The Age Penalty: How AI Prices Older People Without Asking #
She is eighty-five, she lives in the United States, and every morning a small tabletop device with a swivelling head and a soft glowing base tells her it is time to take her medication. It suggests she stretch. It asks how she slept. It is called ElliQ, it is made by the Israeli company Intuition Robotics, and according to the opinion piece that opened with her story in the Korea Times on 19 August 2026, it helps her stay independent.
The author of that piece was Suh Chung-ha, a former Korean ambassador to Singapore and Hungary who now runs the ASEM Global Ageing Center in Seoul. His argument was not that the robot is bad. His argument was subtler and considerably more uncomfortable: that the same broad technological shift that produced her helpful companion is quietly reshaping her access to healthcare, long-term care, credit, insurance and public services, and that it is doing so using systems trained on data which encode decades of assumptions about what old people are worth.
Here is the thing that should keep you awake. There are two artificial intelligence systems in this woman's life. One of them sits on her side table, greets her by name, and was designed with people like her explicitly in mind. The other one has never been in her house, has no name, no face and no voice, and it is scoring her.
She chose the first. The second chose her.
Two Machines, One Household #
The asymmetry is almost perfectly clean, and it is worth stating precisely because it is the structural fact from which nearly everything else in this story follows.
The assistive machine is opted into. It arrives through a screening process, in the New York State case run by county and locality-based Area Agencies on Aging. It is visible: it occupies physical space, it announces itself, it is marketed at older people as a product for older people. If she hates it, she can unplug it and someone will come and take it away.
The discriminatory machine has none of these properties. She did not opt into the underwriting model that prices her travel insurance, the credit decisioning system that assessed her application for a modest overdraft extension, or the triage algorithm that sorted her referral. She cannot see them. She was not consulted in their design, and neither was anyone remotely like her. She cannot unplug them. In most cases she will never learn that a model was involved at all; she will simply receive an outcome, in a letter or on a screen, expressed in the passive voice.
Assistive AI for older people is a market with a sales pitch. Decisional AI about older people is infrastructure with a legal department.
Public debate about AI and ageing has been overwhelmingly captured by the first category. Companion robots photograph well. They allow governments with ageing populations and inadequate care workforces to point at something tangible. Meanwhile the systems that will actually determine whether an eighty-five-year-old can borrow money, buy cover for a trip to see her grandchildren, or get onto a surgical waiting list are being deployed with almost no public conversation about age at all.
Ninety-Five Per Cent of What, Exactly #
Start with the good machine, because the claims made for it deserve the same scepticism we would apply to anything else.
In August 2023 the New York State Office for the Aging announced that its ElliQ rollout had produced a 95 per cent reduction in loneliness among participants. The figure travelled fast, and has been repeated in trade press, vendor materials and policy briefings ever since. That 2023 release reported that more than 800 older New Yorkers were in the programme, that users interacted with the device more than thirty times a day, six days a week, and that over 75 per cent of those interactions related to social, physical or mental wellbeing. Greg Olsen, the agency's director, said the results were “truly exceeding our expectations”.
Thirty interactions a day is not a device gathering dust.
Now read what the same agency published in February 2026. Its year-three project update, covering the programme year from June 2024 to May 2025, reports that 94 per cent of clients say they feel less lonely, up from 93 per cent the year before, that 97 per cent report feeling better overall and 79 per cent feel more connected to the world around them, with customer satisfaction at 4.6 out of 5. The average client is 75. As of May 2025, 834 older adults had joined. More than 3,500 have applied.
Four applicants for every place.
Look at what happened to the sentence. In 2023 the claim was a 95 per cent reduction in loneliness: a delta, a measured change, phrasing that implies an instrument, a baseline and a follow-up. In 2026 it is 94 per cent of clients who say they feel less lonely. That is not a reduction. It is a self-report, and the agency now presents it as one.
The wording did not soften by accident.
Neither version was ever a clinical finding. Both are programme metrics derived from participants who were screened in, who wanted the device, and who were asked how they felt. There is no control group in those numbers. There is no validated loneliness instrument named alongside them in the public materials, no randomisation, no blinding, and no comparison against the obvious alternative intervention, which is a person visiting.
