smaller provider markets.

The national governance task is therefore not to make every locality adopt identical technology. It is to distinguish useful innovation from technology that works only under unusually favourable conditions.

Prediction may become one of AI’s most valuable roles in long-term care

Long-term care generates repeated observations over time.

Changes in mobility, appetite, sleep, continence, mood, medication, falls and care intensity may each appear small individually. Together they can indicate that somebody is becoming less stable.

AI can potentially identify those combinations earlier than a system relying entirely on individual records being reviewed manually.

This creates opportunities around falls, pressure injury, hospitalisation, nutritional risk and functional deterioration.

The attraction is understandable.

China’s elderly-care workforce cannot increase indefinitely in proportion to the number of older people requiring support. A system capable of helping workers distinguish routine information from meaningful deterioration could allow scarce professional attention to be focused more effectively.

But prediction is only useful if the system knows what should happen next.

An algorithm that classifies somebody as “high risk” without defining who reviews the person, what evidence is checked and what response is available merely generates another alert.

This is why AI needs to connect with decision-making and escalation.

The value lies in improving the response pathway, not in producing a more sophisticated risk score.

Operational scenario: an algorithm identifies deterioration that staff have seen only in fragments

A municipal elderly-care institution introduces an AI-supported risk system using information already collected through routine care, including mobility, meals, sleep, falls and changes in assistance.

An 87-year-old resident does not trigger any single urgent concern. She has eaten slightly less over several days, requested more help getting out of bed and become less active in the afternoon.

Different workers have recorded each change, but no individual entry appears serious.

The system identifies the combined pattern as a significant departure from the resident’s usual baseline and prompts professional review.

A nurse assesses the resident and identifies dehydration and signs of an emerging urinary infection. Medical review is arranged before severe deterioration occurs.

The result appears to support the AI system.

Governance nevertheless remains necessary. The institution reviews whether similar alerts are accurate across different residents, whether false positives are creating unnecessary workload and whether staff are beginning to rely on the system so heavily that they stop escalating concerns unless the algorithm agrees.

The technology is therefore treated as a second layer of pattern recognition, not as permission to ignore professional observation.

That distinction is fundamental to responsible AI in elderly care.

AI needs a reliable baseline because older people do not all behave alike

Predictive systems work poorly when they assume that one definition of “normal” applies to every older person.

An 80-year-old who normally walks several kilometres each day has a very different functional baseline from somebody of the same age who uses a wheelchair. A resident with longstanding sleep disruption should not necessarily trigger the same concern as somebody whose sleep changes suddenly.

Personalised baselines can therefore be more useful than generic thresholds.

AI may be particularly valuable where it learns how an individual usually functions and identifies significant deviation.

But personalisation creates another requirement: enough high-quality longitudinal information must exist to establish that baseline.

The wider principle of data quality and performance metrics becomes central because poor input produces unreliable inference.

If frontline records are incomplete, sensors malfunction or different workers describe the same behaviour differently, the algorithm can produce a confident-looking conclusion from weak evidence.

Artificial intelligence does not remove the need for accurate care records.

It increases the consequences of inaccurate ones.

AI can support functional assessment without replacing professional assessment

Functional assessment is becoming more important across China as long-term care insurance expands and elderly-care services become more needs based.

AI may assist this process in several ways.

Computer vision may help analyse movement. Digital tools can structure information about activities of daily living. Longitudinal records may identify changes that suggest reassessment is required. Voice or interaction patterns could eventually contribute to identifying cognitive or functional change.

These possibilities may make assessment more consistent and help systems identify people whose needs have changed between formal reviews.

However, functional assessment is not simply a prediction task.

The same physical limitation can have very different implications depending on housing, family support, cognition, equipment and personal preference.

A person who struggles with stairs in a fifth-floor apartment may require a different response from somebody with identical mobility living in an accessible building.

Professional assessment therefore needs context that an algorithm may not fully capture.

AI can strengthen assessment by highlighting evidence and change.

It should not reduce a multidimensional understanding of the person to a risk category.

Care planning creates a harder test for generative AI

Large language models can summarise records, draft care-plan sections, identify gaps and translate complex information into more accessible language.

This could reduce administrative burden in elderly-care organisations.

Managers and frontline workers currently spend significant time recording and reorganising information. AI may allow some of that effort to shift back towards direct care.

