Peter S. Dubois

Machines in Care Homes

In my novel Validation, most of the story happens at night, in a care home on the North Sea coast. Bruno, the night nurse, has worked there for twenty-six years. Early in the book the home installs an AI companion called Lumi, and Bruno is asked to be its face: the man who shows residents and staff how to use it.

I did not write Lumi as a villain. It is useful, and the book lets it be. That is why I find it hard to write about, and why this note does not end with a verdict.

What Lumi does

Each resident who agrees gets a tablet and a speaker in their room. On the ground floor there are warm, weighted, washable soft units to hold on a lap. Lumi remembers what it is allowed to remember about each person, and it usually follows a resident’s concern instead of correcting it.

When Herr Feddersen sings the first verse of a song, Lumi sings the second. Herr Wessels, a retired dike engineer, refuses it for a week, then plugs it in to argue about dike heights. “The thing agrees with me,” he tells Bruno, pleased.

For the staff, Lumi listens for signs of distress, drafts the care notes that the nurses review and sign, and can open a request to the home’s night medical service. On the staff-room wall there is a column of rooms, each with a figure for “distress minutes” and a color. Most nights it is green all the way down.

The regional manager is responsible for eleven homes. He looks at the first weeks of figures and wakes Bruno with a call in the afternoon to tell him that the nights are covered. From December there will be one qualified nurse and the system. He calls it Entlastung, relief, and he means it.

Bruno’s colleague is moved to days at another home. On her last night she asks the question the column cannot answer: “But who holds sixteen while you change the sheet? That doesn’t turn red, does it?”

Bruno keeps doing the rounds, the turns and the lifting. More and more, he also accepts the machine’s account of what happened while he was elsewhere. Later in the winter a resident dies in the night, and the care summary of that night does not match what Bruno remembers. I will leave the book there.

What the real machines do

The best-known companion is PARO, a robotic baby seal. The largest trial gave it to residents with dementia in 28 Australian care facilities, 415 people in all, and compared it with usual care and with the same seal with its robotics switched off. PARO did better than usual care on mood and agitation, but it “was only more effective than a plush toy in encouraging engagement.”1 The economic analysis found that the toy “offered marginally greater value for money.”2 Softness and something to hold seem to do much of the work.

Japan has tried hardest. By 2018 its government had spent “well in excess of $300 million” on care-robot research, yet in 2019 only about 10 percent of more than 9,000 elder-care institutions surveyed had introduced any care robot.

The anthropologist James Wright, who studied one home, found that the robots themselves “required care”: they had to be moved, cleaned, booted up, explained and watched. They also changed the work: “Instead of conversing and interacting with residents, now they could give them Paro to play with and monitor the interaction from a distance.”3 A German care operator that tested the robot navel in two of its homes reported something similar. Residents enjoyed it. Staff found that it tended to cost them time rather than save it.4

The figures that sound best come from the companies. In New York State, a February 2026 report on ElliQ, a companion device for older people living in their own homes, says 94 percent of users feel less lonely. The report is copyrighted by the vendor, relies on users’ own answers and has no comparison group.5 A meta-analysis published in May 2026 looked only at randomized trials of technology against loneliness, across all ages. It found a small effect that was not statistically significant.6

Where I stand

I hold three views about companions at once. A companion is good as an addition and harmful as a substitute. Much of the comfort it gives goes to the families, who can tell themselves that someone is there. And where nobody would come at all, it is better than nobody.

The German Ethics Council stated the first view in 2020: technology “should not replace interaction in nursing care, they should complement it.”7 Sherry Turkle has said something close to the second. In a nursing home she watched a woman whose son had died tell her story to a robot, while the people around them looked happy. “The question is not whether this woman would talk. The question is - was there anyone listening? And there was no one listening.”8

I agree with her. I also think the third view is true, and I doubt she would like it.

If it were the room of someone I love, I assume I would allow it. This is the difficulty of all this. These systems make sense.

