The room, in the scenario its designer envisages, is small and clean. It contains a chair, a screen, a microphone, and nothing else. The person who has come to die is asked to sit. The screen flickers on. A face appears, rendered in synthetic colour, with a voice that has been trained for warmth. The face asks why the person is there. It asks about pain. It asks about the alternatives the person has considered. It asks about the family. It asks again, in a slightly different way, about the alternatives. The conversation continues for perhaps an hour. At the end the avatar pauses, and a value somewhere in its underlying network resolves into one of two outcomes: capacity granted, or capacity denied. If granted, the door to the next room unlocks. In that room sits a 3D-printed nitrogen capsule. Twenty-four hours later, if the person still wishes to proceed, the capsule will let them die.
That room does not yet exist. It is the proposal the Australian euthanasia advocate Philip Nitschke set out in January 2026, when he announced that he was developing artificial intelligence software to replace psychiatrists in assessing whether a person seeking assisted dying possesses the mental capacity to make the decision. Nitschke is sixty-eight, no longer a registered doctor (his medical licence was permanently suspended in 2015 by the Medical Board of Australia), and the founder of Exit International. He is also the inventor of the Sarco pod, the device used for the first time in Switzerland on 23 September 2024, when a sixty-four-year-old American woman with a severely compromised immune system died inside it in a forest in the canton of Schaffhausen. Swiss authorities arrested Florian Willet, chief executive of the affiliated organisation The Last Resort, on suspicion of inducing and aiding suicide. The serious charge of intentional homicide was withdrawn within weeks. Willet himself died by suicide in Germany in May 2025. The pod has not been used again.
Nitschke's case for the AI assessor is presented as a complaint against human inconsistency. “I've seen plenty of cases where the same patient, seeing three different psychiatrists, gets four different answers,” he told reporters in January 2026. “There is a real question about what this assessment of this nebulous quality actually is.” His proposed alternative is a conversational avatar that interviews the candidate, draws inferences about their reasoning, and arrives at a binary outcome. If the AI grants capacity, the Sarco unlocks after a twenty-four-hour cooling-off period. If it denies, the candidate has no further recourse within the system.
Two months later, on 26 March 2026, Noelia Castillo Ramos died by legal euthanasia at a healthcare centre in Sant Pere de Ribes, in the Province of Barcelona. She was twenty-five. She had survived a suicide attempt in October 2022 that left her paraplegic, and she had been diagnosed with obsessive-compulsive disorder and borderline personality disorder. Her euthanasia request had been approved on 18 July 2024 by the Catalonia Guarantee and Evaluation Commission. It had then been delayed for 601 days by her father's appeals, which travelled through a Barcelona court, the High Court of Justice of Catalonia, the Spanish Supreme Court, the Constitutional Court and finally the European Court of Human Rights. Every one of those bodies, at every level, found that she had the capacity to decide. Uniladtech, reporting on the case in March 2026, noted that Castillo's twenty-month legal battle had revived a debate that until recently had been hypothetical: whether, in a system where capacity is the gate through which the entire procedure passes, the gate-keeper might one day be a machine.
In the jurisdictions that permit assisted dying (Switzerland, the Netherlands, Belgium, Luxembourg, Spain, Canada, Colombia, New Zealand, parts of Australia, ten US states plus the District of Columbia), the law requires that the person making the request have decision-making capacity. The form of the requirement varies. In Spain it is set out in Organic Law 3/2021 and assessed by the responsible physician and a consulting physician, with a Guarantee and Evaluation Commission as procedural backstop. In the Netherlands and Belgium, two decades of practice have produced a clinical literature in which capacity is most often presumed and only formally tested when doubt arises. In Canada, the Medical Assistance in Dying regime requires a capacity assessment by two practitioners. The United Kingdom's most recent attempt at a statute, Kim Leadbeater's Terminally Ill Adults (End of Life) Bill, would have written capacity testing on at least five separate occasions into the procedure, including a panel review by a psychiatrist, a social worker and a senior judge. That bill ran out of parliamentary time in 2025 and did not become law.
