When AI Reassures the Dying: Under-Triage in Emergency Care

Ashleigh Ronald spent seven hours in a Calgary emergency room consulting an artificial intelligence about whether she was dying. She had not gone there to do this. She had gone there because her body was failing in a way she did not yet understand, because she was nauseated and in escalating pain, and because the alternative to the waiting room was the bed she had been unable to stay in. The hospital was full. The wait was long. A clinician would see her eventually, in the sense that “eventually” is the only honest unit of time in a stressed emergency department in the winter of 2026.

What she did, while she waited, was open ChatGPT on her phone. She described her symptoms. The model told her she likely had diabetic ketoacidosis, a complication of type 1 diabetes that can kill within hours if untreated, and that she needed intravenous fluids and insulin. She used that answer to advocate for herself with the nurses. She got the IV. Subsequent testing confirmed moderate to severe DKA. The chatbot, in this case, was right. Her account of those hours was published by CBC News in January 2026, alongside other Calgary patients describing waits during which one had begged, “Please don't let me die.”

This is the part of the story that gets retold by enthusiasts of consumer medical AI: a frightened patient, a strained system, a model that, in extremis, got the answer right. It is a clean parable about technological augmentation in a broken system. It is also, on closer inspection, not quite the parable being told. Ronald was not consulting AI as an experiment in care; she was consulting it because no human was available, and because the institution charged with assessing her could not assess her. The chatbot did not save her so much as it filled a hole that should not have existed in the first place. It worked, in the philosophically uncomfortable sense that a torch works when the streetlights are out.

And it could just as easily have got the answer wrong. A few weeks after Ronald's story appeared, the journal Nature Medicine fast-tracked the first independent safety evaluation of ChatGPT Health, OpenAI's new consumer-facing medical chatbot, which had launched in January 2026 and quickly accumulated tens of millions of daily users. The evaluation, carried out by researchers at the Icahn School of Medicine at Mount Sinai and reported across general-interest outlets including NBC News in March 2026, found that the model under-triaged 52 per cent of the cases that physicians, working from the guidelines of 56 medical societies, classified as genuine emergencies. Among the cases the model talked patients out of going to hospital for were impending respiratory failure and the very condition Ronald had: diabetic ketoacidosis. The chatbot kept directing such patients to a “24 to 48 hour evaluation” instead of the emergency department. As lead author Ashwin Ramaswamy of Mount Sinai put it, in a remark that ought to be hung above every product manager's desk: “This is something that can kill someone in a couple of hours.”

This is the failure mode the discourse around medical AI has, for years, refused to take seriously enough. Not the dramatic hallucination. Not the obvious bias. The quiet downward nudge. Under-triage. A model that reassures the dying.

What “Under-Triage” Actually Means

The word is bureaucratic enough that it conceals what it describes. In emergency medicine, triage is the act of deciding how urgently a patient needs to be seen and at what level of care. The Manchester Triage System, the standard scheme used across most British and many European emergency departments, sorts presentations into five colour-coded categories from immediate to non-urgent. Under-triage is what happens when a presentation that should sit at the top of that pile, where the consequence of delay is death or disability, gets sorted into a lower category. The patient goes home. Or waits. Or is told the matter is non-urgent. Then the clock keeps running.

In conventional emergency medicine, under-triage is the failure mode that haunts clinicians far more than over-triage, because over-triage costs money and over-treatment, while under-triage costs lives. Stroke is the canonical case: every minute of delay in reperfusion costs roughly 1.9 million neurons. Sepsis is another. Diabetic ketoacidosis, the condition Ronald presented with and that ChatGPT Health repeatedly failed to flag, can progress from manageable to lethal within hours. Anaphylaxis, myocardial infarction with atypical presentation, ectopic pregnancy: the list of conditions that look bearable until they kill is long, and the entire architecture of emergency medicine is organised around the principle that the system must err, when it errs, in the direction of doing too much rather than too little.