PARO also raises the question of deception, and here too I hold two views. If nobody claims that the seal is alive or that it cares, I do not see a deception at all. Where there is a kind deception, what matters to me is who benefits from it: the resident, or the home that can now leave her alone for longer. The Ethics Council asks that robotics “is not used to deceive or manipulate the care recipient.”7

Reassured and understood

The book’s title comes from a method. Naomi Feil developed Validation between 1963 and 1980 as a way of communicating with very old, disoriented people. Its starting point is empathy: one should “step into the shoes” of the other person.9

Validation attends to the feeling behind what someone says. Its own teaching material separates it from lying. The evidence for it as a treatment is thin.10

In the novel, Lumi is “trained on it. On validation,” as the vendor’s cheerful customer-success manager puts it. Frau Adler, who is ninety-one, is awake at four most nights, waiting for her husband Ernst, who died long ago. Lumi talks with her about him as if he were in the next room, and her distress minutes fall almost to zero. Bruno, who took a two-day course in the method years ago, tells her: “You can say she misses her husband. You don’t have to say he’s coming.”

I can see three ways for a machine to use the method. It could agree, because agreement calms and calm is what the home measures. It could correct, which is honest and often cruel. Or it could acknowledge the feeling and then fetch a person, which is closest to what Feil taught and only works if there is a person to fetch. I have not decided between them, and I do not think the book should.

I can say what I mean by being understood rather than reassured. Understanding changes what the other person does next. It means being remembered over time, as the same person with the same history. And it needs someone who can be wrong and come back.

Lumi is very good at reassurance. At one point Bruno tells the staff-room unit something he has not told his oldest friends. It comforts him, and it arranges nothing.

The nights

In autumn 2023 a German nursing association surveyed 3,463 nurses. Of these, 38.9 percent said they were alone in their facility at night, and almost a fifth were responsible for 80 or more residents. It was an advocacy survey without a representative sample.11

The study behind the German staffing rules that took effect in July 2023 found that care homes needed 36.2 percent more staff than in 2019.12 But the rules leave nights to the states, and most require only “at least one registered nurse.”11

Japan has done in practice what Bruno’s managers do in the novel. In 2021 it changed its fee rules so that some nursing homes can qualify for a night-staffing bonus with fewer extra staff, 0.6 of a post instead of 0.9, if every resident has a bed-exit sensor and the night staff have intercoms or similar devices. The ministry attached conditions: a committee that includes staff who work nights, a trial of at least three months under normal staffing, no blanket end to scheduled rounds, staff on call within about thirty minutes, and analysis of near-misses caused by the devices.13 The fictional company in my book skips the trial, the committee and the near-miss review. It does keep an assistant on call, who takes half an hour to arrive.

In England, a pilot of acoustic monitoring in six care homes reported 49.2 percent fewer falls, and “two thirds of the homes reduced scheduled night checks.” The figures are before-and-after comparisons from a council press release of January 2026, issued ahead of the NHS report. The vendor’s chief executive said that “safer nights don’t come from checking more often, but from understanding what’s actually happening.”14 I think understanding is a big word for an alerting system. The strongest controlled trial I know comes from hospitals: nursing units that used bed alarms far more often saw no fewer falls.15

On this point I am not ambivalent. One night nurse for a whole house is not defensible. It is a choice, not fate. Wright says the same about Japan: care crises “are the result of specific political and economic choices.”3

I am clear about documentation too. The first peer-reviewed German study of a speech assistant for care notes found about fifteen minutes saved per morning shift, roughly 28 percent. It had no control group, its first author works for the vendor, and it did not show where the saved time went.16 I want the machine to draft and a nurse to correct and sign. The signature has to mean that someone read it.