The opportunity is substantial, but so is the danger of plausible error.

A generative system can produce language that sounds professional even when an inference is wrong or unsupported.

An AI-generated care plan may therefore appear more coherent than the underlying evidence justifies.

Human review needs to remain mandatory where generated content affects care decisions.

The wider field of digital care planning is relevant because automation should improve the quality and usability of the record rather than simply accelerate documentation.

The appropriate division of labour is increasingly clear: AI can organise, summarise and suggest; accountable professionals need to verify, decide and personalise.

Administrative AI may deliver faster gains than autonomous care

Some of the most valuable uses of AI in elderly care may initially occur away from direct care decisions.

Scheduling can be optimised around geography and worker availability. Documentation can be summarised. Repetitive administrative queries can be automated. Demand patterns can support workforce planning. Service platforms can analyse requests and help direct people towards appropriate resources.

These applications may improve productivity with less direct risk than allowing an algorithm to determine clinical or care eligibility decisions.

The broader principle of automation, workflow and operational productivity is therefore particularly relevant.

China’s ageing challenge makes this distinction strategically important.

The question is not only whether AI can replace a human task.

It is which human tasks should be protected.

Automating repetitive data entry may give an elderly-care worker more time to observe, converse and support independence. Automating a complex judgement about whether somebody requires more care may create risk if contextual factors are missed.

Productivity should therefore be pursued where technology removes low-value work without removing necessary human judgement.

AI-supported workforce planning could help local systems anticipate pressure

Artificial intelligence can also operate above the individual care-plan level.

Provincial, municipal and county systems increasingly hold information about population ageing, long-term care insurance, service utilisation, provider supply and workforce capacity.

Analytical models could help identify where demand is likely to rise fastest or where provider capacity may become insufficient.

A county experiencing rapid growth in the number of older people with substantial functional impairment may need additional home-care capacity before current services become visibly overwhelmed.

Similarly, repeated workforce vacancies in particular neighbourhoods may reveal that existing service geography is unsustainable.

AI could support these planning decisions by modelling several variables together.

Organisations examining comparable scenarios can use the Digital Twin Scenario Modeller to explore how demand, workforce and service stability may interact. It is not a China-specific AI planning system, but the scenario-based discipline is relevant.

The important governance point is that prediction should inform planning rather than create an illusion of certainty.

Demographic assumptions, workforce behaviour and provider responses can change. Models should therefore support judgement, not disguise uncertainty behind a numerical forecast.

AI-supported care planning needs stronger safeguards than AI-supported administration

The distinction between low-risk and high-risk automation should shape how elderly-care organisations deploy artificial intelligence.

Using AI to summarise meeting notes, identify duplicated records or prepare a draft rota has different consequences from allowing it to recommend whether somebody needs more care, whether a resident is at increased risk of falling or whether a person should be prioritised for review.

The more directly an AI system influences health, safety, eligibility or restriction, the greater the need for professional verification.

This creates a practical hierarchy of use.

Administrative automation may be suitable for relatively broad deployment where accuracy can be checked easily. Decision-support tools require stronger validation, defined professional ownership and clear routes for overriding the output. Fully autonomous decisions affecting care should face the highest threshold because the consequences of error are greater.

This distinction can help providers avoid two opposite mistakes: treating every use of AI as inherently dangerous, or treating every application as equally suitable for rapid automation.

Human oversight needs to be designed into the workflow

Statements that AI will remain “human supervised” are useful only when the supervision process is clear.

An elderly-care organisation needs to know who reviews an algorithmic recommendation, what information they see, whether they can disagree and whether disagreement is recorded.

Human oversight also needs to occur at the right level.

A care worker may be able to challenge an obviously inaccurate observation about daily activity. A nurse or clinician may need to review a health-related risk prediction. A manager may need to investigate repeated errors across many people. Local authorities or healthcare-security bodies may need to examine systems where AI influences publicly funded service allocation.

The relevant principle is that accountability should remain attached to the decision, not transferred to the software.

The wider theme of governance and leadership is therefore fundamental to responsible AI. Intelligent systems can inform judgement, but organisations still need identifiable people responsible for how those systems are used.

Automation bias can make a technically accurate system unsafe

One of the less obvious risks of artificial intelligence is not that the system is always wrong.

It is that people may begin trusting it too readily because it is often right.

This is known as automation bias.