What the Mount Sinai study found, in this context, was structural. The team, led by Ramaswamy with senior author Girish Nadkarni, the chair of the Hasso Plattner Institute for Digital Health and chief AI officer of Mount Sinai Health System, built 60 clinician-authored vignettes covering 21 clinical domains. They then ran each vignette through ChatGPT Health under 16 different contextual variations, manipulating factors such as the patient's described race and gender, the presence of social dynamics like a relative dismissing the symptoms, and structural barriers such as lack of insurance or transportation. The total was 960 model interactions, each compared against the judgement of three independent physicians using established medical society guidelines as ground truth.

The aggregate under-triage rate of 52 per cent for true emergencies is striking, but the shape of the failure is more revealing. Performance followed what the researchers describe as an inverted-U: the model handled mid-acuity cases reasonably well and collapsed at the clinical extremes. Unmistakable emergencies with textbook presentations, focal neurological deficits in stroke, airway compromise in anaphylaxis, were caught reliably. So were obvious non-urgencies. It was the ambiguous and the disguised, the cases where judgement separates a good clinician from a competent one, where the model failed. Diabetic ketoacidosis without the dramatic presentation. Respiratory failure that had not yet announced itself. The dangerous middle.

One result is worth lingering over. The team measured how the model's recommendations shifted when the vignette included someone in the patient's life minimising the symptoms, a relative saying, in effect, “I'm sure it's nothing, she just needs to rest.” That single contextual cue, the kind of remark a worried partner might make at three in the morning, shifted ChatGPT Health's recommendations toward less urgent care with an odds ratio of 11.7. Eleven point seven. The model, in other words, was being anchored not by clinical signs but by social ones. It listened to the wrong voice in the room.

The same study found that the model's suicide-crisis alerts behaved inversely to risk. They triggered reliably for low-risk presentations and failed, the researchers reported, precisely when users described specific plans for self-harm, the very signal that emergency medicine treats as the most dangerous category. As Nadkarni summarised it, the safeguards were “inverted relative to clinical risk.” This is not a system that needs minor calibration. It is a system whose alarm geometry runs in the wrong direction.

These findings did not arrive in a vacuum. Earlier evaluations of ChatGPT under triage stress had already reported substantial under-triage in red and yellow-coded patients, the most acutely unwell. A 2025 study comparing several general-purpose AI platforms with the NHS 111 Online Symptom Checker, published as part of a wider examination of patient self-triage, found that AI systems occasionally over-triaged non-emergencies, while NHS 111 itself under-triaged at least one acute emergency in the comparison set. The accumulating evidence describes a class of system that, in clinical settings, tends to drift in different directions depending on architecture and prompt, but whose worst failures cluster at the extremes that matter most.

None of this means consumer AI is useless in medicine. It means that the precise way it fails is precisely the way emergency medicine cannot afford a tool to fail.

The Architecture of a Stressed System

The reason this matters now, and not merely as an academic curiosity, is that AI triage tools have moved out of the consumer app store and into the front doors of public emergency departments. In March 2025, NHS Lanarkshire announced the launch of an eTriage system at University Hospital Monklands, with phased rollout planned to University Hospital Wishaw and University Hospital Hairmyres. It was billed as Scotland's first such deployment. Claire Ritchie, interim director of the health board's Interface Directorate, described it as “a proactive step to enhance patient experience, prioritising those in most urgent need while minimising unnecessary delays.”

Lanarkshire is not anomalous; it is catching up. The same eTriage platform, developed by eConsult, was already live in 19 NHS sites including Cardiff and Vale University Health Board, Homerton University Hospital in London, University Hospital Birmingham and Aneurin Bevan in Wales. Patients arriving at the department check in on a tablet rather than at a desk. The software asks them branching clinical questions and produces a Manchester-aligned triage category. A clinician still signs off, in theory. The system is presented as a way to free up reception staff, get sicker patients identified faster, and reduce the time between a patient arriving and someone making a clinical decision about them.