Note too how little the figure moves. Ninety-three per cent, then 94, across a self-selected cohort that grew by hundreds in between. Effect estimates wander when the sample changes; satisfaction scores do not. The 4.6 out of 5 sitting in the same document is from the same family of measurement.
The wider literature is more honest and less exciting. A systematic review and meta-analysis by Lihui Pu, Wendy Moyle, Cindy Jones and Michael Todorovic, published in The Gerontologist, screened more than two thousand articles and found thirteen from eleven randomised controlled trials, nine of which entered the meta-analysis. Social robots appeared to have positive effects on agitation, anxiety and quality of life, but the meta-analysis found no statistical significance. The authors flagged a relatively high risk of bias in allocation concealment and blinding, and concluded that firm conclusions were limited by the shortage of high-quality studies. A later meta-analysis published in the Journal of the American Medical Directors Association, focused on residents of long-term care facilities and drawing on eight trials, did report significant reductions in depression and loneliness with large effect sizes.
So: promising, contested, and nowhere near the confidence implied by a round 95.
None of this means ElliQ is useless. It plainly is not; the engagement figures describe a device in near-constant use, and 834 people have one while thousands wait. It means that the most widely circulated statistic about AI and older people is a self-reported satisfaction figure from a self-selected group, and that we have collectively decided to treat it as proof of concept for an entire policy direction.
The Awkward Protected Characteristic #
Now the harder half, and it requires conceding something that age-discrimination advocates often skate past.
Age is a genuinely strange thing to build fairness engineering around. Race and sex, as legal categories, are treated as characteristics that should almost never bear on how you are priced or assessed. Age is not treated that way, and not merely out of prejudice. Age correlates with mortality and morbidity in ways that are real, measurable and actuarially load-bearing. An insurer pricing life cover without reference to age is not being fair; it is being incompetent.
The law recognises this explicitly, and the contrast is sharper than most people realise.
When the Equality Act 2010 extended the ban on age discrimination to the provision of goods, facilities and services in April 2012, it carved out financial services. Schedule 3 permits providers to use age in connection with a financial service, provided that any risk assessment involving age is carried out by reference to information relevant to the assessment and from a source on which it is reasonable to rely. Insurers and banks can price by age. They just have to be able to show the evidence is relevant and the source reasonable.
Compare this with what happened to sex. In the Test-Achats case the Court of Justice of the European Union struck down Article 5(2) of the Gender Directive, the provision that had allowed insurers to differentiate by gender, ruling it incompatible with the principle of equal treatment and invalid from 21 December 2012. The wording of the age exception is close to identical to the gender exception the court destroyed. It survives anyway.
Age, in other words, occupies a legal position no other protected characteristic holds: formally protected, substantively negotiable, with a standing statutory permission to price on it as long as you can point at a table.
There is a second awkwardness. Age is continuous, not categorical. Most fairness metrics in machine learning are built for discrete groups: compare outcomes for group A against group B, measure the gap, minimise it. Continuous attributes require binning, and binning is a modelling choice that quietly determines what you will find. Is the relevant comparison 65 and over against under 65? 80 and over against everyone else? Each decade separately? A model can look admirably fair across coarse bands and be brutal at the eighty-fifth birthday.
That the technical community is working on this is not in doubt. A paper posted to arXiv on 7 April 2026 by Fernando López, Paula Delgado-Santos, Pablo Gómez, David Solans and Jordi Luque examined demographics-agnostic training for bias mitigation in wake-up word detection, evaluating fairness across sex, age and accent, and reported that one technique reduced predictive disparity by 83.65 per cent for age. Note what that result implies about the baseline. A voice interface, the very modality most often proposed as the accessible option for people who struggle with screens, had an age disparity large enough that removing four fifths of it counted as a headline.
Where the Actuarial Defence Runs Out #
Take the actuarial argument seriously and it still only covers a fraction of the territory.
It works for life insurance, where the outcome predicted is death and age is causally implicated in death. It works, with more strain, for annuities and some health cover. It does not work for the vast and expanding class of decisions where age enters not as a causal variable but as a learned correlation with something the model was never asked to think about.
