AI and Long-Term Care in Taiwan: Opportunities, Governance and the Human Dimensions of Care

An older person does not experience artificial intelligence as a national technology strategy. They experience a lifting device that helps a care worker reposition them more safely, a rehabilitation system that adapts exercises to their movement, a digital service that answers a question about long-term care, or a monitoring system that notices something unusual. Whether any of those technologies improves care depends less on the label “AI” than on what happens around it.

That distinction is becoming increasingly important in Taiwan. Long-Term Care 3.0 began in 2026 with smart care established as one of its major directions, while the Ministry of Health and Welfare has encouraged technology adoption in community and residential long-term care settings. Taiwan’s wider capabilities in artificial intelligence, electronics, robotics and digital health create significant possibilities for an ageing society. Within the Taiwan Ageing, Long-Term Care & Community Support Knowledge Hub, however, AI needs to be understood as part of the care system rather than as a separate technological story.

The central question is not whether Taiwan can develop increasingly capable care technologies. It can. The more difficult question is how those technologies should be integrated into services where decisions concern dignity, disability, cognition, privacy, family relationships, clinical risk and everyday human dependence.

AI can help people work differently. It can recognise patterns, automate repetitive processes, support physical tasks and extend access to information. But long-term care also involves judgement, reassurance, touch, trust and understanding what matters to an individual. Taiwan’s strongest opportunity therefore lies not in automating care indiscriminately, but in identifying where technology can increase human capability while keeping responsibility clearly human.

Long-Term Care 3.0 moves smart care into mainstream policy

Taiwan’s development of AI-enabled long-term care is occurring within a much larger reform of the care system. LTC 2.0 concentrated heavily on expanding service coverage and building home and community infrastructure. LTC 3.0 retains that foundation but places greater emphasis on quality, medical and long-term care integration, healthy ageing, rehabilitation, workforce sustainability and technology.

Smart care is explicitly one of the programme’s major directions. In 2026, national policy began supporting technology adoption in community and residential services, including incentives for day-care centres and residential institutions to introduce smart-care technologies intended to improve safety, autonomy and workforce efficiency.

This matters because technology adoption in long-term care has often occurred through isolated projects. One institution purchases sensors. Another tests a robot. A municipality develops a digital service. A university conducts a rehabilitation trial. Useful innovation can emerge, but fragmentation makes it difficult to determine what should become routine care.

National policy creates an opportunity to move from experimentation towards a more coherent model. That requires distinguishing at least four different purposes:

• technology that directly supports an older or disabled person’s independence;
• technology that assists workers with physical or administrative tasks;
• analytics that support professional or organisational decisions;
• automation that helps citizens navigate information and services.

These applications create different risks. A robot assisting with physical movement requires strong safety controls. An algorithm prioritising information for professional review raises questions about bias and explainability. A generative AI service answering public questions must manage inaccurate responses and escalation. A sensor inside somebody’s home raises privacy and consent questions.

Treating all of them simply as “smart care” can obscure those differences.

AI should solve care problems rather than search for care applications

Taiwan’s technological strength creates a particular strategic temptation: starting with what technology can do and then finding somewhere in care to deploy it.

Long-term care requires the opposite discipline.

The starting point should be an operational problem. Care workers may experience musculoskeletal strain from physically demanding tasks. Families may struggle to obtain information outside office hours. Rehabilitation capacity may be limited. Staff may spend excessive time on repetitive documentation. Managers may find it difficult to recognise patterns across large volumes of quality information.

Technology can then be assessed against that defined problem.

This approach matters because an impressive technical demonstration is not necessarily a sustainable care service. A device can work accurately in controlled testing but be difficult to clean, maintain or operate in a busy residential setting. An AI system can generate fluent answers while occasionally producing confidently incorrect information. A monitoring system can identify hundreds of alerts but increase rather than reduce staff workload.

The wider field of AI and automation in care therefore needs an outcome discipline. The question is not simply whether a task can be automated. It is whether doing so improves safety, independence, continuity, workforce capacity or experience sufficiently to justify its costs and risks.

This is particularly important as Taiwan’s long-term care workforce expands while demographic change increases demand. Technology that genuinely removes low-value workload could release more time for human interaction. Technology that adds another system, login, alert stream or recording requirement may do the opposite.

Robotics could reduce physical burden without removing the care worker

Robotics is one of the areas where Taiwan’s industrial and research capabilities intersect directly with long-term care.

