An older person does not experience artificial intelligence as an algorithm. They experience whether somebody notices that they are becoming frailer, whether a clinician identifies a change in risk, whether medication is reviewed before harm occurs, whether a home-care worker has enough time to provide meaningful support, and whether decisions affecting their life can be understood and challenged.
That distinction is particularly important in Israel. The country combines mature digital-health infrastructure, extensive longitudinal health data, sophisticated technology capabilities and a growing artificial-intelligence ecosystem with a long-term care system distributed across health plans, hospitals, the National Insurance Institute, welfare services, municipalities, providers and families. The wider Israel Aging, Long-Term Care & Community Support Knowledge Hub shows how those responsibilities intersect across later life.
AI could strengthen that system in several ways. It may help identify deterioration earlier, prioritize people requiring assessment, support clinical decisions, reduce repetitive administrative work, improve service planning and assist professionals who are working with increasingly complex needs. Israel's Ministry of Health is already developing governance approaches for AI-enabled medical technology, including published development principles and a regulatory sandbox for breakthrough healthcare applications. The direction is therefore moving beyond theoretical discussion toward supervised real-world adoption.
But aging and long-term care create a demanding test of responsible AI. Predictions can shape access to professional attention, influence risk judgments, affect autonomy and alter how scarce human care is distributed. The central question is not whether algorithms can identify patterns. It is whether those patterns can be translated into fair, explainable and accountable decisions that improve the lives of older people rather than simply making organizations more efficient.
AI enters an already complex care system
Artificial intelligence does not arrive in Israel as a stand-alone service. It enters pathways that already contain multiple actors and different forms of accountability.
A health plan may use analytics to identify an older member at increased risk of hospitalization. A hospital may use decision-support technology during diagnosis or treatment. A community clinician may receive an algorithm-generated risk flag. A home-care organization may use AI to optimize scheduling. A municipality may use data to understand population need. A technology company may develop predictive tools for falls, dementia or deterioration.
Each application may be technically distinct, but the operational challenge is similar: somebody has to decide what the information means and what happens next.
This is why AI in later-life care should be understood as part of wider health and social care coordination rather than merely as a technology initiative. A highly accurate prediction has limited value if the person who can act on it sits in another organization, does not receive the information or lacks the capacity to respond.
Israel's fragmented long-term care architecture makes that interface particularly important. Health plans possess significant clinical information and deliver extensive community healthcare, while National Insurance long-term care benefits relate mainly to functional dependency and support needs. Municipal welfare services, families and providers may hold different information again. AI can potentially strengthen coordination, but it can also reproduce existing boundaries if algorithms are developed around whichever dataset happens to be easiest to access.
Prediction matters only when it changes the pathway
One of the most attractive uses of AI in aging is prediction. Large datasets may reveal combinations of information that indicate rising risk before deterioration becomes obvious through conventional assessment.
Possible applications include identifying increased likelihood of:
- hospital admission or emergency department use;
- falls and functional decline;
- medication-related harm;
- frailty progression;
- readmission following discharge;
- caregiver breakdown; or
- increasing need for home or community support.
These possibilities connect strongly with Israel's existing digital-health capabilities. But predictive performance is only the first part of the intervention.
If a system identifies 500 older people as having increased hospitalization risk, a health plan still needs to determine what different response those people will receive. Additional primary care review? Home nursing? Medication reconciliation? Rehabilitation? Social support? A geriatric assessment? Telephone contact?
Without that service pathway, risk stratification becomes an information exercise rather than a care intervention.
The principle is closely related to wider risk stratification and triage. Predictive models are useful when they change the sequence, intensity or timing of support in a way that can be evaluated.
Organizations examining similar implementation questions can use the Quality Improvement Action Plan Builder to structure the practical gap between identifying a risk and changing the operational response. It is not an Israeli regulatory instrument, but the discipline is directly relevant: evidence should show not only that a risk was detected, but that somebody acted and that the action improved the pathway.
Operational scenario: an algorithm identifies rising frailty
Consider an 84-year-old woman living independently in central Israel. She has several chronic conditions and receives healthcare through her health plan, while her daughter provides most day-to-day assistance. Over several months, health-plan data show more frequent primary care contacts, reduced attendance at routine appointments and two recent falls.