What unites these regimes is that the moment of capacity assessment is the load-bearing column of the entire structure. Everything else, the prognosis, the suffering, the documentation, the medical opinion, the cooling-off period, depends on the prior finding that the person before the clinician understands what they are choosing and can hold the choice steady. To propose that this assessment be performed by a machine is to propose that the column itself be replaced. The question is not whether such a substitution is technically possible. The question is what standard of evidence, accountability and explainability it would have to meet, who would set that standard, and who would be liable when the system was wrong.
What Capacity Actually Is
The clinical standard for decision-making capacity is older than most AI systems by several decades. The MacArthur Competence Assessment Tool for Treatment (MacCAT-T), developed by Thomas Grisso and Paul Appelbaum at the University of Massachusetts Medical School and published in 1997, identifies four abilities a person must demonstrate: the ability to communicate a choice; the ability to understand the relevant information; the ability to appreciate the situation and its likely consequences; and the ability to reason with the information in a way that is internally coherent. The MacCAT-T is administered as a semi-structured interview, takes fifteen to twenty minutes, and is calibrated against the patient's own clinical situation rather than a generic script. Its inter-rater reliability is high. It is the closest thing the field has to a gold standard, and it is what most of the formal clinical literature on capacity assessment for assisted dying assumes.
What the MacCAT-T cannot do, and what no successor instrument has succeeded in doing, is remove the human judgement at its centre. The clinician administering the interview has to decide whether the patient's articulation of their understanding is genuinely their own; whether their appreciation of consequences extends to the morbidity of their own affect; whether their reasoning is shaped by a depression that is itself a treatable condition. The Dutch literature on assisted dying for psychiatric suffering is unsparing on this point. A 2016 study in JAMA Psychiatry by Scott Kim and colleagues at the United States National Institutes of Health, reviewing sixty-six cases of euthanasia for psychiatric reasons in the Netherlands, found that in only a minority were the capacity assessments documented in any structured form. Survey research published among Dutch psychiatrists found that sixty-five per cent believed they could determine capacity in a patient with a psychiatric disorder requesting assisted dying; twelve per cent thought they could not; twenty-three per cent had doubts.
Nitschke takes this variability as evidence that the existing assessment is incoherent and that an AI could do better by being consistent. The inference is half right. The variability is real. The conclusion that consistency is the same as correctness, however, is the mistake at the centre of his proposal. A model that returns the same answer every time can be reliably wrong. The variability between psychiatrists is, in part, a feature of a genuinely contested judgement being made under uncertainty. To collapse that variability into a deterministic algorithm is to mistake the noise of human judgement for the signal of the underlying problem. Codifying the disagreement away does not resolve it. It only conceals it inside a model.
There is then the related problem of what the AI would actually be measuring. A capacity assessment is not a quiz. It is a relational interaction in which the clinician reads the patient's affect, hesitations, repetitions and changes of mind across time. The Dutch psychiatrists writing in Frontiers in Psychiatry in 2022 describe capacity in psychiatric euthanasia cases as a temporally extended judgement: not a snapshot but a moving picture, sometimes assembled over months. An avatar that speaks to a candidate for an hour cannot perform that kind of assessment, regardless of how richly trained its conversational model. Even a system fine-tuned on transcripts of human capacity assessments would inherit the structural limits of its training distribution: it would replicate the documented patterns of those assessments rather than independently verify the underlying capacity. If a substantial portion of the training data records cases in which capacity was presumed without rigorous test, the model will learn to presume.