Research and development already includes technologies designed to assist rehabilitation, movement and physically demanding health-care tasks. In 2026, Taiwan also strengthened its wider AI robotics infrastructure through collaboration between government technology and economic-development bodies, with health care identified as a significant application field.

The long-term care opportunity is substantial because physical workload is a real workforce issue. Repositioning people with high support needs, transfers, mobility assistance and repetitive manual tasks can contribute to worker fatigue and injury. Mechanical and robotic assistance may help reduce some of that burden.

Yet the objective should not be to remove people from intimate care.

Consider repositioning. A robotic or sensor-assisted system may help determine position and provide controlled physical assistance. The care worker still needs to understand pain, skin integrity, anxiety, communication preferences, clinical restrictions and whether the person wants the intervention to proceed.

The machine may provide force. The worker provides context.

This is an important workforce principle. AI and robotics can change the distribution of work without eliminating professional responsibility. In some cases, they may actually increase the skills required of frontline staff because workers need to understand when technology is appropriate, recognise malfunction and know when human judgement should override an automated recommendation.

Organisations considering such transformation can use the Digital Transformation Readiness Assessment to structure questions about leadership, workforce readiness, cyber resilience and implementation. It is not a Taiwanese approval mechanism, but its underlying principle is transferable: technological capability and organisational capability need to develop together.

Scenario: a lifting technology changes the work but not the responsibility

A residential long-term care institution introduces an intelligent transfer-assistance system for residents who require substantial physical help. One resident has limited mobility following a stroke and becomes anxious during transfers because of a previous fall.

The equipment can reduce the physical effort required from staff and provide greater consistency in movement. Initially, however, workers concentrate heavily on operating the technology correctly. The resident reports that transfers feel more mechanical and that staff sometimes speak to one another about the device rather than explaining what is happening to her.

The institution treats this as a quality issue rather than evidence that the technology itself has failed.

Training is revised. Staff are expected to explain the transfer before beginning, confirm the resident’s readiness, observe discomfort throughout and stop when circumstances change. Competency assessment covers both safe operation and communication. Equipment faults and near misses are reviewed alongside resident feedback and worker injury information.

Several months later, physical strain among staff undertaking these transfers has reduced and the resident reports greater confidence because the process has become predictable without becoming impersonal.

The scenario illustrates what successful automation can look like in long-term care. Technology performs part of the physical task, but accountability remains with the organisation and its workforce. The relevant outcome is not that a robot was deployed. It is whether the combination of worker and technology produced safer, more dignified care.

AI-assisted rehabilitation may extend capability, but evidence matters

Rehabilitation is another important area of development. Taiwan’s research institutions are exploring advanced systems that combine artificial intelligence, robotics, sensing and neurological technologies to support recovery of movement.

Some emerging systems are highly experimental. Research involving brain-computer interfaces and robotic walking assistance, for example, illustrates the direction in which personalised rehabilitation may develop, but clinical validation remains part of the development process. Such technologies should therefore not be described as established national long-term care provision.

More mature digital and robotic rehabilitation technologies can nevertheless help services deliver repeated exercises, capture movement data and provide structured feedback. This may be particularly valuable where specialist workforce capacity is limited.

The operational question is whether technology supplements therapeutic expertise or becomes a reason to reduce it.

A rehabilitation professional does more than prescribe repetitions. They interpret pain, motivation, fatigue, cognition, environmental barriers and whether a functional goal remains meaningful. An algorithm may adjust an exercise according to measured performance without understanding that the person’s actual priority is being able to reach a neighbourhood market independently.

AI-enabled rehabilitation should therefore remain connected with outcomes-focused support. Improvement in a technical performance measure is useful, but its strongest significance lies in what the improvement enables the person to do.

Generative AI is already changing service navigation

One of the most immediate applications of generative AI in Taiwan’s long-term care system is not physical care but information.

The 1966 long-term care service hotline has traditionally provided an important route through which citizens and families can ask questions about services and access. Demand is uneven across the day, however, because many family caregivers make contact during lunch breaks or after work. Human staffing cannot be expanded indefinitely simply to match every demand peak.

In 2026, the Ministry of Health and Welfare publicly explored an AI-enabled digital avatar approach for long-term care customer service, using large language models and conversational technology to create more natural interaction. The model illustrates a potentially valuable use of AI: extending access to routine information while allowing human staff to concentrate on questions requiring judgement or individual problem-solving.