A worker who repeatedly sees accurate fall-risk alerts may gradually stop questioning the system. A clinician may give disproportionate weight to an AI-generated summary because it appears comprehensive. A manager may assume that a resident not identified as high risk does not need review.

These behaviours can weaken professional judgement even when the technology performs reasonably well overall.

Training therefore needs to teach staff not only how to use AI but how to challenge it.

A responsible organisation expects people to compare the algorithmic output with direct observation, history and contextual information.

Operational scenario: a low-risk classification conflicts with what the worker sees

A home-care provider uses an AI-supported system to prioritise supervisory review across a large caseload. The system analyses missed visits, changes in support time, incident reports and care-record information.

An older man with mild dementia remains classified as relatively low risk because there have been no incidents, missed visits or major changes in recorded care.

During several visits, however, one care worker notices that his kitchen contains less food than usual and that he repeatedly asks what day it is. His daughter, who normally visits frequently, has also been absent.

The worker escalates the concern despite the low-risk classification.

Review identifies that the daughter has been admitted to hospital and the man’s informal support has suddenly reduced. His needs are reassessed and additional temporary support is arranged.

The provider subsequently adds loss of informal support as an important contextual variable but does not assume the algorithm can identify every social change automatically.

The scenario demonstrates why frontline judgement remains indispensable. AI is strongest when it helps professionals see patterns they might miss, not when it discourages them from acting on evidence the model does not contain.

Bias can enter AI systems through the data long before a decision is made

Artificial intelligence learns from historical information.

If that information is uneven, incomplete or unrepresentative, the model can reproduce those weaknesses.

This matters in a country as geographically and socially diverse as China.

Data generated in large urban hospitals or well-resourced elderly-care institutions may not reflect the experience of older people in rural counties. People who receive little formal care may leave fewer digital traces than those using intensive services. Older people with limited digital access may be underrepresented in data generated through smart platforms.

An algorithm trained predominantly on one population can therefore perform less reliably elsewhere.

The problem is not necessarily intentional discrimination.

It is structural bias created by unequal data.

Responsible implementation therefore requires organisations to test performance across different groups rather than relying only on overall accuracy.

Rural AI should not depend on data and infrastructure that rural services do not have

Artificial intelligence is sometimes presented as a way to compensate for weaker service capacity in rural areas.

It may contribute to that objective, but only if the technology fits local conditions.

A model requiring continuous high-quality sensor data, extensive electronic records and rapid specialist review may work in a large urban elderly-care institution while being impractical in a village-based support network.

Rural AI may initially create greater value through simpler applications: assisting remote clinical triage, supporting county-level planning, improving scheduling, helping frontline workers document care or identifying patterns within information that is already available.

The strategic principle is similar to rural telemedicine.

Technology should extend local capability rather than create a new dependence on infrastructure that local services cannot sustain.

Explainability becomes more important when AI affects consequential decisions

Not every AI system can explain its reasoning in a way that is meaningful to frontline staff or older people.

That creates difficulty when a recommendation affects care.

If a system identifies somebody as being at high risk of hospital admission, professionals should ideally understand which information contributed materially to that conclusion.

If an eligibility or prioritisation process were ever to use AI more directly, the requirement for understandable reasoning would become stronger still.

Explainability does not mean every worker needs to understand the mathematical architecture of the model.

It means they need enough information to judge whether the output makes sense in the person’s circumstances.

A system saying “high risk” is less useful than one identifying that declining mobility, increasing night-time assistance and recent hospital use contributed to the classification.

AI-generated summaries can improve continuity but can also hide uncertainty

Long records create real practical problems in elderly care.

Important information can become buried among repetitive entries, making it difficult for workers beginning a shift or professionals joining a case to identify what has changed.

AI-generated summaries can help by drawing together recent developments.

But summarisation introduces selection.

The system decides which information appears important and which detail is omitted.

A concise summary may therefore look more certain than the underlying record actually is.

Organisations using generative AI for documentation need ways for staff to check the source information where a summary influences an important decision.

The original record should not disappear behind a generated interpretation.

AI can support rehabilitation, but physical progress still requires direct assessment

Artificial intelligence is also entering rehabilitation through computer vision, adaptive exercise systems and intelligent rehabilitation equipment.

These technologies can potentially analyse movement, adjust exercise difficulty and provide feedback between professional contacts.

