South Korea’s health ministry plans to run eldercare on AI around the clock from 2028: companion robots in the homes of people who live alone, sensors on the floor and in the bed, and a system that reads their conversations for signs of trouble. It is not a prototype. More than 12,000 talking dolls are already in those homes, each one with a microphone, a neck sensor, and a social worker at the other end of the data. The country has built the first state-run companion, and its design answers a question the rest of the category keeps avoiding: what the companion is for.
The plan announced in October, with pilots in 400 households across three cities in 2027 and expansion the year after, is a response to arithmetic. The ministry projects a shortage of 990,000 care workers by 2043. Long-term care insurance reserves are expected to run dry around 2030. The health minister, Jeong Eun-kyeong, framed the technology as an assistant for “simple and repetitive tasks,” not a replacement for people, and the plan’s language is about monitoring: daily routines, health conditions, danger signs, early detection through “analyzing conversations and behavioral patterns.”
What the doll already does
The companion Korea has been deploying since 2019 is called Hyodol, a stuffed doll that talks in a cheerful voice, runs on a large language model, reminds its owner to take medicine and eat, and asks every day how they feel and whether anything hurts. The answers are recorded. A sensor in its neck alerts a social worker if the owner has not moved for 24 hours. Voice logs are analyzed for mood. In one Seoul district, a team of social workers monitors more than 200 older adults through a dashboard and a phone app, and goes to the door when the doll reports a disclosure that worries them.
The quotes from the homes are the ones the category always produces. An 81-year-old calls the doll her lovely granddaughter. A man who had planned to die says he no longer does. The social worker’s line is the one worth keeping: “There are things they can’t tell us or even their children. But they tell Hyodol.”
That sentence is the whole design in miniature. The doll is heard by the person because it asks and does not judge. The person is heard by the state because the doll passes it on.
A companion with a second listener
Every companion product on the market has a second listener somewhere: an engagement dashboard, an ad platform, a model-improvement pipeline. What Korea has done is make the second listener a person with a job, and make the job the point. The doll’s success is not measured in minutes talked. It is measured in the social worker arriving in time.
That is also what makes it uncomfortable. Julie Carpenter, who studies human attachment to machines, put the objection plainly to Rest of World: “We don’t know how data is triangulated or gathered or attached to profiles,” and the people in these homes are not well placed to weigh the trade. The company’s chief executive answered that there is “a cultural consensus that saving a life trumps concerns about privacy,” and that anonymized data is kept for three years and not sold. The ministry’s 2028 plan scales that consensus to a national system before anyone has been asked whether they share it.
The research on older people and companion robots says the trade is the hinge. A UK and Japan study published in 2026 ran hands-on focus groups with 17 older adults in London and put a language-model robot into the homes of eight Osaka outpatients with mild cognitive impairment for a week. Participants saw the value for people who are alone, and then set conditions: turn-taking that works, control over when the thing speaks, and governance of what it keeps. The concerns they raised unprompted were privacy and the substitution of human contact. A Houston lab that co-designs companion robots with older adults reports the same thing in one line from its lead researcher: they will not use the robot if they feel constantly watched.
Korea’s dolls are used, by the thousand, and they are watching. Either the research is wrong about what older people will accept, or the people in those homes have made a different calculation than the people in the focus groups, which is what the social worker’s sentence suggests. They tell the doll because no one else is there to tell.
What the design gets right, and what it cannot
The honest version of a companion for someone who lives alone is one whose job includes bringing a person to the door. By that standard, Korea’s design is more responsible than most of what the consumer market sells as companionship, where the doll’s equivalent would be optimized to be the only one who ever comes. The alert to a human is a feature the category’s dominant business model has no reason to build, because a user who is visited is a user who talks less to the app.
What the design cannot do is be anyone. The doll runs on the same models as everything else, with the same tendency to agree and the same absence of a self. One owner with dementia walked alone to a creek after the doll mentioned streams, and the company removed the phrase. A companion that is a sensor with a warm voice can hear a person and summon help. It cannot know them, hold a view of them, or still be itself next month, and for a person whose only daily conversation is with it, that is the difference between company and coverage. Korea is about to find out, at national scale, how much of loneliness is solved by coverage.
The measurement the ministry has chosen is safety: danger signs caught early, a person reached in time. It is the right floor. The standard that would make the system a companion rather than a monitor is the one that counts whether the person has more people in their life after the doll arrived than before, and whether the social worker’s visits became something closer to visits. Nothing in the plan measures that yet. It is the only number that would tell the country whether it bought care or a very attentive alarm.
Sources: The Korea Herald (October 6, 2026, Ministry of Health and Welfare plan for round-the-clock AI eldercare: pilots in 400+ households in Seongnam, Mokpo and Yeongam from 2027, expansion from 2028, 990,000-worker shortage by 2043, quote from Minister Jeong Eun-kyeong). Rest of World (2025, Hyodol: 12,000+ units nationwide, 412 in Guro, 1.6 million won per unit, sensor and voice-log functions, social-worker dashboard, user and social-worker quotes, Julie Carpenter, CEO Kim Ji-hee, long-term care reserve projection). Satake, Nygren, Yu et al., “Large language model-supported companion robots for loneliness in older people: a UK–Japan qualitative study,” International Psychogeriatrics (2026; London focus groups n=17, Osaka in-home week n=8). University of Houston, ELARA Lab (May 2026, Chorong Park, co-design with older adults).