But long-term care enquiries are not always routine.

A person may begin by asking how to apply for home care and then reveal that an exhausted family caregiver is no longer able to cope. Somebody asking about residential services may actually be describing an urgent safety concern. A migrant care worker may need information but find Mandarin text interaction difficult.

Generative AI therefore needs escalation architecture.

The system should be able to distinguish between information that can reasonably be provided automatically and situations that require human intervention. It should also make its limitations clear. A fluent conversational interface can create greater trust than a traditional menu-driven chatbot, which makes incorrect answers potentially more consequential.

Accessibility is equally important. Voice, multilingual capability and simple interfaces may improve inclusion, while a text-only system can reinforce barriers for some older people and migrant caregivers.

Scenario: the question is simple until it is not

A daughter contacts a digital long-term care information service late in the evening. She initially asks whether her father could qualify for home-based support after his mobility has deteriorated.

An AI assistant can explain the general application route and describe the role of long-term care assessment. During the conversation, however, the daughter says that her father has fallen twice in the past week and is now afraid to walk to the bathroom without assistance. She also explains that she cannot remain at his home overnight.

The conversation has moved from routine service navigation to potential immediate risk.

A well-governed system should not continue generating increasingly detailed generic advice as though nothing has changed. Its escalation rules should recognise the risk indicators, explain that the situation may require prompt professional assessment and direct the user towards an appropriate human or urgent-care route according to the circumstances.

The interaction should also be recorded proportionately so that, where a formal referral follows and information can lawfully be transferred, the family does not have to reconstruct the entire situation repeatedly.

The example demonstrates why conversational quality is not the same as service quality. A chatbot that sounds empathetic but fails to recognise escalation can be more dangerous than a basic system whose limitations are obvious.

AI service navigation should therefore be governed through clear decision-making and escalation, with defined boundaries between information, triage and professional judgement.

Smart assistive technology is moving closer to funded long-term care

Taiwan is also developing a more formal infrastructure around smart assistive technology within the long-term care system.

In 2026, the Ministry of Health and Welfare established arrangements for specified smart assistive technology products to be registered through the Long-Term Care Assistive Technology Information Platform. Product applications are checked against defined specifications, functions and relevant safety documentation before eligible products are included on the platform.

This is significant because one of the challenges in technology-enabled care is the gap between consumer innovation and publicly supported care.

A product may be commercially available without being appropriate for public reimbursement or formal care use. Conversely, a highly controlled purchasing framework can become outdated if it cannot respond to technological change.

Taiwan’s approach creates a mechanism through which specified technologies can be considered within the long-term care benefit environment while maintaining a structured product-registration process.

The broader policy challenge will be ensuring that approval of a product does not become confused with suitability for every individual.

An intelligent mobility or monitoring device may meet technical standards yet be inappropriate for somebody with particular cognitive, sensory or environmental circumstances. Assessment therefore remains necessary.

The relevant person-centred principle is person-centred technology: the device should fit the person’s goals, abilities, environment and preferences rather than requiring the person to adapt to the technology.

AI can strengthen workforce capacity without becoming a workforce strategy

Taiwan’s ageing population creates a powerful economic argument for automation. More people are likely to require long-term support while growth in the traditional working-age population is constrained. Formal long-term care already employs more than 100,000 care workers, alongside a substantial migrant caregiving workforce and unpaid family care.

It would nevertheless be misleading to frame AI primarily as a substitute for workers.

Long-term care includes many tasks that can potentially be redesigned:

• routine scheduling and administrative processing;
• summarising information for professional review;
• physical assistance with selected high-strain tasks;
• monitoring equipment and environmental conditions;
• supporting documentation and information retrieval;
• identifying patterns in quality, demand or workforce data.

Reducing time spent on these activities could increase the proportion of the workforce available for direct relationships, observation, rehabilitation and coordination.

But productivity gains do not appear automatically.

If staff have to document care in both old and new systems, automation creates duplication. If sensors generate excessive alerts, monitoring creates work. If AI-generated notes require extensive correction, documentation becomes slower. If equipment is unreliable, workers spend time troubleshooting it.

The correct workforce measure is therefore not how many technologies have been introduced but whether total workload has changed and whether the change improves care.