For China, this may be valuable as rehabilitation demand rises with population ageing and survival following stroke, fracture and chronic disease.

AI-supported rehabilitation can increase repetition and make selected exercises more accessible at home or in community settings.

However, it should not be assumed that a camera or algorithm can replace all physical assessment.

Pain, fatigue, balance, fear, environmental hazards and subtle neurological changes may require professional judgement.

The strongest model therefore combines intelligent technology with periodic direct review, adjusting the digital programme when the person’s condition changes.

Intelligent elderly-care robots are moving into real-world testing

China’s national intelligent elderly-care robot pilot for 2025–2027 marks an important shift from laboratory capability towards testing in homes, communities and elderly-care institutions.

The programme is intended to bring technology developers and elderly-care organisations together so that robots can be refined through practical use rather than designed in isolation.

Potential functions include mobility support, rehabilitation, monitoring, assistance with daily activities, logistics and forms of companionship.

This implementation approach is significant because elderly care is an unusually demanding environment for robotics.

Homes are cluttered and inconsistent. Older people move unpredictably. Physical assistance carries injury risk. Cognitive impairment can affect interaction. Robots also need to operate around family members and care workers rather than inside a controlled industrial workspace.

Real-world testing is therefore essential before claims about workforce substitution can be taken seriously.

Robotics should be evaluated task by task

The category “elderly-care robot” can obscure major differences between applications.

A robot transporting laundry or meals within an institution performs a comparatively structured task. A rehabilitation robot may operate under professional supervision. A mobility-assistance robot physically interacting with a frail person carries greater safety implications. A companionship robot introduces different ethical and relational questions.

Evaluation should therefore focus on the task rather than the novelty of the device.

Useful tests include whether the robot performs reliably, whether it reduces physical strain, whether staff time is genuinely released, whether older people accept the interaction and what happens when the technology cannot complete the task.

A robot that requires a worker to supervise it continuously may add less workforce capacity than expected.

The national pilot approach is therefore valuable precisely because operational evidence is still being generated.

AI companionship deserves particularly careful scrutiny

Conversational systems and robots may provide reminders, entertainment and a sense of interaction for older people who spend long periods alone.

For some people, this could be positive.

AI may support communication, stimulate memory or provide an accessible way to obtain information.

But digital companionship should not be confused with human relationship.

An older person experiencing loneliness may benefit from technology while still needing family, neighbours, community activity and meaningful human contact.

There is also potential for emotional dependence if systems are designed to simulate intimacy in ways users do not fully understand.

The ethical question is not whether older people should be prevented from forming attachments to technology.

It is whether the technology is transparent about what it is and whether its use supplements rather than quietly replaces human connection.

Operational scenario: a companionship robot is popular but reveals a service gap

An elderly-care institution pilots a conversational robot in a communal area. Several residents enjoy using it for music, games, reminders and simple conversation.

One resident begins spending long periods with the device and tells staff that it is the only “person” who has time to listen.

The technology itself has not malfunctioned.

Instead, it has revealed something important about the service.

Staff review the resident’s daily experience and identify that workforce routines have become highly task focused, leaving limited time for meaningful conversation. The resident is supported to join a smaller activity group and receives more regular one-to-one contact alongside continued voluntary use of the robot.

The institution keeps the technology because it adds value, but it changes how success is measured.

High engagement with the robot is no longer interpreted automatically as evidence of positive social impact. It is considered alongside loneliness, participation and human contact.

The scenario demonstrates how AI can reveal unmet need as well as address it.

Workforce redesign matters more than workforce replacement

China’s elderly-care workforce pressures make automation attractive.

But the most credible near-term opportunity is not a care system with substantially fewer humans.

It is one in which workers spend less time on repetitive administration, unnecessary walking, routine monitoring and predictable logistics while concentrating more attention on observation, skilled support, rehabilitation, communication and complex judgement.

This changes workforce design.

Some roles may require greater digital competence. Supervisors may need to interpret AI-generated trends. Frontline workers may need to validate automated records. Technical support roles may become more important inside larger providers.

The wider field of digital skills, training and workforce adoption therefore becomes increasingly important as AI moves into care settings.

Technology implementation should include training in limitations as well as functionality.

Older workers should not be excluded by the digital transformation of care

The elderly-care workforce itself includes workers with different educational backgrounds and levels of digital confidence.