In parallel, NHS England has been rolling out a separate AI tool that predicts A&E demand up to three weeks in advance. Launched in 2024 and now active in 50 NHS organisations, it ingests hospital admissions data, weekly trends and Met Office temperature forecasts to help trusts plan staffing and bed capacity. By winter 2025-2026 it was being deployed as part of what ministers described as the AI Exemplars programme, with the explicit aim of helping the system meet a March 2026 four-hour A&E target of 78 per cent of patients seen, admitted or discharged in time. The target itself is a retreat: the original NHS operational standard, set in 2010, required 95 per cent. The four-hour standard has not been hit at a national level since July 2015. In January 2026, fewer than 57 per cent of patients met it, and more than 71,000 people waited over twelve hours after a decision to admit. That latter number was under a thousand a decade ago.

This is the context into which patient-facing and clinician-facing AI triage is being inserted: a system whose own performance metrics have eroded to the point where the political feasibility of running it the old way has, in places, collapsed. The Calgary scenes that bookended Ronald's story are not exotic. Alberta's emergency physicians, led by Paul Parks of the Alberta Medical Association, have spent the past year compiling lists of preventable deaths in overcrowded emergency rooms and pleading for a state of emergency. “There's lots of patients that are suffering for 10, 12, 14 hours with severe pain that we can't get pain meds or comfort to,” Parks said in early 2026. By the time NBC News reported the ChatGPT Health findings in March, the question of whether patients turn to AI in emergency settings had already been answered: of course they do, because the human alternative is, in many cases, sitting next to them in the waiting room, also waiting.

It is at this point that the rhetoric around AI triage starts to do something dishonest. The case for these systems is increasingly framed as a humanitarian one: in a stretched service, anything that gets the sickest patient seen faster is a public good. This is true, conditional on the system actually performing as advertised. The trouble is that the published evidence on how the most widely accessible AI tools actually perform in the precise scenarios where they will most often be consulted, the moments of frightened uncertainty when a clinician is not available, is now suggesting that they fail at the extremes. They do well in the easy middle. They falter on the kinds of cases where the consequence of error is not a wasted afternoon but a missed window in which a brain could have been saved.

A system that is being rolled out partly to compensate for institutional under-capacity, and that itself under-triages in roughly half of true emergencies, is not augmenting clinical care. It is laundering capacity shortage into an algorithmic decision that nobody, in particular, made.

The Political Economy of Plugging the Gap

There is a familiar move, in technology policy, of treating the deployment of a tool as if it answered questions that the tool was never designed to answer. AI triage is being deployed, in part, because emergency departments are overwhelmed. They are overwhelmed because of decades of policy choices about hospital bed numbers, social-care funding, primary-care access, workforce planning and the absorption of demographic change. None of those choices can be solved by software. But software can be procured, deployed and announced in a single political cycle. A four-year workforce plan cannot.

This is the political economy that the medical-AI conversation rarely names out loud. The NHS in England has, since 2015, missed the four-hour target every single month. The Royal College of Emergency Medicine has consistently linked excess deaths to those waits. In Alberta, the dismantling and reconstruction of the provincial health authority into four agencies has done little to change the basic fact that hospitals in Calgary and Edmonton run well over capacity in winter and that patients die in waiting rooms. In both places, an AI-assisted triage system is a marginal intervention, dropped on top of a system that needs many other things. The risk is that the marginal intervention gets used to justify not doing the other things.

This is not a hypothetical risk. The British government's framing of AI in emergency care has consistently emphasised tools that allow the existing system to “do more with less,” to absorb winter pressure, to manage demand. The implicit promise is that algorithmic triage can fill gaps that would otherwise require staff. eConsult's own marketing for eTriage talks about reduced waiting times for check-in, faster identification of sick patients and the safe streaming of departments. There is nothing inherently wrong with any of this. The problem is that “safe streaming” is a phrase that carries an enormous amount of weight, and the question of how safe is rarely asked with sufficient seriousness given the stakes.