An AI-supported risk model identifies her as having a significantly increased probability of hospital admission within the next several months.
The weak response would be to place a flag in an electronic record and assume that prediction itself constitutes prevention.
A stronger model connects the flag to a defined community pathway. A nurse reviews the information, contacts the woman and identifies that she has become less confident walking outdoors. Medication review reveals treatment that may be contributing to dizziness. Physiotherapy and functional assessment are arranged, and the family is given clearer information about what changes should trigger further contact.
The important evidence is not that the algorithm was correct in predicting risk. The important evidence is whether intervention changed the person's trajectory.
If similar cases repeatedly show that certain risk flags do not produce useful action, the model or pathway needs refinement. Governance should therefore review false positives, missed deterioration, professional response rates and actual outcomes rather than treating algorithmic accuracy as the sole measure of success.
AI can support clinical judgment without replacing it
Healthcare AI is often framed as decision support because many of the strongest applications do not make an autonomous final decision. Instead, they analyze information, identify patterns, classify images, summarize records or generate recommendations that a clinician considers alongside other evidence.
This distinction is central to the Israeli Ministry of Health's emerging regulatory approach. The Ministry has published guiding principles for AI and machine-learning medical technologies and principles for evaluating AI-based interventional clinical trials, while its regulatory sandbox is designed to support supervised testing of innovative healthcare AI. [oai_citation:1‡Gov.il](https://www.gov.il/en/Departments/General/digital-medical-technology-gmlp-1?utm_source=chatgpt.com)
For older people, human interpretation is particularly important because complexity often reduces the usefulness of a single predictive signal. Frailty, multimorbidity, dementia, polypharmacy, functional limitation, social circumstances and personal preferences interact.
An algorithm may identify a treatment as statistically beneficial but not understand that the person's priority is avoiding sedation so they can continue caring for a spouse. It may identify increased fall risk but not know that the person values continuing to walk independently and accepts some degree of risk. It may classify somebody as medically stable while missing a rapidly deteriorating caregiving situation at home.
Responsible AI therefore requires professional judgment capable of integrating the machine-generated recommendation with context that may not exist in the dataset.
Administrative AI could release capacity for human care
Some of the most immediately useful applications of AI may be less dramatic than clinical prediction.
Health and long-term care systems devote substantial workforce time to documentation, scheduling, information retrieval, coding, correspondence, referral administration and repetitive data entry. AI can potentially reduce some of that burden.
For an aging society, the opportunity is important. Workforce capacity is constrained not only by the number of professionals and care workers available, but by how much of their time is consumed by work that does not require their full skills.
AI may assist with summarizing long records before an assessment, drafting routine correspondence, extracting information from documents, supporting scheduling or identifying incomplete administrative workflows.
This links with wider workforce innovation and role redesign. The strongest objective is not to replace professional roles but to shift scarce human time toward activities in which empathy, physical presence, clinical reasoning and relationship matter.
That benefit cannot be assumed, however. Poorly implemented systems may simply create a new checking burden. If every AI-generated summary needs extensive correction, or staff must operate several disconnected platforms, workload can increase rather than fall.
AI-generated documentation requires accountability
Generative AI creates a particularly important governance issue because it can produce plausible text that contains errors.
In aging and long-term care, inaccurate documentation can have practical consequences. A generated summary might omit a medication allergy, misstate whether a person lives alone, confuse a historical diagnosis with a current one or present uncertain information as established fact.
Human review therefore cannot become ceremonial. If a professional signs an AI-generated clinical or care record, responsibility for the final record remains human.
Organizations need clarity about which uses of generative AI are permitted, what information can be entered into particular systems, how outputs are checked and which decisions require direct professional assessment.
The distinction is especially important where staff begin using easily available general-purpose tools rather than systems procured and governed by their organization. Sensitive health and care information should not be transferred casually into external AI platforms whose data practices have not been assessed.
The governance issue therefore extends beyond AI performance into data governance and information accountability.