Rapid introduction of complex AI systems can create anxiety or widen capability differences inside organisations.

Good implementation therefore needs accessible interfaces, practical training and sufficient time for staff to learn new workflows.

Introducing an AI system and then treating workers who struggle with it as resistant to change is poor workforce governance.

Adoption should be judged through whether the technology fits real practice.

Frontline feedback can reveal unnecessary complexity that technical teams may overlook.

Digital inclusion also applies to the older person interacting with AI

Artificial intelligence can make technology easier to use in some circumstances.

Voice interfaces may be more accessible than complex menus. Natural-language interaction can reduce the need to navigate multiple screens. Intelligent systems may adapt to individual preferences.

But AI can also create new forms of exclusion.

Speech systems may perform differently across accents or dialects. People with hearing, speech or cognitive impairments may find conversational interfaces difficult. Older people may be uncertain whether they are interacting with a person, an automated system or a combination of both.

China’s continuing emphasis on digital age-friendliness is therefore directly relevant. Accessible design needs to remain part of AI development rather than being added after systems are deployed.

Privacy risks increase as AI combines multiple sources of personal data

Traditional care records contain sensitive information.

AI systems may combine far more.

Video, voice, movement, sleep, health measurements, location, service history and family information can collectively create a detailed picture of someone’s private life.

The analytical power comes partly from combining these datasets.

That same combination increases privacy risk.

The broader principle of digital safeguarding and technology-enabled harm becomes particularly important where AI operates inside private homes or monitors behaviour continuously.

Organisations need to ask not simply whether data can improve prediction, but whether collecting them is proportionate to the purpose.

Data minimisation remains relevant even when more data could theoretically improve the model.

Cybersecurity and AI safety increasingly overlap

An AI-enabled elderly-care system depends on software, networks, data and often cloud infrastructure.

A security failure can therefore become a care failure.

Compromised monitoring systems may stop generating alerts. Manipulated data could affect decision-support outputs. Unavailable platforms can interrupt scheduling or access to records.

Cyber resilience needs to sit alongside model accuracy as part of system assurance.

The risk increases as providers connect multiple devices and platforms.

An intelligent elderly-care environment is only as secure as its weakest significant connection.

AI governance needs evidence from real care environments

Laboratory accuracy is not enough.

AI used in elderly care needs evaluation under the conditions in which it will actually operate.

Those conditions include incomplete records, changing staff, different home environments, cognitive impairment, unreliable connectivity and people whose behaviour does not fit standard patterns.

Organisations examining comparable implementation can use the Digital Transformation Readiness Assessment to test whether technology, governance, workforce, resilience and operational processes are developing together. It is not a China-specific AI framework, but the implementation discipline is relevant.

Responsible scaling should depend on evidence from real service use rather than demonstrations alone.

AI assurance needs to continue after a system goes live

Artificial intelligence cannot be treated as a conventional technology purchase that is tested once and then assumed to continue performing in the same way.

Models can become less reliable as populations, workflows and data change. A risk model trained on one group of residents may perform differently when introduced into another service. Recording practices may change after a new digital system is installed. Software updates can alter outputs. Workers may also change how they behave because they know an algorithm is monitoring particular variables.

Responsible AI therefore requires continuing assurance.

For an elderly-care organisation, this means examining not only whether the system is technically functioning but whether it remains useful and safe in practice. Managers need to know whether alerts are accurate enough to justify the workload they create, whether particular groups are generating disproportionately high or low risk classifications and whether professionals are routinely overriding recommendations.

Frequent overrides do not necessarily mean staff are resisting technology. They may indicate that the model does not understand the care environment well enough.

Organisations examining comparable assurance questions can use the Governance Maturity Assessment to structure responsibility, escalation and oversight around significant technology-enabled decisions. It is not a China-specific AI framework, but the principle that accountability should remain visible throughout implementation is directly relevant.

AI performance should be measured through care outcomes as well as technical accuracy

A model can achieve impressive technical performance while creating little meaningful improvement for older people.

A falls algorithm may identify risk accurately but fail to reduce injuries because the service lacks rehabilitation, environmental adaptation or sufficient staff to respond. An intelligent scheduling system may optimise travel time while producing poorer continuity because older people see more unfamiliar workers. A companionship robot may record high engagement while loneliness remains unchanged.