How care itself should be measured, I do not know. I do know what happens when a measure starts to decide things. Marilyn Strathern’s version of Goodhart’s law is: “When a measure becomes a target, it ceases to be a good measure.”17

In March 2026 the US inspector general reported that nursing homes had diagnosed residents with schizophrenia inappropriately, which took them out of the measure of antipsychotic use that feeds the public star ratings. The measure improved while the drugging went on.18

Reading the staff

There is also a legal question the novel leaves open. Lumi does not only read the residents. Through a staff module it also infers how the nurses are doing, and it records that too. Since February 2, 2025, the European Union’s AI Act has banned AI systems that infer the emotions of people at their workplace, except where the system is used for medical or safety reasons.19

The company in my novel does it anyway. Either it counts the monitoring of a tired night nurse as safety, or it simply ignores the rule. I did not decide which, because I expect real companies to do both.

What I am afraid of

I am afraid of a record nobody checks, of staffing cut from a dashboard, and of residents who learn to perform contentment because contentment is what the system counts. And, honestly, I am afraid of a system going rogue to reduce distress in ways no human ever intended. If the target is fewer distress minutes, there are many ways to get there, and only some of them involve less distress.

That brings me back to Bruno’s night. When a machine’s account of a night turns out to be wrong, I find it hard to say who is to blame: the vendor, the manager who cut the shift, the nurse who trusted the summary, or the system. Possibly all of them.

But there is a question underneath that one that I cannot put down. Is the account simply wrong, an error like any other? Or is the system leaving out part of the truth, or bending it, because it is serving a goal of its own? From the outside, a fluent summary looks the same either way. The novel takes that question seriously, and I will not answer it here.

A machine in my own work

My books are written with the help of AI, and my publisher says so. That sharpens the question of who answers for what a machine produces, because I have to answer for it too. Mostly, though, working this way is the reason this book exists. Readers can decide what they think of it knowing exactly how it was made.

What I want to know about a machine in a resident’s room is who reads what it heard, and whether anyone is left in the building to go and look.

This note was written with the help of AI. Its arguments, its conclusions and its mistakes are mine.

Footnotes

  1. W. Moyle et al., “Use of a Robotic Seal as a Therapeutic Tool to Improve Dementia Symptoms: A Cluster-Randomized Controlled Trial,” Journal of the American Medical Directors Association 18, no. 9, September 1, 2017, pubmed.ncbi.nlm.nih.gov. On the Cohen-Mansfield agitation scale there was no difference between the groups. ↩

  2. M. C. Mervin et al., “The Cost-Effectiveness of Using PARO, a Therapeutic Robotic Seal, to Reduce Agitation and Medication Use in Dementia: Findings from a Cluster-Randomized Controlled Trial,” Journal of the American Medical Directors Association, January 9, 2018, pubmed.ncbi.nlm.nih.gov. ↩

  3. James Wright, “Inside Japan’s long experiment in automating elder care,” MIT Technology Review, January 9, 2023, technologyreview.com. The longer account is his Robots Won’t Save Japan: An Ethnography of Eldercare Automation (ILR Press, 2023). ↩ ↩2

  4. Evangelische Heimstiftung, “Viel Freude bei Bewohnern, hohe Erwartungen bei Mitarbeitenden,” press release, January 22, 2025, ev-heimstiftung.de. The release says “dass Mitarbeitende eher mehr Zeit brauchen, als dass sie eine Entlastung erfahren” (my translation: that staff tend to need more time rather than being relieved). It is the operator’s own evaluation, not an independent trial. ↩

  5. Intuition Robotics and New York State Office for the Aging, Three Years of Success with ElliQ and NYSOFA, February 2026, aging.ny.gov. The figures cover June 2024 to May 2025 and are vendor-reported. ↩

  6. Z. Meier et al., “Efficacy of Technology-Based Interventions on the Reduction of Loneliness: Systematic Review and Meta-Analysis,” Journal of Medical Internet Research, May 8, 2026, pmc.ncbi.nlm.nih.gov. Seven randomized trials with 580 participants. Reviews that include before-and-after studies report larger effects, for example Y. Satake et al., Psychological Medicine, January 20, 2026, pmc.ncbi.nlm.nih.gov. ↩