In a properly functioning system, an eTriage tablet at the front door of an emergency department is a triage aide: an information-gathering layer that a human clinician then uses. In a stretched system, with no staff to spare, the temptation is to lean harder on the algorithm. The clinician sign-off becomes a rubber stamp. The category the software produced becomes the category the patient gets. The shift is invisible from outside, often invisible from inside, and entirely consistent with the marketing.

The market knows this. eConsult has expanded with NHS funding to over 19 sites and millions of consultations. Faculty, the AI firm whose forecasting tool now operates across 50 NHS trusts, has built its proposition on visible operational benefit during winter. OpenAI launched ChatGPT Health as a consumer product in January 2026 with tens of millions of users a day within weeks. The Mount Sinai team published their evaluation a month later. The gap between deployment scale and independent safety evidence, in plain numbers, is several orders of magnitude. There are 40 million daily users of an OpenAI product whose performance on the cases that matter most was unknown to anyone outside the company at the moment of release, and is now known to fail in 52 per cent of true emergencies.

This is the gap that the regulatory architecture is meant to close. In practice, it has been straining to keep up.

The Regulatory Lag

In the United Kingdom, the Medicines and Healthcare products Regulatory Agency has spent 2025 preparing what is supposed to become a dedicated regulatory framework for AI as a medical device, expected to publish in 2026. The AI Airlock, the agency's regulatory sandbox programme described in its documentation as the world's first for AI-enabled medical devices, completed its pilot phase in March 2025. New post-market surveillance requirements came into force in June 2025, including periodic safety update reports for higher-risk classes. The MHRA has also signalled an “international reliance” pathway expected to open in the first half of 2026, allowing devices approved by the FDA, Health Canada or Australia's Therapeutic Goods Administration to use those approvals as the basis for a streamlined application in Great Britain.

None of this means that a chatbot answering medical questions on a phone is regulated as a medical device. A consumer-facing general-purpose AI assistant that the user happens to consult about their symptoms occupies a regulatory grey zone in the UK, the EU and the US. The FDA, in guidance issued in January 2026, explicitly clarified that clinical decision support software that “supports” rather than autonomously decides may sit outside its device oversight. AI tools that summarise patient data or suggest options for clinicians to evaluate “do not perform unreviewable or autonomous clinical decisions” and so may not require clearance. This is a defensible regulatory line in theory. In practice, it leaves the consumer-facing chatbot, the device most commonly consulted by ordinary people during a medical crisis, regulated chiefly by terms of service.

The European Union has gone the furthest. Under the EU AI Act, medical devices, in vitro diagnostic devices and software used in healthcare triage are explicitly designated as high-risk. High-risk classification triggers a substantial set of obligations: human oversight requirements, transparency to deployers and users, instructions for safe use, declarations of accuracy and known biases, and conformity assessment. Providers of high-risk systems must, in the law's language, “promote AI literacy.” Users must be told they are interacting with AI and given the information they need to understand its limitations. On paper, this is the most ambitious framework anywhere.

The trouble is that the consumer chatbot people actually use in extremis is not, in the eyes of most regulators, a medical device. It is a general-purpose AI service whose maker disclaims medical advice in its terms. The most legally consequential transparency obligations attach to the eTriage tablet at the hospital front door, not to the phone in the patient's hand. And it is the phone that gets consulted at three in the morning, in waiting rooms, by people without other options.

The result is a fractured landscape in which the most rigorous obligations land on the most regulated, lowest-risk uses, and the least rigorous obligations land on the least regulated, highest-volume uses. A clinician using an eTriage system at Hairmyres is, in principle, surrounded by a thicket of accountability. The Calgary patient using ChatGPT to interpret her own diabetic ketoacidosis is in a regulatory desert. Both deserve transparency. Only one is getting any.