Bias in aging care can hide inside apparently neutral data
Algorithmic bias is sometimes discussed as though it results only from flawed programming. In practice, bias can emerge from the data used to build a model, the populations represented in that data, the outcome selected for prediction and the way a prediction influences later decisions.
This matters greatly in Israel because the older population is diverse. Differences in language, socioeconomic circumstances, geography, cultural patterns, health-service use and digital engagement can all influence what appears in administrative and clinical datasets.
If an AI system learns mainly from people who use digital health services frequently, it may perform differently for people who engage less through digital channels. If healthcare utilization is used as a proxy for need, a population with historically weaker access may appear to have lower need simply because fewer services were received.
A model can therefore reproduce inequality without containing any explicitly discriminatory rule.
This connects AI implementation directly with data-led equity planning. Validation should examine whether performance differs across relevant groups rather than relying only on an overall accuracy measure.
The governance questions include whether false-negative and false-positive rates vary meaningfully by age, sex, population group, language, geography or other relevant characteristics, and whether the consequences of those errors are equally serious.
Operational scenario: prediction works differently across populations
A health organization introduces an AI model designed to identify older adults at increased risk of deterioration so that community teams can intervene earlier.
Overall evaluation shows good predictive performance. Several months later, however, clinicians working with a minority population report that people they consider highly vulnerable are not being flagged at the expected rate.
Further analysis shows that the model relies partly on patterns of previous healthcare use. Members of this population historically use some preventive and specialist services less frequently, partly because of access and language barriers. The model has therefore learned that lower recorded service use indicates lower risk when in this population it can sometimes indicate unmet need.
The response is not simply to abandon AI. The model is re-evaluated using stratified performance measures, community professionals contribute contextual information and alternative indicators of vulnerability are tested.
Governance also changes. Future models cannot proceed from global accuracy results alone. Performance across relevant population groups becomes part of routine validation.
The scenario illustrates a fundamental principle: fairness cannot be assessed only by examining the algorithm's intentions. It has to be assessed through the distribution of its errors and consequences.
Older age itself can become an inappropriate proxy
AI can also reinforce ageism if age is treated as a convenient predictor without sufficient attention to individual function, preferences and potential benefit.
Chronological age is often correlated with health risk, but people of the same age can have profoundly different capabilities and priorities. A system that increasingly deprioritizes interventions merely because age predicts poorer average outcomes may inadvertently transform statistical association into discriminatory practice.
In long-term care, functional status is often more relevant than age alone. The question is not simply how old somebody is, but what they can do, what support they need and what outcomes matter to them.
This matters especially where AI is used for triage or resource prioritization. Efficiency objectives can create pressure to rank people, but those rankings need ethical and professional oversight.
The strongest systems therefore distinguish legitimate clinical risk prediction from inappropriate assumptions that older age automatically reduces the value of intervention.
Explainability has to be meaningful to the decision being made
Not every AI system can explain its reasoning in a way that resembles a human decision. Some complex models identify statistical relationships that are difficult to reduce to a simple narrative.
That does not mean every system must be rejected. It means the required level of explainability should reflect the significance of the decision.
An algorithm used to suggest more efficient staff scheduling raises different concerns from one used to influence clinical treatment, eligibility assessment or a decision about whether somebody can safely remain at home.
Where AI materially influences a high-impact decision, professionals need enough understanding to recognize the model's purpose, inputs, limitations and circumstances in which its recommendation should be questioned.
Older people and families do not necessarily need technical explanations of machine-learning architecture. They need understandable explanations of the decision affecting them: what information was considered, whether automated analysis played a role, who made the final decision and how concerns can be raised.
That connects with wider trust, transparency and ethical data use.
Human oversight must have the power to change the outcome
“Human in the loop” can become meaningless if the human is expected simply to approve whatever the technology recommends.
Real oversight requires authority, time and competence to disagree.
If professional performance systems subtly penalize staff for overriding algorithmic recommendations, or if caseloads are so high that accepting the recommendation is always the fastest option, formal human review may coexist with practical automation.
Organizations therefore need to examine behavior as well as policy. How often are AI recommendations overridden? Which professional groups challenge them? What happens when an override occurs? Are some workers reluctant to question the system because they believe the algorithm is assumed to be superior?