This is closely connected with workforce planning. Organisations need to understand which roles technology changes, what new competencies are required and whether savings in one part of the workflow create additional tasks elsewhere.

A Predictive Workforce Risk Module can help organisations examining similar questions structure analysis of vacancy, retention and continuity pressures. It does not predict Taiwan’s workforce requirements, but it illustrates how technology decisions can be considered alongside workforce risk rather than as a separate digital programme.

AI-generated records require human ownership

Generative AI could eventually reduce one of the most persistent administrative burdens in care: turning observations, conversations and professional decisions into structured records.

Speech-to-text systems can already transcribe conversations. Large language models can summarise information and draft structured text. Combined carefully, these technologies could reduce repetitive writing and allow professionals to spend more time with people.

However, a care record has consequences.

An inaccurate summary may influence the next worker’s understanding of a person. A generated statement can appear authoritative even when the original conversation was ambiguous. Important nuance can disappear. A description of distress may become more categorical than the worker intended.

AI-generated documentation should therefore be treated as draft information requiring accountable human review where it enters an official record.

This is particularly important when records influence assessment, health-care decisions, safeguarding or service eligibility. Efficiency should not weaken evidential integrity.

The wider discipline of digital records and information governance remains applicable regardless of how the text was produced. Organisations need clarity about who verified information, what source material informed it and how errors can be corrected.

Predictive AI should support attention, not automate entitlement

Taiwan’s expanding digital infrastructure creates the possibility of applying AI to large health and care datasets. At population level, this could help identify emerging demand, geographic differences or patterns associated with deterioration.

At individual level, the ethical threshold becomes higher.

A predictive model could potentially recognise combinations of service use, functional change or health events associated with increasing care need. That information might prompt a professional review. It should not automatically determine that somebody is eligible or ineligible for support.

There are several reasons for caution.

Historical data describe previous patterns of service use as well as underlying need. If some communities have historically accessed services less frequently, an algorithm trained on utilisation can interpret under-access as lower need. Informal family care may also make formal demand appear lower even where the underlying level of dependency is substantial.

Prediction can therefore reproduce existing inequalities unless the meaning of the data is examined carefully.

The stronger use of AI is to identify questions requiring attention: why is demand changing, why is a group under-represented, where might capacity become constrained, or which cases might benefit from professional reassessment?

AI can narrow the field of inquiry. Human governance remains responsible for deciding what follows.

Scenario: a risk model identifies the wrong explanation

A local long-term care team pilots an analytical model designed to identify people whose patterns suggest an increasing likelihood of needing additional support. An older man is flagged because of repeated health-care contacts and declining mobility.

The model interprets the pattern as increasing dependency. A care manager reviewing the information speaks with the man and discovers something more complicated. He recently underwent treatment that temporarily reduced his mobility, but he is improving through rehabilitation. His primary concern is not additional personal care; it is regaining enough confidence to use public transport and resume community activities.

If the algorithm directly increased his care package, resources might be directed towards an intervention he neither needs nor wants.

Instead, the risk flag triggers reassessment. The professional combines clinical information, functional progress and the man’s own goals before determining the next step.

The organisation also learns from the case. It examines whether temporary post-treatment patterns are creating systematic false positives and adjusts the model or its interpretation guidance accordingly.

This is the difference between predictive support and automated decision-making. The AI successfully identified a case worth reviewing even though its implied explanation was incomplete.

Good governance allows both things to be true: the technology added value, and the technology was not entitled to the final decision.

Consent becomes more complex when technology observes continuously

Traditional care interactions are relatively visible. A worker enters a home. A clinician conducts an assessment. A person knows that an interaction is taking place.

AI-enabled monitoring can be less visible.

Sensors may analyse movement continuously. Cameras or computer-vision systems can interpret physical activity. Voice technology may listen for commands. Wearable devices can generate data throughout the day.

These technologies can support safety, but they change the relationship between observation and privacy.

A person may consent to fall detection without wanting detailed behavioural monitoring. They may accept a sensor in one room but not another. Family members may value reassurance more highly than the person being monitored does.

Consent therefore needs to be specific enough to remain meaningful.

Where cognition changes, the issues become more difficult. Families and professionals may reasonably prioritise safety, but technology should not automatically become a way of imposing continuous surveillance simply because somebody is considered vulnerable.

Long-term care needs a proportionate approach that considers benefit, intrusion, alternatives and the individual’s preferences. This aligns with wider principles of safeguarding, consent and human rights in later life.