Consider hiring. In August 2023 the United States Equal Employment Opportunity Commission settled with the tutoring company iTutorGroup for $365,000, in what was widely described as its first settlement involving an AI-driven hiring tool. The company's application software had been configured to automatically reject female applicants aged 55 and over and male applicants aged 60 and over. More than 200 qualified applicants were rejected on that basis. The discrimination came to light in an almost novelistic way: an applicant submitted two applications identical in every respect except the date of birth, and only the younger one got an interview. The consent decree included five years of EEOC monitoring and an injunction against requesting applicants' birth dates.
There is no actuarial defence for that. It was a hard filter, and it was illegal.
The more consequential case is messier. In Mobley v. Workday, Derek Mobley alleges that the applicant screening tools supplied by Workday systematically disadvantaged older job seekers; he says he submitted more than a hundred applications through the platform and was rejected every time. Judge Rita Lin of the United States District Court for the Northern District of California dismissed his intentional discrimination claim but allowed the disparate impact allegation to proceed, and on 16 May 2025 granted preliminary certification of a nationwide collective action under the Age Discrimination in Employment Act, covering applicants aged 40 and over denied recommendations through the platform since 24 September 2020. The court had earlier accepted the theory that an AI vendor could be directly liable for employment discrimination as an “agent” of the employer, and a March 2026 ruling rejected Workday's argument that the ADEA does not cover job applicants.
Two rulings since have sharpened it, in opposite directions. On 28 May 2026 the court held that AI bias-testing data can be protected from discovery by attorney-client privilege, shielding Workday's own testing material while accepting that it had probative value as evidence of disparate impact, on the basis that counsel had been substantively involved in curating it. The most useful evidence for establishing whether a screening model disadvantages older applicants is the vendor's own bias testing, and a vendor that routes that testing through its lawyers may be able to keep it from the people the model rejected.
Then on 22 June 2026 the court granted in part and denied in part Workday's motion to dismiss. It declined to dismiss an Americans with Disabilities Act claim built on a proxy-discrimination theory, the allegation being that the screening tools inferred health status. It declined to dismiss claims under California's Fair Employment and Housing Act, finding sufficient allegations that Workday designed, developed, maintained and controlled the tools from its California headquarters. It did dismiss a Title VII race-based disparate impact claim brought by a plaintiff who had not sought authorisation to add it, and struck a newly asserted theory that Workday was itself the direct employer. The allegations remain allegations; the case is unresolved.
What makes Mobley the important one is that nobody claims a birth date field was set to reject anyone. The claim is that a model, trained on which past applicants got hired, learned the shape of the people who tend to get hired, and that this shape has an age.
That is not actuarial risk. That is a machine reproducing a hiring market's existing prejudice at industrial throughput and calling it a recommendation.
The Ghost in the Feature Set #
The standard corporate response is to remove age from the model. Anyone who has worked on this knows why that fails, but the mechanism deserves spelling out, because it is where the eighty-five-year-old actually gets caught.
Age is one of the most redundantly encoded attributes in consumer data. It leaks through everything.
The landmark demonstration of how much can be inferred from almost nothing is the study by Tobias Berg, Valentin Burg, Ana Gombović and Manju Puri, published in the Review of Financial Studies, which analysed over 250,000 purchases at a German e-commerce firm. Their finding was that a handful of trivially available “digital footprint” variables matched the predictive power of a credit bureau score for consumer default. The variables were not financial. They included the device type and operating system the customer was using, characteristics of their email address, the channel through which they arrived at the site, and the time of day the order was placed.
Every one is age-correlated. Operating system and device age track purchasing power and upgrade behaviour, which skew by generation. Email domain is a near-fossil record: certain providers cluster heavily among people who set up an address in a particular decade and never changed it. Whether you arrived via a search engine, a price-comparison site or by typing the address directly is a behavioural signature that varies sharply with digital fluency. Time of day correlates with employment status and with sleep patterns that shift with age.
Add the signals a modern web session captures without asking: typing speed and correction rate, scroll behaviour, time per form field, zoom level, whether accessibility settings are enabled, session length, abandonment patterns, whether the customer switched to the telephone halfway through.