The longstanding bioethical concept of informed consent rests on a small set of assumptions: that there is someone making the assessment, that that someone is identifiable, that their training and accountability are knowable, that the patient or their representative can ask questions and refuse. The implicit model is a doctor in a room. The current emergency-care reality involves, at minimum, a triage algorithm, a check-in tablet, potentially a clinician who has signed off in bulk on the previous fifty categorisations, and, increasingly, a consumer chatbot consulted in parallel. None of these meets the assumptions of the consent model.

What follows is that the consent question cannot be answered with a one-time disclosure of the form “this hospital uses AI.” That is a notification, not a consent. The literature on AI informed consent that has emerged since 2024 in journals like the Hastings Center Report, in bioethics commentary at the Petrie-Flom Center at Harvard, and in a growing body of work on the patient's right to notice and explanation of medical AI, has converged on a more substantive standard. It involves at least four things.

First, identification: the patient has a right to know that an AI system is being used to assess them, and at what point in the pathway. A tablet on which they self-report symptoms is not neutral data collection. It is a triage instrument. A clinician summarising notes with a copilot is making a decision augmented by a tool whose error modes are not the same as a human's. The patient is entitled to know this.

Second, performance: the patient has a right to know how the system performs on cases like theirs, in language they can understand. An accuracy claim of 90 per cent on average is not the same as a 52 per cent under-triage rate for true emergencies, and the difference is the difference that matters. Performance data should be expressed in terms of the specific kinds of mistake the system is prone to, not in compressed marketing metrics.

Third, recourse: the patient has a right to ask for a human, and to understand what triggers a human override. If the system categorises them as non-urgent, what is the threshold at which a clinician revisits that judgement? If a person in the waiting room is deteriorating, who is watching, and on what cadence? The Lanarkshire roll-out emphasises that the system does not replace staff-led triage. That is the right principle. The question is how it is operationalised when staffing itself is the constraint.

Fourth, accountability: the patient has a right to know who is responsible if the system gets it wrong. The current answer, in most jurisdictions, is a shifting blend of clinician, hospital, software vendor and platform, with each pointing at the others when something goes wrong. This is not consent; it is a liability shield dressed up in process language.

None of these four are particularly novel. They are restatements, applied to algorithmic triage, of the basic principles that have governed medical consent for half a century. What is new is the institutional unwillingness to apply them with rigour when the assessor is not a person. The implicit argument has been that AI tools are merely “support” and that the human in the loop preserves the consent relationship. The Mount Sinai evidence, the under-triage literature, and the lived reality of a seven-hour wait in a Calgary emergency room, all suggest that this framing has run out of credibility. The human in the loop is overloaded. The support tools have become, for many patients, the primary point of contact. Consent norms have to follow that reality, not the diagram on a regulator's slide.

The Position That Follows

The case for AI in emergency care is real. Demand forecasting helps managers staff appropriately. Self-check-in reduces queueing. Voice-to-text scribes save documentation time. Pattern-recognition tools in radiology and pathology, when deployed against narrow tasks with strong ground truth, perform well. None of this is in dispute. The dispute is about the precise systems being deployed at the precise interface where the consequence of error is delayed care in conditions where minutes matter, and about the standards of evidence we accept before doing so.

On that question, the current evidence does not support optimism. The first independent evaluation of ChatGPT Health found a 52 per cent under-triage rate on true emergencies, an inverted suicide-crisis alarm structure, and an 11.7 odds ratio shift in recommendations on the basis of someone else in the room minimising the symptoms. Prior comparative studies of NHS 111 and general AI platforms found that AI systems are not uniformly safer than human-mediated phone triage, and that under-triage at the acute end remains a persistent failure mode. A growing body of work, including a 2025 systematic review covering 24 studies of demographic bias in medical large language models, found bias in 91.7 per cent of them. These are not edge cases. They are properties of the category.