Technology can sometimes enable greater freedom. A proportionate sensor may allow somebody to remain at home with less intrusive physical supervision. The same technology can become restrictive if used primarily for organisational convenience.

The ethical question is therefore not whether monitoring is inherently good or bad. It is whether the particular monitoring arrangement increases the person’s safety and autonomy without imposing unnecessary intrusion.

Scenario: monitoring offers reassurance but begins to reshape family relationships

An older woman with early dementia lives alone. Her daughter worries that she may leave home during the night and proposes a combination of movement sensors and door monitoring.

The woman accepts a limited system because remaining in her own home is important to her. Initially, the arrangement works well. Her daughter receives defined alerts when unusual night-time door activity occurs rather than monitoring every movement continuously.

Over time, however, the daughter becomes increasingly anxious and begins checking the system frequently. She calls her mother whenever the movement pattern appears unusual, including occasions when her mother is simply awake early or moving around the house normally.

The technology has not malfunctioned. The problem is how its information is being used.

During review, the family and care team reconsider the purpose of monitoring. Alert thresholds are refined, the daughter is encouraged not to treat ordinary movement as an emergency, and the woman confirms which monitoring she remains comfortable with.

The system continues because it supports her goal of living at home, but its use becomes more proportionate.

This illustrates a distinctive governance issue in AI-enabled care. Technology changes behaviour not only through automated decisions but through the information it gives people. Families may become more confident, or more anxious. Staff may become more attentive, or overly dependent on alerts.

Good technology governance therefore examines human behaviour around the system as well as technical performance.

Cyber security becomes a care-continuity issue

As long-term care becomes more connected, cyber resilience moves from an information-technology concern into service continuity.

A paper record being unavailable affects information access. A connected care environment failing can affect alarms, communication, medication workflows, monitoring or access to essential information simultaneously.

AI systems also introduce additional dependencies. They may rely on cloud services, external vendors, software updates, data connections or proprietary models. Organisations need to understand what happens when those dependencies fail.

A smart-care service should therefore have a credible degraded mode. Workers need to know how essential care continues during an outage. Safety-critical equipment requires appropriate fallback arrangements. Providers need clarity about vendor responsibilities, update processes and how vulnerabilities are addressed.

This is where cyber security and digital resilience become inseparable from operational care quality.

The issue is particularly important for smaller providers. Large hospitals and technology organisations may have specialist cyber teams; community and long-term care services can have much more limited technical capacity. National encouragement of smart-care adoption therefore needs to consider implementation support as well as product availability.

Quality assurance must extend beyond technical accuracy

AI evaluation in long-term care needs a broader definition of quality than model accuracy.

A system can perform its technical function correctly while producing poor care outcomes. A highly accurate monitoring system can generate too many alerts for staff to manage. A documentation tool can summarise accurately while encouraging workers to spend less time listening to people. A navigation chatbot can answer standard questions correctly while failing during unusual or high-risk conversations.

Quality assurance should therefore consider several dimensions together:

• technical safety and reliability;
• accuracy and known limitations;
• impact on the person’s autonomy and experience;
• effects on workforce workload and behaviour;
• equitable performance across different populations;
• service continuity when technology fails;
• whether measurable outcomes improve after implementation.

The Quality Dashboard Builder offers organisations examining comparable innovations a way to structure operational, quality and outcome evidence rather than relying on implementation activity alone.

This distinction will become increasingly important as public funding supports technology adoption. Counting devices purchased or institutions participating may demonstrate programme reach. It does not establish whether residents became safer, workers experienced less strain or people maintained greater independence.

Local innovation needs national learning

Taiwan’s 22 cities and counties create valuable opportunities for local experimentation. Different areas are already sharing experiences of introducing smart technology into long-term care services.

Local variation can accelerate innovation because organisations can adapt technology to their service environment. But it can also produce repeated experimentation in which similar problems are solved separately without enough shared learning.

LTC 3.0 creates an opportunity for a stronger learning architecture.

National bodies can establish broad requirements around safety, information governance, interoperability and evidence while allowing local services to test different operating models. Municipal and county experience can then help identify which technologies deserve wider adoption and which have produced limited value.

This requires willingness to learn from unsuccessful implementation as well as success.

A technology may fail because the product is unsuitable, because the workforce was not trained, because the workflow was not redesigned or because the problem itself was poorly defined. Those are different findings with different implications.