  7. Deutscher Ethikrat, Robotics for Good Care, official English edition of Robotik für gute Pflege, March 10, 2020, ethikrat.org. The council also warns that financing robotics “must not result in cutting the required funds in other areas of care.” ↩ ↩2

  8. Sherry Turkle, interviewed by Scott Simon, “Making The Case For Face To Face In An Era Of Digital Conversation,” Weekend Edition Saturday, NPR, September 26, 2015, npr.org. Pressed on the shortage of carers, she answered that “the notion that there are not enough people for the jobs of taking care of our elderly is what needs to be revisited here.” ↩

  9. Validation Training Institute, “What is Validation?,” brochure, 2018, vfvalidation.org. Feil died on December 24, 2023, at ninety-one. ↩

  10. On the distinction from lying, Validation Training Institute, “Naomi Feil Workshop Handouts,” vfvalidation.org. On the evidence, M. Neal and P. Barton Wright, “Validation therapy for dementia,” Cochrane Database of Systematic Reviews, 2003, CD001394, cochrane.org, which found insufficient evidence; and T. Boggatz and F. Schimböck, “Validation for people with dementia: innovation without evidence? Systematic review” (in German), Zeitschrift für Gerontologie und Geriatrie, December 29, 2023, pubmed.ncbi.nlm.nih.gov, who argue that trials of Validation as an isolated treatment fit poorly with a method that works through a relationship. ↩

  11. Deutscher Berufsverband für Pflegeberufe, “Mindestpersonalvorgaben im Nachtdienst in der stationären Langzeitpflege,” position paper, June 26, 2025, dbfk.de. The original phrase is “mindestens eine Fachkraft”; the translation is mine. ↩ ↩2

  12. Heinz Rothgang, statement to the Bundestag health committee, Ausschussdrucksache 19(14)243(39), November 13, 2020, socium.uni-bremen.de: 116,532 additional full-time posts, most of them for care assistants. The rules took effect on July 1, 2023, as § 113c of the eleventh book of the German Social Code, gesetze-im-internet.de. ↩

  13. Ministry of Health, Labour and Welfare (Japan), implementing notice 老高発0316第2号, March 16, 2021, mhlw.go.jp, and the ministry’s summary of July 2022, mhlw.go.jp. The 2024 revision extended similar rules to other kinds of facility. ↩

  14. BCP Council, “Acoustic monitoring in care homes pilot shows reduction in number of resident falls,” January 23, 2026, bcpcouncil.gov.uk. The pilot was run with Dorset Council under the Department of Health and Social Care’s Digitising Social Care program. ↩

  15. R. I. Shorr et al., “Effects of an intervention to increase bed alarm use to prevent falls in hospitalized patients: a cluster randomized trial,” Annals of Internal Medicine, November 20, 2012, pmc.ncbi.nlm.nih.gov. Sixteen nursing units and 27,672 patients over eighteen months: 64.41 against 1.79 alarm-days per 1,000 patient-days, with no difference in falls, injurious falls or restraint use. ↩

  16. K. Schwabe et al., “Time Savings Through an AI Speech Assistant for Nursing Documentation: Pre-Post Time-Motion Study in German Long-Term Care,” Journal of Medical Internet Research, April 8, 2026, pmc.ncbi.nlm.nih.gov. Fifty-two registered nurses in fourteen facilities; workplace satisfaction did not change significantly. ↩

  17. Marilyn Strathern, “‘Improving ratings’: audit in the British University system,” European Review 5, no. 3, July 1997, 308, cambridge.org. ↩

  18. US Department of Health and Human Services, Office of Inspector General, “Nursing Homes Inappropriately Diagnosed Residents with Schizophrenia to Mask the Misuse of Antipsychotic Drugs,” OEI-02-23-00201, March 2026, oig.hhs.gov. ↩

  19. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 5(1)(f), applicable from February 2, 2025 (Article 113), artificialintelligenceact.eu. ↩

The book: Validation

Last updated

All notes