These questions are part of clinical governance and accountability, not merely digital governance.
The Governance Maturity Assessment can help organizations structure similar questions about decision rights, oversight and escalation. It does not define Israeli AI requirements, but it can help leaders test whether responsibility remains clear when technology becomes embedded in operational decisions.
Privacy is foundational because AI depends on data
AI systems become more capable when they can access large quantities of relevant information. In health and aging services, however, those datasets may contain some of the most sensitive information a person has.
Israel's Privacy Protection Authority is responsible for protecting personal information held in digital databases, and the broader framework includes the Privacy Protection Law and associated data-security obligations. Privacy-enhancing technologies are increasingly being considered as part of responsible AI governance in Israel. [oai_citation:3‡Gov.il](https://www.gov.il/en/departments/the_privacy_protection_authority?utm_source=chatgpt.com)
The practical question is not simply whether an organization possesses lawful access to information. AI development creates additional issues around purpose, reuse, linkage, model training and retention.
Data originally collected for direct healthcare may become valuable for model development. Information from several sources may become far more revealing when combined. De-identification reduces some risks but does not automatically eliminate re-identification concerns where datasets are rich and highly granular.
Older people should not be treated as a passive source of training data simply because their health system has accumulated extensive longitudinal records.
Responsible implementation requires clarity about why information is being processed, what governance applies, how access is controlled, what safeguards are used and whether the proposed data use is proportionate to the intended benefit.
Cybersecurity becomes a patient-safety and care-continuity issue
AI adoption also expands the digital attack surface of organizations.
Systems may depend on cloud infrastructure, external vendors, application programming interfaces and large datasets. A cybersecurity failure can therefore affect not only confidentiality but continuity of service.
If a clinical decision-support system becomes unavailable, staff need a safe fallback. If an AI-enabled home-care scheduling platform fails, visits still need to occur. If models rely on compromised or corrupted data, incorrect outputs may become a safety issue.
This means AI resilience should sit within broader operational continuity rather than a separate technical plan.
Israel's national cybersecurity environment gives this issue particular salience. The country's 2025 national cybersecurity strategy emphasizes growing digital dependency and the need to strengthen resilience, while healthcare remains an especially sensitive domain. [oai_citation:4‡Gov.il](https://www.gov.il/BlobFolder/news/cyber_strategy_2025/he/israel_national_cybersecurity_strategy_feb2025.pdf?utm_source=chatgpt.com)
Organizations considering major AI deployment can use the Digital Transformation, AI and Cybersecurity Readiness Assessment to examine whether digital ambition is matched by organizational controls, workforce competence and continuity arrangements.
AI can support the long-term care workforce, but workforce effects need monitoring
Israel's long-term care workforce includes healthcare professionals, rehabilitation staff, social workers, Israeli care workers, foreign caregivers and substantial unpaid family care. AI will affect these groups differently.
Clinical professionals may gain decision support and record summarization. Managers may gain better demand forecasting. Home-care organizations may optimize scheduling. Family caregivers may receive automated prompts or deterioration alerts.
But efficiency in one part of the system can shift workload elsewhere.
A predictive model that identifies more high-risk people may create additional community nursing demand. Remote monitoring may reduce routine visits while generating more alerts for family caregivers. Automated scheduling may improve organizational efficiency but create fragmented working patterns that workers dislike.
Workforce impact should therefore be measured across the whole pathway.
One of the strongest uses of AI may be capacity planning. Combining demand, acuity, geography and workforce data could help organizations anticipate where staffing pressure is likely to emerge. The Digital Twin Scenario Modeler provides a structured way to explore similar relationships between workforce capacity, quality and service stability. It is not an Israeli workforce-planning system, but the principle is relevant to an aging society in which care demand and labor availability need to be considered together.
Operational scenario: AI scheduling improves efficiency but damages continuity
A home-care organization introduces an AI scheduling platform intended to reduce travel time and improve use of available caregiver hours. The system analyzes worker availability, geography and visit requirements and produces schedules more quickly than manual planning.
Initial performance looks positive. Travel time falls and a greater proportion of contracted hours is filled.