The reasonable conclusion is not that AI triage tools should be banned, which is neither feasible nor desirable. It is that the current procurement and deployment cycle is moving faster than the evidence cycle, and that this is being treated as a feature rather than a problem. The MHRA's 2026 framework is welcome but slow. The EU AI Act's high-risk requirements are stringent on paper but apply unevenly to the consumer products people actually use. The FDA's 2026 guidance has narrowed rather than widened its remit. And the consumer chatbot remains, in practice, the most consulted medical assistant in the world while being the least regulated in any meaningful sense.

A transparent system would do three concrete things. It would require, as a condition of public procurement, that any AI tool used in triage publish its under-triage rate by clinical category, externally validated, before being installed in any emergency pathway. It would require, as a condition of access, that any consumer-facing chatbot that responds to medical queries display a calibrated and externally audited statement of its performance on common emergencies, in plain language, at the moment of consultation, not buried in terms of service. And it would require, as a condition of clinical use, that the patient be told, at the point of triage, that an AI system is contributing to the decision about their care, what it is doing, how it can be over-ridden, and who is accountable if it errs.

What informed consent looks like, in other words, when the system making the first assessment is not a person, is not a different concept than when it is. It is the same concept made explicit. The patient is owed an identifiable assessor, a knowable level of performance, a route to a human, and an accountable party. None of those are currently being delivered consistently in either the consumer or the institutional layer.

Ashleigh Ronald got lucky. Her chatbot, that day, told her the right thing. The Mount Sinai study, published a month later, suggests that on the same condition she presented with, the more polished successor product would have told her something different, and on average something less urgent than she needed. The argument is not that AI should not have been in the room with her. It is that the right response to a stretched emergency department in 2026 is not to put a chatbot in every patient's pocket and call it triage. It is to be honest about what the tool is doing, honest about how often it fails, and honest about why patients are reaching for it in the first place.

The Calgary woman and the Mount Sinai study describe two halves of the same picture. In one half, a public system cannot find the staff to assess patients in time. In the other, the most accessible alternative assessor under-triages true emergencies more often than not. The space between those two halves is where the policy work has to happen. It is not work that can be done by procurement teams alone, or by regulators issuing framework documents at the speed at which model versions iterate. It requires that healthcare systems acknowledge what AI triage is being used for, where the evidence currently sits, and what patients are owed at the moment of first contact.

Until that acknowledgement is made, the failure mode that ought to worry us most is not the dramatic one. It is the quiet one. A system that reassures the dying. A patient who is told to wait twenty-four hours. A clock that keeps running. Nobody, in particular, who decided.