The wider discipline of continuous improvement is therefore especially relevant to smart care. Innovation should be iterative rather than a one-off procurement decision.

AI governance needs clear human accountability

The more autonomous technology becomes, the more important it is to establish who remains responsible.

A technology supplier is responsible for aspects of product design and performance. A long-term care organisation is responsible for how a system is deployed within its service. Professionals remain responsible for decisions within their roles. National and local authorities establish policy, funding and oversight arrangements.

Problems arise when responsibility disappears between these actors.

A provider should not assume that purchasing an approved product transfers responsibility for how it is used. A professional should not treat an algorithmic recommendation as inherently correct because the underlying model is technically complex. A supplier should not define care policy through default software settings.

Organisations examining similar questions can use the Governance Maturity Assessment to structure accountability, escalation and assurance around technological change. Again, the framework does not replace Taiwanese requirements; its relevance lies in ensuring that ownership remains visible as systems become more complex.

Human accountability is not an obstacle to AI adoption. It is one of the conditions that makes responsible adoption possible.

The human dimensions of care remain technologically difficult

Long-term care contains forms of intelligence that are difficult to encode.

An experienced worker may notice that a normally talkative person has become unusually quiet. A daughter may understand that her father saying “I’m fine” has a different meaning on a particular day. A rehabilitation professional may recognise that somebody has the physical ability to complete an exercise but has lost confidence after a fall.

These judgements draw on relationship, history, context and subtle communication.

AI will become better at analysing language, movement and behavioural patterns. That does not make relational knowledge irrelevant. It may make the distinction more important.

The danger is not only that technology makes mistakes. It is that organisations gradually redesign services around what technology can measure.

If easily captured indicators dominate, outcomes such as belonging, confidence, cultural identity, trust and feeling known can receive less attention precisely because they are harder to quantify.

Person-centred long-term care therefore requires a deliberate commitment to forms of evidence that include the individual’s own experience. AI can enrich that evidence environment, but it should not define it.

What Taiwan’s experience can offer internationally

Taiwan enters the AI era with an unusual combination of strengths: substantial technology capability, extensive digital-health infrastructure, national long-term care reform, a growing formal care network and strong policy interest in smart care. Those conditions are specific to Taiwan and cannot simply be reproduced elsewhere.

Several underlying principles nevertheless have wider relevance.

First, AI adoption is stronger when it begins with a defined care problem rather than a technology looking for a use case.

Second, workforce productivity should be assessed across the complete workflow. Automating one task has little value if it creates additional work elsewhere.

Third, public adoption of smart technology needs evidence beyond technical performance. Safety, experience, equity, workload and human outcomes all matter.

Fourth, predictive systems are most defensible when they support attention and professional review rather than automatically determining access to care.

Fifth, local experimentation needs mechanisms for national learning. Scaling innovation should depend on evidence about implementation as well as enthusiasm for novelty.

Finally, the more sophisticated AI becomes, the more explicit human accountability needs to be. Technology can distribute information and capability across a system, but responsibility for care cannot simply disappear into an algorithm.

Conclusion

Artificial intelligence is becoming part of Taiwan’s long-term care development at a significant moment. LTC 3.0 has moved smart care into the national reform agenda, public support for smart assistive technologies is becoming more structured, municipalities are sharing technology-enabled practice, and Taiwan’s wider robotics and AI capabilities are creating new possibilities for rehabilitation, physical assistance, service navigation and operational intelligence.

The strategic opportunity is considerable, but it is not primarily about replacing care workers. The stronger model uses AI to remove avoidable administrative burden, reduce physically demanding work, extend specialist capability, identify information requiring attention and help people maintain greater independence. Human time can then be concentrated where relationships, judgement, reassurance and individual understanding matter most.

That requires governance to develop alongside innovation. Consent must remain meaningful. Predictive models should not become automatic gatekeepers to support. AI-generated records need accountable verification. Connected technologies require cyber resilience and fallback arrangements. Public investment needs evidence of outcomes rather than simply evidence of adoption.

Taiwan’s central challenge is therefore not whether long-term care will become more technological. That direction is already visible. It is whether technological capability can be integrated without allowing care to become less human.

If LTC 3.0 can maintain that balance, AI can strengthen Taiwan’s response to population ageing not by automating human relationships, but by giving people, families and care workers better tools with which to sustain them.