AI evaluation therefore needs a wider evidence base.

Depending on the application, decision-makers may need to consider whether the technology changes:

• timeliness of assessment or intervention;
• functional independence and preventable deterioration;
• falls, hospital transfers or other significant safety events;
• worker time, administrative burden and continuity;
• family-care burden and confidence;
• older people’s experience, autonomy and willingness to use the system.

This broader approach prevents technological performance from becoming a substitute for care-system performance.

Organisations considering similar evidence structures can use the Quality Dashboard Builder to bring technology, workforce, safety and outcome information into a more balanced view. It does not validate an AI system or determine Chinese regulatory compliance, but it illustrates the value of evaluating technology alongside the operational outcomes it is intended to improve.

Operational scenario: an accurate risk model produces the wrong operational response

A large elderly-care organisation introduces an AI system that predicts which residents are most likely to fall during the following month.

Testing shows that the model identifies a large proportion of residents who subsequently fall, and managers initially regard the implementation as successful.

Frontline practice then begins to change.

Workers become increasingly cautious with residents classified as high risk. Some discourage independent walking and begin providing more hands-on assistance because they fear an incident occurring after the system has issued a warning.

Several residents become less active.

A later quality review identifies an unexpected pattern: fall numbers have changed only modestly, while mobility and independence have deteriorated among part of the high-risk group.

The organisation changes the response model. A high-risk classification now triggers multidisciplinary review of mobility, medication, environment, footwear and rehabilitation rather than automatic restriction. Where appropriate, workers are explicitly encouraged to support safe movement and retain existing ability.

The AI model had performed the statistical task it was designed to perform. The weakness was the human response built around it.

The scenario demonstrates why responsible AI requires governance of consequences as well as accuracy. Prediction is valuable only when it leads to proportionate action that improves the person’s overall outcome.

AI should not quietly become an eligibility system without appropriate safeguards

The growth of long-term care insurance and more standardised functional assessment creates considerable potential for data-driven administration in China.

AI could eventually help identify assessment inconsistencies, detect unusual claims patterns, support reassessment prioritisation or identify people whose changing information suggests that their needs should be reviewed.

These applications could improve efficiency and fund stewardship.

A much higher threshold should apply if algorithmic systems begin directly influencing whether somebody qualifies for publicly financed support or how much care they receive.

Eligibility decisions have significant consequences for older people and families. Data may not fully describe informal support, housing conditions, cognitive difficulties or the practical consequences of functional limitation.

A model can also institutionalise historical variation if trained on previous decisions that were themselves inconsistent.

AI may therefore support assessors by highlighting relevant evidence or unusual patterns. Final decisions affecting entitlement should remain subject to accountable processes, understandable reasoning and appropriate review.

Efficiency should not turn administrative prediction into an invisible barrier to support.

Public authorities will increasingly need to govern AI markets as well as individual projects

China’s large technology sector means elderly-care organisations are likely to face a rapidly expanding range of AI products.

The market may include predictive systems, robots, intelligent beds, digital assistants, automated documentation platforms, monitoring products and integrated smart-care environments.

Not every product will have equivalent evidence.

This creates a procurement and system-governance challenge for Civil Affairs departments, elderly-care organisations, healthcare institutions and other public bodies.

Decision-makers need to distinguish between products that are technically innovative and those that can operate reliably within real care pathways.

Useful scrutiny includes the quality of training data, performance across relevant populations, compatibility with existing systems, cybersecurity, maintenance arrangements, accessibility, workforce requirements and the supplier’s ability to support the product over time.

Public procurement can also influence market behaviour. If purchasing decisions reward demonstrable outcomes, interoperability and age-friendly design, suppliers have stronger incentives to develop technology around service need rather than novelty.

Local AI pilots should produce evidence that can travel beyond the pilot site

China’s capacity for local experimentation can accelerate responsible AI development if pilots are evaluated carefully.

Technology-rich areas such as Beijing’s Haidian District can test models involving smart communities, intelligent elderly-care institutions, demand analysis, robotics and health-management technologies. Other cities and provinces are developing their own combinations of AI-enabled community and home support.

The policy value lies partly in discovering what does not work.

A pilot may show that an application requires too much staff supervision, that older people reject a particular interface or that savings assumed during procurement do not materialise after maintenance and training costs are included.

That information is valuable.