A model given these features and told to predict default, or churn, or fraud, or claim frequency, will find age whether or not you have deleted the birth date column. It will not label the pattern “age”. It will simply learn that a slow-typing user on an old Android device who zoomed the page, took eleven minutes over a form and then rang the call centre belongs to a cluster with a particular outcome rate. The cluster is old people. The model does not know this and does not need to.
This is proxy discrimination, and age is the characteristic most vulnerable to it, because unlike race or sex it is continuously written into behaviour rather than occasionally into a form field. You can decline to state your sex. You cannot decline to type at the speed you type.
It is also, since June 2026, a theory a federal court has agreed to hear. The proxy-discrimination claim that survived Workday's motion to dismiss is this argument made in a courtroom rather than a conference paper: that a system can sort people by a protected characteristic it was never given and could not name if asked.
A World Where They Were Not There #
There is a deeper problem underneath the proxy problem, and it is the one the World Health Organization identified with unusual bluntness in February 2022 in its policy brief “Ageism in artificial intelligence for health”.
The brief made a point that is easy to nod along to and hard to fully absorb: the datasets used to train AI models frequently exclude older people, who often sit within a minority subset for technologies not explicitly designed as gerontechnology. It set out eight considerations, among them participatory design of AI by and with older people, age-diverse data science teams, age-inclusive data collection, investment in digital infrastructure and digital literacy for older people and their carers, rights for older people to consent and to contest, and governance frameworks with teeth.
Under-representation in training data is not neutral. This is the part that gets lost.
If a population is thinly represented in the data, the model has less signal about them, so its predictions for them are less accurate and typically more conservative. Less accurate prediction for a group means more errors in both directions, but the consequences of those errors are asymmetric. A false negative for an older applicant means a declined loan or a rejected application. A false positive means an accepted risk. Institutions tune thresholds to avoid the second kind of error, so noisier estimates for a group systematically produce more refusals for that group. Uncertainty gets priced as risk. Now compound it across time. Because older people were less present in digital life during the decades when the training corpora were accumulating, they generated fewer digital records. Because they generated fewer records, models are worse at assessing them. Because models are worse at assessing them, they are more often refused or steered towards manual, slower, more expensive channels. Because they are pushed off the digital rails, they generate still fewer records. This is the thin-file problem, and for older people it runs in the opposite direction from the intuitive one: a person can have fifty years of impeccable financial history and still be functionally invisible to a model that mostly reads behavioural exhaust from the last eighteen months.
The past is not a neutral training set. It is a record of who was allowed to participate.
The Recursive Trap #
Here is where the digital divide stops being a story about access and becomes a story about power, and it is the thread that runs directly back to the woman with the robot on her side table.
The numbers are not ambiguous. Ofcom's Adults' Media Use and Attitudes report, published on 2 April 2026, found that 6 per cent of UK adults still have no home internet access, and that 83 per cent of that group are aged 65 or over, with 66 per cent aged 75 and above. Among those offline, 68 per cent said they were not interested or felt no need, 38 per cent found it too complicated and 25 per cent cited cost. Age UK reported in July 2025 that 2.4 million older people, nearly one in five, use the internet less than once a month or not at all, that 920,000 had reduced their internet use in the previous twelve months, and that 4.3 million, a third, do not use a smartphone. Exclusion was higher among older Black people at 32 per cent and older Asian people at 26 per cent, and among older people living alone at 30 per cent. Caroline Abrahams, Charity Director at Age UK, has warned repeatedly that people who cannot or will not go online must still be able to reach services offline.
In the United States, Pew Research Center reported in January 2026, drawing on a survey conducted between February and June 2025, that 78 per cent of adults aged 65 and over own a smartphone against 97 per cent of adults under 50, that 70 per cent have home broadband, and that 14 per cent are online almost constantly compared with 63 per cent of those aged 18 to 29.
Now put those two facts side by side.
Fact one: older people are the population most likely to be scored by systems they cannot see, and most likely to be misscored because of thin data and proxy leakage. Fact two: older people are the population least equipped to use the machinery that exists for challenging an automated decision.