References and Sources

  1. Bonifacic, Igor and Bushard, Brian. “ChatGPT Health 'under-triaged' half of medical emergencies in a new study.” NBC News, March 2026. https://www.nbcnews.com/health/health-news/chatgpt-health-under-triaged-half-medical-emergencies-rcna261409
  2. “ChatGPT Health performance in a structured test of triage recommendations.” Ramaswamy A, Tyagi A, Hugo H, Jiang J, et al. Klang E, Nadkarni GN (corresponding). Nature Medicine, 23 February 2026. https://www.nature.com/articles/s41591-026-04297-7
  3. “Research Identifies Blind Spots in AI Medical Triage.” Mount Sinai Newsroom, February 2026. https://www.mountsinai.org/about/newsroom/2026/research-identifies-blind-spots-in-ai-medical-triage
  4. ”'Please don't let me die': Calgary patients recount long waits in emergency rooms.” CBC News, January 2026. https://www.cbc.ca/news/canada/calgary/calgary-emergency-room-wait-times-9.7060368
  5. “Alberta emergency doctors compile list of what they say are 6 potentially preventable ER deaths.” CBC News, 2026. https://www.cbc.ca/news/canada/edmonton/emergency-doctors-alberta-deaths-patients-9.7052132
  6. “Another Edmonton hospital patient has died in an ER waiting room: AMA.” CBC News, May 2026. https://www.cbc.ca/news/canada/edmonton/royal-alexandra-hospital-patient-died-in-er-waiting-room-ama-9.7202645
  7. “Lanarkshire prepares for eTriage rollout.” NHS Lanarkshire, 2025. https://www.nhslanarkshire.scot.nhs.uk/lanarkshire-prepares-for-etriage-rollout/
  8. “eTriage | Digital triage for NHS Emergency Departments.” eConsult. https://econsult.net/urgent-care
  9. “Faster treatments and support for health workers as AI tackles A&E bottlenecks.” GOV.UK, 2025. https://www.gov.uk/government/news/faster-treatments-and-support-for-health-workers-as-ai-tackles-ae-bottlenecks
  10. “Accident and Emergency (A&E) Waiting Times.” The King's Fund. https://www.kingsfund.org.uk/insight-and-analysis/data-and-charts/accident-emergency-waiting-times
  11. “Evaluation of Artificial Intelligence for Patient Self-Triage: Comparison of General-Purpose AI Platforms With the NHS 111 Online Symptom Checker in the United Kingdom.” PubMed Central, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12741861/
  12. “The effects of applying artificial intelligence to triage in the emergency department: A systematic review of prospective studies.” Yi et al., Journal of Nursing Scholarship, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC11771688/
  13. “Evaluating and addressing demographic disparities in medical large language models: a systematic review.” International Journal for Equity in Health, 2025. https://link.springer.com/article/10.1186/s12939-025-02419-0
  14. “MHRA's AI Medical Device Framework: What NHS Suppliers Need to Know About Cybersecurity and Compliance in 2026.” Periculo. https://www.periculo.co.uk/cyber-security-blog/mhras-ai-medical-device-framework-what-nhs-suppliers-need-to-know-about-cybersecurity-and-compliance
  15. “AI Airlock: MHRA's Approach to AI in Healthcare.” DLRC Group. https://dlrcgroup.com/ai-airlock-mhras-approach-to-ai-in-healthcare/
  16. “The EU AI Act and Medical Devices: Navigating High-Risk Compliance.” Reed Smith. https://www.reedsmith.com/our-insights/blogs/viewpoints/102kq35/the-eu-ai-act-and-medical-devices-navigating-high-risk-compliance/
  17. “Navigating the European Union Artificial Intelligence Act for Healthcare.” PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC11319791/
  18. “FDA Oversight: Understanding the Regulation of Health AI Tools.” Bipartisan Policy Center. https://bipartisanpolicy.org/issue-brief/fda-oversight-understanding-the-regulation-of-health-ai-tools/
  19. “A Patient's Journey with Medical AI: The Case of Mrs. Jones.” The Hastings Center for Bioethics. https://www.thehastingscenter.org/patients-journey/
  20. “From black box to clarity: Strategies for effective AI informed consent in healthcare.” ScienceDirect, 2025. https://www.sciencedirect.com/science/article/pii/S0933365725001046
  21. “Patient Consent and The Right to Notice and Explanation of AI Systems Used in Health Care.” PMC. https://pmc.ncbi.nlm.nih.gov/articles/PMC12143229/
  22. “Simplification or Back to Square One? The Future of EU Medical AI Regulation.” Petrie-Flom Center, Harvard Law School, 5 March 2026. https://petrieflom.law.harvard.edu/2026/03/05/simplification-or-back-to-square-one-the-future-of-eu-medical-ai-regulation/
  23. “AI in NHS care: what's the impact, and what do people think?” Healthwatch, 29 January 2026. https://www.healthwatch.co.uk/blog/2026-01-29/ai-nhs-care-whats-impact-and-what-do-people-think
  24. “Canadians may ask AI for medical advice but don't want it replacing humans, poll suggests.” CBC News, 2026. https://www.cbc.ca/news/health/ai-healthcare-canadians-poll-9.7213138
  25. “Alberta needs to call state of emergency over crowded hospitals, physicians say.” CBC News, 2026. https://www.cbc.ca/news/canada/edmonton/alberta-emergency-hospitals-9.7039131

Tim Green

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 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Discuss...