A mature innovation system should be willing to stop or redesign weak models rather than interpreting every pilot as something that must subsequently scale.

Where a project performs well, evaluation should identify which features made the difference. Was success dependent on unusually strong technical infrastructure? Did the locality have specialist staff unavailable elsewhere? Was extensive public subsidy required? Could a simpler version generate most of the same benefit?

This helps provinces and national authorities distinguish transferable principles from conditions that are specific to the demonstration site.

National standards can help AI scale without requiring identical local systems

China does not need every province or municipality to deploy the same artificial intelligence platform.

It does need sufficiently clear expectations around safety, data, interoperability, accessibility and accountability if AI becomes increasingly embedded in elderly care.

Common standards can make local innovation safer while preserving room for different service models.

The intelligent elderly-care robot programme already reflects this direction by combining practical application testing with development of standards and evaluation around issues including safety, reliability, age-friendliness and economic viability.

A similar principle can apply more widely to AI-enabled elderly care.

National guidance can establish boundaries and evidence expectations, while local systems determine which applications fit their population and infrastructure.

This offers a more realistic route than either uncontrolled experimentation or premature national standardisation of rapidly changing technology.

The 15th Five-Year Plan makes responsible AI part of mainstream ageing policy

The 2026–2030 national development direction substantially increases the strategic relevance of artificial intelligence to health and elderly care.

The 15th Five-Year Plan calls for the wider implementation of the “AI+” agenda while promoting orderly use of digital and intelligent technologies in areas including assisted diagnosis, precision medicine, health management, medical insurance, elderly care and support for people with disabilities.

This wording is important because it combines technological expansion with the concept of orderly application.

For elderly care, the next phase is therefore likely to involve more real-world deployment rather than AI remaining primarily an innovation-sector subject.

That creates an opportunity to improve productivity and personalisation at scale.

It also raises the consequences of weak implementation.

If algorithms become embedded in care planning, resource allocation, monitoring or professional workflow, data quality and accountability become part of mainstream elderly-care governance. Digital competence becomes a workforce requirement. Interoperability becomes more important because AI is more useful when relevant information can connect across services. Privacy questions become harder as systems combine multiple datasets.

The next stage of Chinese AI elderly care will therefore be determined as much by institutional capability as by technical capability.

What China’s development of AI in elderly care offers international systems

China’s technology industry, population scale, administrative structure and approach to local policy experimentation differ significantly from those of many other countries. Its specific AI programmes should therefore not be treated as models that can simply be replicated elsewhere.

The transferable lessons lie at a deeper level.

Artificial intelligence creates greatest value when it strengthens a defined care pathway rather than when organisations search for uses simply because the technology exists.

Administrative automation can often be scaled with less risk than automated judgement. Prediction requires a response pathway. Personalisation requires reliable longitudinal data. Robotics should be assessed task by task. AI companionship needs to be judged alongside human relationships. Local pilots need evidence that explains why a model succeeded, not only evidence that it operated.

Most importantly, human oversight needs to remain meaningful.

A system in which professionals routinely accept algorithmic output without understanding or challenging it is not genuinely human supervised. Responsible AI requires people with sufficient authority, information and competence to disagree.

Conclusion

Artificial intelligence could become an important part of China’s response to population ageing. Predictive analytics can help identify deterioration earlier, generative systems can reduce documentation burden, intelligent scheduling can improve productivity, rehabilitation technology can extend professional support and robotics may increasingly assist with selected physical, logistical and social tasks.

The strategic opportunity is not autonomous elderly care. It is a more capable care system in which technology helps people identify patterns, remove low-value work and direct scarce human expertise towards the situations in which judgement and relationship matter most.

That opportunity depends on discipline. AI needs reliable data, age-friendly design, continuing evaluation and cybersecurity. High-impact recommendations need understandable reasoning and accountable human review. Local pilots should demonstrate meaningful outcomes before wider adoption, while national standards can provide common expectations without suppressing useful regional experimentation.

During the 15th Five-Year Plan period, AI is likely to move further from demonstration projects into ordinary health and elderly-care services across China. The quality of that transition will depend less on how intelligent individual technologies appear and more on whether the surrounding system remains intelligent enough to question them. Responsible practice means preserving a clear principle as automation expands: artificial intelligence can inform, predict and assist, but accountability for the life and wellbeing of an older person must remain recognisably human.