Because that machinery is digital. All of it. The right of appeal lives behind a login. The “why was I declined?” explanation is a link in an email. The subject access request is a web form. The complaint goes to a chatbot that triages before a human sees it. The regulator's guidance is a PDF. The decision notice arrives in an app.
Article 22 of the General Data Protection Regulation is supposed to be the backstop. It gives people the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects, and where such decisions are permitted, Article 22(3) requires safeguards including the right to obtain human intervention, to express a point of view and to contest the decision. Legal scholars have criticised the provision for years, principally over the word “solely”, which invites institutions to insert a nominal human who rubber-stamps the model output, and over the vagueness of what counts as meaningful information about the logic involved.
But there is a failure mode more basic than any of those doctrinal complaints. A right you have to go online to exercise is not a right for someone who is not online. It is a courtesy extended to people who can already reach it.
This is the recursive trap. The people most likely to be wrongly assessed by an algorithm are, by the same underlying cause, the people least able to find out that an algorithm assessed them, least able to obtain an explanation, and least able to appeal. Digital exclusion does not merely sit alongside algorithmic harm. It is the mechanism that makes algorithmic harm unaccountable. The error and the inability to contest the error have a common origin, and that common origin is age.
What Seoul Cannot Fix by Declaration #
On 2 September 2026, the ASEM Global Ageing Center will convene the 6th ASEM Forum on the Human Rights of Older Persons in Seoul, under the theme “Artificial Intelligence and the Human Rights of Older Persons: Toward Age-inclusive AI Transformation”, bringing together experts from Asia and Europe. Suh's Korea Times piece was, transparently, a curtain-raiser for it.
The forum arrives at a moment of unusual institutional motion. On 3 April 2025 the UN Human Rights Council adopted by consensus a decision to establish an intergovernmental working group to begin drafting a legally binding convention on the human rights of older persons, after more than a decade of stalled effort in the open-ended working group on ageing. The new body held an organisational meeting in Geneva in February 2026, convened its first substantive session from 13 to 17 July 2026, and will hold its second session from 26 to 30 October 2026 at the Palais des Nations in Geneva, with the early sessions devoted to purpose, general principles and scope before any drafting of articles. The timetable beyond that is now known: a discussion of an outline of the convention's main elements is expected in July 2027, and a first zero draft around October 2027. The International Telecommunication Union published its own report in July 2026 on artificial intelligence and ageing, examining the unequal distribution of AI benefits across age, gender and region.
This is real progress and it will take years. An outline of main elements in mid-2027, a zero draft in late 2027, and then the negotiation of text that states have to agree, sign and ratify before it binds anyone. The woman in the opening paragraph is eighty-five now.
Meanwhile, the regulation that already exists treats age oddly. The EU AI Act prohibits, under Article 5(1)(b), AI systems that exploit vulnerabilities due to age, disability or specific socio-economic circumstances in order to materially distort behaviour, with those prohibitions applicable from 2 February 2025. That is a genuine protection against the payday-lender-targeting-the-cognitively-impaired scenario. It is not a protection against underwriting. Credit scoring and life and health insurance pricing sit in the high-risk category under Annex III, which brings documentation, data governance and human oversight duties, but lawful risk assessment carried out for legitimate purposes with proportionate safeguards falls outside the Article 5 prohibition. Which is to say: the Act catches predation and regulates underwriting, but it does not question whether age-based underwriting is itself the problem, because European law has already decided that it is not.
The American position is different and, in its way, weaker. The Age Discrimination in Employment Act of 1967 covers employment and does so reasonably robustly, as Mobley is testing. Outside employment, there is no federal equivalent of the Equality Act's general services provision. The Equal Credit Opportunity Act does prohibit age discrimination in credit transactions, and Regulation B is on its face protective: a creditor may use age as a predictive variable in an empirically derived, demonstrably and statistically sound scoring system, but only provided the age of an elderly applicant is not assigned a negative factor or value, and applicants aged 62 and over must be treated at least as favourably as those under 62. Read that carefully and you can see exactly where it fails. The rule polices the explicit age variable. It says nothing about a model that has never seen a date of birth and has inferred one from an operating system, a typing cadence and an email domain. You cannot audit for a negative factor assigned to elderly applicants if the model does not know which applicants are elderly, and neither, formally, does the institution running it.