Several months later, however, older people and families begin reporting more frequent changes of caregiver. The algorithm has optimized geographic efficiency but has not given enough weight to continuity of relationship.
For some recipients, this is inconvenient. For people with dementia, communication difficulties or highly personal care needs, it becomes a quality issue. Workers also report that schedules sometimes look efficient on a map but underestimate the complexity of particular visits.
The organization changes the model. Continuity is added as a weighted outcome alongside travel and utilization. Certain people are designated as requiring smaller caregiver teams. Staff can flag visits where algorithmic assumptions do not match operational reality.
The organization also broadens its dashboard. Instead of measuring only filled hours and travel reduction, it monitors caregiver continuity, missed visits, complaints, worker feedback and outcomes.
The lesson is that AI optimization inevitably reflects the objectives it is given. If cost and efficiency are the only objectives, the system may become highly effective at producing the wrong version of care.
Evaluation needs to move beyond technical accuracy
Healthcare AI is often evaluated through sensitivity, specificity, predictive value or other technical performance measures. Those measures matter, particularly for clinical applications, but they do not tell decision-makers whether implementation improved care.
A mature evaluation framework should examine several layers:
- technical performance and reliability;
- clinical or operational usefulness;
- differences in performance across population groups;
- human response to recommendations;
- effects on workforce workload;
- older-person and caregiver experience; and
- actual outcomes, including unintended consequences.
This aligns with Israel's wider work on evaluating digital-health technology. The Ministry of Health has previously developed health-technology assessment guidance for digital interventions, emphasizing the need to evaluate implementation and value rather than focusing solely on technological novelty. [oai_citation:6‡Gov.il](https://www.gov.il/BlobFolder/generalpage/digital-health-guide-062021/he/files_publications_digital_health_digital-health-hta-062021.pdf?utm_source=chatgpt.com)
AI deserves the same discipline.
An algorithm that predicts hospital admission accurately but does not reduce admissions may still have research value, but its service value is limited. A documentation tool that saves clinicians five minutes while introducing frequent inaccuracies may create unacceptable risk. A scheduling model may save money while worsening continuity.
Value is therefore multidimensional.
Governance should follow AI throughout its lifecycle
AI governance cannot end when a technology is procured.
Models may behave differently after deployment because the population changes, clinical practice changes or data patterns shift. Vendors may update systems. Generative models may change substantially between versions.
Organizations need ongoing oversight of:
- what AI systems are in use and for what purpose;
- who owns the associated risk;
- which datasets and external suppliers are involved;
- how performance and bias are monitored;
- how incidents and near misses are investigated;
- how material model changes are reviewed; and
- when a system should be restricted or withdrawn.
This is where traditional quality governance and technology governance need to converge.
An AI-related safety event should not remain solely with an IT department if it affected clinical or care decisions. Likewise, a privacy or cybersecurity concern cannot be separated from service quality where loss of trust changes people's willingness to use the technology.
Organizations can use the Quality Dashboard Builder to structure a broader evidence set that brings technology measures together with safety, experience, workforce and outcome indicators.
Generative AI raises a different category of challenge
Much healthcare AI has historically been designed for a relatively narrow purpose: detect an image feature, calculate a risk, classify a signal or support a specific workflow.
Generative AI is different because it can produce open-ended language, summaries, recommendations and conversational responses.
This creates opportunities in aging care. A multilingual conversational system might help people navigate services. AI could simplify complex information, draft accessible explanations or help professionals synthesize lengthy histories.
But fluent language creates a specific risk: people may attribute understanding or authority to a system that is generating statistically plausible text.
That matters particularly for older people who may be seeking health advice, navigating eligibility or making emotionally difficult decisions. A conversational AI should not quietly become an unregulated substitute for professional assessment simply because it feels easier to talk to.
The Ministry of Health's current description of digital mental-health applications offers a useful principle beyond mental health: AI tools may support assessment or intervention, but professional supervision and mechanisms for escalating to human care remain important. [oai_citation:7‡אתרי בריאותי](https://me.health.gov.il/en/mental-health/information-and-updates/mental-health-care/digital-care/?utm_source=chatgpt.com)
The transferable principle is human availability. AI should make professional support more reachable rather than create a digital barrier between the person and a professional.