Conference declarations do not touch any of this. The gap between what Seoul will resolve and what a German e-commerce model does with an old Android handset is the whole subject.
Four Things That Would Actually Change the Arithmetic #
Strip away the declaratory language and there are perhaps four interventions that would materially alter the position of the woman in the opening scene. None of them are technically hard. All of them are expensive, which is why they have not happened.
The first is a legally enforceable right to a non-digital channel. Not a helpline that exists at the discretion of the provider, not a branch that survives until the next cost review, but a statutory obligation on any organisation providing an essential service to maintain a route to a competent human being that requires no internet connection, no smartphone and no app. Age UK has been arguing this for years, most pointedly in its work on the collapse of high street banking, where the shift to online-only services has left people who do not trust or cannot use digital channels struggling to manage their own money. Offline access is currently a courtesy. It needs to be a licence condition.
The second is age-disaggregated auditing as a compliance requirement rather than a research exercise. Institutions deploying models in credit, insurance, employment and healthcare triage should be required to publish outcome rates by fine-grained age band, not by a single crude over-65 bucket that conceals everything interesting, and to do so alongside approval rates, appeal rates and appeal success rates. You cannot regulate a disparity that nobody has to measure.
The third is co-design that is not decorative. The WHO brief called for participatory design by and with older people and for age-diverse data science teams, which sounds like boilerplate until you consider how few people building consumer risk models have ever watched an eighty-five-year-old complete an online form.
The fourth is human review that is genuinely reachable and genuinely empowered: a named person, contactable by telephone, with the authority and the information to overturn a model, and a duty to record why. Article 22 gestures at this. Nobody has made it real.
The Grandchild That Isn't #
Return, finally, to the room with the robot in it, because there is a question underneath the discrimination question that the fairness metrics cannot reach.
Sherry Turkle, the MIT professor who has spent decades studying what happens to people around machines, argued in Alone Together that sociable technology “will always disappoint because it promises what it cannot deliver. It promises friendship but can only deliver performances.” Her worry was never that robots do things for us. It was that they do things to us: that we attach to what we nurture, and that outsourcing the nurturing dissolves the attachment.
Recent research suggests this is not merely philosophical. A study posted to arXiv on 26 February 2026 by Tianqi Song, Black Sun, Jingshu Li, Han Li, Chi-Lan Yang, Yijia Xu and Yi-Chieh Lee examined AI-generated influencers on Chinese short-video platforms that adopt kinship personas, presenting themselves as virtual grandchildren, using visual and conversational cues to enact family roles. Through social media analysis and interviews with older adults, the researchers found that these relationships met real informational and emotional needs, and also identified risks including emotional displacement and unequal emotional investment.
Unequal emotional investment. That phrase should be read slowly. It describes a relationship in which one party gives everything and the other party is a product roadmap.
The care-substitution risk is not that a family decides to buy a robot instead of visiting. It is subtler and more institutional. It is that a care system under fiscal pressure, looking at a 95 per cent loneliness reduction figure and a device that costs a fraction of a support worker's hourly rate, makes an entirely rational commissioning decision. Nobody need intend the substitution for it to happen. It is a budget line, not a betrayal.
The Bill Arrives in the Post #
So here she is, eighty-five years old, at the intersection of both machines.
The one on her side table knows her medication schedule, notices when she has not moved for a while, and asks about her day. It has probably improved her life; the engagement data suggests she uses it constantly. It was designed for her, sold to a state agency on her behalf, and screened to her through a public programme.
The other one does not know she exists as a person. It knows a vector: an operating system four versions behind, a form completion time in the ninety-ninth percentile, an email domain that has not been fashionable since the Clinton administration, a preference for the telephone channel, a session at three in the afternoon. It has never been told her age. It does not need to be told her age. It has learned the residue that age leaves on everything a person touches, and it has priced it.
When it declines her, or loads her premium, or drops her below a referral threshold, the letter will not say why. If she wants to know why, she will be directed to a portal. If she cannot use the portal, she will be offered a chatbot. If the chatbot cannot help, she will be given a number that leads to a menu. And at the end of that process, if she reaches it, a human being will look at a screen showing the model's output and a confidence score, and will decline to overturn it, because overturning it requires a reason and the reason is buried in a feature interaction that nobody at the institution can articulate either.
And if anyone ever tested the model for age bias, the test may be privileged.
Suh's article carried the line that no one should be considered too old for AI. It is a good line. But the more precise formulation of the problem is that nobody is too old for AI, because AI does not require your participation. It requires only your data exhaust, and it will make decisions about your money, your health and your access to public life whether or not you have ever touched a keyboard.
The robot on the side table is the part of this she agreed to. Everything else is happening to her, in rooms she will never see, in a language nobody will translate, on the basis of a life lived mostly before the data started being collected.
Sources and References #
- Suh Chung-ha (2026) “No one should be too old for AI,” The Korea Times, 19 August 2026. Available at:https://www.koreatimes.co.kr/opinion/20260819/no-one-should-be-too-old-for-ai - World Health Organization (2022) Ageism in artificial intelligence for health: WHO policy brief, WHO, 9 February 2022, ISBN 978-92-4-004079-3. Available at:https://www.who.int/publications/i/item/9789240040793 - New York State Office for the Aging (2023) “NYSOFA's Rollout of AI Companion Robot ElliQ Shows 95% Reduction in Loneliness,” press release, 1 August 2023. Available at: https://aging.ny.gov/news/nysofas-rollout-ai-companion-robot-elliq-shows-95-reduction-loneliness - New York State Office for the Aging (2026) Transforming Care for Older Adults Across the State of New York: ElliQ Project Update, February 2026. Available at:https://aging.ny.gov/system/files/documents/2026/02/nysofa-elliq-project-update-2026.pdf - Pu, Lihui, Moyle, Wendy, Jones, Cindy and Todorovic, Michael (2019) “The Effectiveness of Social Robots for Older Adults: A Systematic Review and Meta-Analysis of Randomized Controlled Studies,” The Gerontologist, 59(1), pp. e37-e51. Available at:https://academic.oup.com/gerontologist/article/59/1/e37/5036100 - Yen, Hsin-Yen, Huang, Chia-Wei, Chiu, Hsiao-Ling and Jin, Gaoxiang (2024) “The Effect of Social Robots on Depression and Loneliness for Older Residents in Long-Term Care Facilities: A Meta-Analysis of Randomized Controlled Trials,” Journal of the American Medical Directors Association, 25(6), 104979. Available at:https://www.jamda.com/article/S1525-8610(24)00176-2/fulltext - U.S. Equal Employment Opportunity Commission (2023) “iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit,” press release, 9 August 2023. Available at: https://www.eeoc.gov/newsroom/itutorgroup-pay-365000-settle-eeoc-discriminatory-hiring-suit - Holland & Knight (2025) “Federal Court Allows Collective Action Lawsuit Over Alleged AI Hiring Bias,” Insights, May 2025. Available at:https://www.hklaw.com/en/insights/publications/2025/05/federal-court-allows-collective-action-lawsuit-over-alleged - Duane Morris Class Action Defense (2026) “California Federal Court Grants In Part And Denies In Part Workday's Motion To Dismiss In Mobley v. Workday,” 24 June 2026. Available at: https://blogs.duanemorris.com/classactiondefense/2026/06/24/california-federal-court-grants-in-part-and-denies-in-part-workdays-motion-to-dismiss-in-mobley-v-workday/ - Norton Rose Fulbright (2026) “Behind the privilege shield: Safeguarding AI bias-testing data in employment decisions,” Inside Tech Law, June 2026. Available at:https://www.insidetechlaw.com/blog/2026/06/behind-the-privilege-shield-safeguarding-ai-bias-testing-data-in-employment-decisions - Berg, Tobias, Burg, Valentin, Gombović, Ana and