Older people should participate in the design of AI systems affecting them
Responsible AI cannot be achieved only through technical committees.
Older people and caregivers can identify problems that developers and administrators miss. They may explain that a question is confusing, that an alert feels stigmatizing, that a technology creates anxiety or that an apparently efficient pathway removes valued human contact.
Participation should therefore occur before deployment as well as through complaints afterwards.
Co-design can test accessibility, language, trust, consent and the practical acceptability of proposed automation. It can also reveal different preferences. Some older adults may welcome predictive monitoring; others may regard it as intrusive.
There will rarely be one universal threshold of acceptable automation.
That is why person-centered AI should preserve meaningful alternatives where possible. Digital efficiency should not become compulsory digital participation.
Israel's innovation strength creates both opportunity and responsibility
Israel has a substantial healthcare AI ecosystem. A 2026 Israel Innovation Authority review identified hundreds of Israeli healthcare companies using forms of automation or AI-supported decision-making, reflecting the depth of technological activity in the sector. [oai_citation:8‡רשות החדשנות](https://innovationisrael.org.il/en/ai-in-healthcare/?utm_source=chatgpt.com)
The country is therefore well placed to test how AI can move beyond isolated innovation into routine care.
But precisely because development capacity is strong, Israel also has an opportunity to demonstrate that responsible implementation is part of innovation rather than an obstacle to it.
A technically impressive product that cannot demonstrate fairness, workflow fit, security or measurable benefit is not mature simply because its underlying model is sophisticated.
The difficult transition is from demonstration to dependable infrastructure.
That challenge connects directly with pilot evaluation and learning loops. Real-world deployment should generate evidence capable of changing the product, the operating model or the decision to scale.
What international systems can learn from Israel
Israel's institutional architecture cannot be copied directly. Its health plans, national health insurance arrangements, National Insurance long-term care benefits and particular technology ecosystem are specific to the country.
The transferable lesson lies in the relationship between data infrastructure and service design.
Countries may spend heavily developing predictive tools without first defining what happens after a prediction. They may procure AI to relieve workforce pressure without measuring whether work is simply transferred elsewhere. They may assess average model accuracy while missing unequal performance across communities.
Israel's emerging experience highlights several principles with wider relevance: AI should enter a defined care pathway; human decision rights must remain real; performance should be tested across populations; administrative efficiency should not displace person-centered outcomes; and technology governance needs to connect with ordinary quality governance.
The model itself is only one component of responsible implementation.
From AI adoption to responsible practice
The next stage of AI in Israeli aging and long-term care should therefore be judged less by how many organizations use artificial intelligence and more by the maturity with which it is governed.
That means distinguishing low-risk administrative assistance from systems influencing clinical or care decisions. It means validating algorithms among the populations who will actually be affected. It means ensuring that older people know when automated analysis is materially shaping decisions about them.
It also means measuring outcomes after deployment.
A responsible AI system should be able to answer not only “Did the algorithm perform as designed?” but “Did people receive better care because of it?”
That second question is harder. It is also the one that matters.
Conclusion
Artificial intelligence could become an important part of Israel's response to population aging. The country's digital-health infrastructure, longitudinal data, technology capabilities and emerging healthcare AI governance create significant opportunities to identify deterioration earlier, support professional judgment, reduce administrative burden and plan services more intelligently.
Yet aging and long-term care expose the limits of technology-first thinking. Older people rarely have single problems, and the most important decisions frequently involve trade-offs between safety, independence, health, relationships and personal preference. Those decisions cannot be reduced safely to prediction alone.
The stronger direction for Israel is therefore responsible augmentation rather than automated substitution. AI should help professionals see relevant information sooner, help organizations deploy scarce capacity more intelligently and help older people access support more effectively. It should not make opaque decisions about people's lives simply because those decisions can be calculated.
Implementation will determine whether AI strengthens or fragments care. That requires clear decision rights, population-specific validation, privacy and cybersecurity safeguards, meaningful human oversight, transparent escalation and evidence about real outcomes. If those disciplines develop alongside technological capability, Israel can move from experimenting with AI to integrating it responsibly into the everyday architecture of aging and long-term care.