Puri, Manju (2020) “On the Rise of FinTechs: Credit Scoring Using Digital Footprints,” The Review of Financial Studies, 33(7), pp. 2845-2897. Available at:https://academic.oup.com/rfs/article-abstract/33/7/2845/5568311 - Pinsent Masons (2024) “Age discrimination and insurance,” Out-Law Guide. Available at:https://www.pinsentmasons.com/out-law/guides/age-discrimination-and-insurance - UK Government (2012) The Equality Act 2010 (Amendment) Regulations 2012, SI 2012/2992, legislation.gov.uk. Available at:https://www.legislation.gov.uk/uksi/2012/2992/made - Information Commissioner's Office (2026) “Rights related to automated decision making including profiling,” UK GDPR Guidance and Resources. Available at:https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/individual-rights/rights-related-to-automated-decision-making-including-profiling/ - European Union (2024) “Article 5: Prohibited AI Practices,” Regulation (EU) 2024/1689 (Artificial Intelligence Act). Available at:https://artificialintelligenceact.eu/article/5/ - Ofcom (2026) Adults' Media Use and Attitudes Report 2026, published 2 April 2026. Available at:https://www.ofcom.org.uk/siteassets/resources/documents/research-and-data/media-literacy-research/adults/adults-media-use-and-attitudes-2026/adults-media-use-and-attitudes-2026-report.pdf - Age UK (2025) “Age UK warns 2.4 million digitally excluded older people are at risk of being left behind in an increasingly digital world,” press release, 29 July 2025. Available at: https://www.ageuk.org.uk/latest-press/articles/age-uk-warns-22.4-million-digitally-excluded-older-people-are-at-risk-of-being-left-behind-in-an-increasingly-digital-world/ - Age UK (2023) “You can't bank on it anymore”: The impact of the rise of online banking on older people, May 2023. Available at:https://www.ageuk.org.uk/siteassets/documents/reports-and-publications/reports-and-briefings/money-matters/the-impact-of-the-rise-of-online-banking-on-older-people-may-2023.pdf - Pew Research Center (2026) “Internet use, smartphone ownership, digital divides in the US: What we know,” Short Reads, 8 January 2026. Available at:https://www.pewresearch.org/short-reads/2026/01/08/internet-use-smartphone-ownership-digital-divides-in-u-s/ - Office of the High Commissioner for Human Rights (2026) “Second session of the Intergovernmental Working Group on the human rights of Older Persons, 26-30 October 2026,” OHCHR. Available at: https://www.ohchr.org/en/events/sessions/2026/second-session-intergovernmental-working-group-human-rights-older-persons-26 - International Telecommunication Union (2026) Artificial intelligence and ageing: Inclusive pathways for older persons in a digital world, ITU, July 2026. Available at:https://www.itu.int/hub/publication/d-phcb-ai_age-2026/ - Song, Tianqi, Sun, Black, Li, Jingshu, Li, Han, Yang, Chi-Lan, Xu, Yijia and Lee, Yi-Chieh (2026) “Understanding Older Adults' Experiences of Support, Concerns, and Risks from Kinship-Role AI-Generated Influencers,” arXiv preprint 2602.22993, 26 February 2026. Available at:https://arxiv.org/abs/2602.22993 - López, Fernando, Delgado-Santos, Paula, Gómez, Pablo, Solans, David and Luque, Jordi (2026) “'OK Aura, Be Fair With Me': Demographics-Agnostic Training for Bias Mitigation in Wake-up Word Detection,” arXiv preprint 2604.05830, 7 April 2026. Available at:https://arxiv.org/abs/2604.05830 - Turkle, Sherry (2011) Alone Together: Why We Expect More from Technology and Less from Each Other, Basic Books, New York. - Consumer Financial Protection Bureau (2026) “12 CFR § 1002.6 Rules concerning evaluation of applications,” Regulation B (Equal Credit Opportunity Act), eCFR. Available at:https://www.ecfr.gov/current/title-12/chapter-X/part-1002/subpart-A/section-1002.6
Tim Green UK-based Systems Theorist & Independent Technology Writer
Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.
His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.
**ORCID:** [0009-0002-0156-9795](https://orcid.org/0009-0002-0156-9795)
**Email:** [tim@smarterarticles.co.uk](mailto:tim@smarterarticles.co.uk)
Listen to the free weekly [SmarterArticles Podcast](https://www.smarterarticles.fm)