The Bias That Lives in the Data
Nitschke's claim that AI is “less subject to personal bias” than a human clinician is the part of the proposal that has aged worst in the seven years since the most authoritative work on AI bias in medicine was published. The position is not novel. It is the same claim that has been made for AI in criminal sentencing, hiring, child welfare and visa adjudication, and in each domain the claim has not survived contact with the data. Models do not invent their judgements from first principles. They infer them from training distributions that reflect the prejudices of the institutions whose records they were trained on. The 2018 Gender Shades study by Joy Buolamwini and Timnit Gebru documented commercial facial classification systems with error rates of up to 34.7 per cent on darker-skinned women, against 0.8 per cent on lighter-skinned men, an asymmetry that arose not from any flaw in the architectures but from the demographic skew of the data on which they had been trained.
The clinical AI literature has reproduced the pattern in fine detail. A 2025 systematic review in Oxford Open Digital Health found that of 390 clinical AI studies examined, eighty-four per cent failed to report the racial composition of their training data and thirty-one per cent failed to report sex. A 2025 study in npj Digital Medicine on racial bias in AI psychiatric diagnosis found that large language models propose differential treatments when patient race is implicitly indicated, and that descriptive language describing Black male patients diverges in ways that align with documented patterns of involuntary hospitalisation. None of these findings is exotic. They are now baseline expectations of the field.
If the AI that Nitschke proposes were trained on the records of past capacity assessments, it would inherit any structural patterns those assessments contained. Spanish psychiatric data, Dutch end-of-life records, Belgian dossiers: each carries the demographic, linguistic and cultural particularities of the system that produced it. A model trained on European data and asked to assess capacity in a candidate whose first language is not the language of the training corpus, whose cultural framing of illness or family or suffering differs from the modal record, will not be neutral. It will be biased in ways that the model itself cannot articulate. The relational competence that a human psychiatrist brings to a difficult bilingual capacity assessment, the ability to ask the question in a different register, to wait for the second answer, to read silence as a signal rather than a missing data point, is precisely the competence that the model has not been trained to perform.
The Paper on Trust
On 29 April 2026, three authors from the Ukrainian computer-science community, Serhii Zabolotnii, Viktoriia Holinko and Olha Antonenko, posted to arXiv a paper that addresses the structural question Nitschke's proposal raises without ever naming his project. The paper, “From Black-Box Confidence to Measurable Trust in Clinical AI: A Framework for Evidence, Supervision, and Staged Autonomy”, argues that clinical AI trustworthiness cannot be inferred from accuracy benchmarks, fluency of generation, or the subjective confidence of human users. Trust, the authors write, must be engineered as a measurable property of the system, with explicit evidence trails, supervised escalation, and graduated action rights that depend on demonstrated calibration.
The paper's substantive proposal is the framework named in its title. A trustworthy clinical AI system, on this account, is built from a deterministic clinical logic core (the parts of the decision rule that can be written as code and audited line by line), a patient-specific assistant that validates the deterministic decision against the patient's individual context, a multi-tier escalation mechanism that routes uncertain cases upwards through a hierarchy of models and humans, and a human supervision layer that retains the right of final adjudication. Around these structural elements the paper specifies a set of trust metrics drawn from metrology: measurement uncertainty, calibration error, evidence trail completeness, autonomy boundary compliance, operational stability. The point is that an AI is not granted autonomy by fiat. It is granted autonomy by demonstrating, on instruments that can be inspected, that it deserves it.
The phrase the paper deploys for the governing principle is “staged autonomy”. A system begins life under tight human supervision, with its decisions advisory only. It progresses, if and only if its performance on the trust metrics warrants the progression, through stages of expanded autonomy. At each stage the evidence threshold is higher. The right to act without immediate human review is earned, not assumed. The framework is not specific to assisted dying, and the authors are careful not to claim domain-particular expertise. The framework is, however, exactly the framework against which a proposal like Nitschke's most usefully fails. A capacity-assessment AI deployed at the highest tier of autonomy, granting or denying access to an irreversible procedure on its own authority, would, on the paper's logic, need to clear an evidence threshold that no clinical AI to date has cleared, in a domain where the metrics themselves are contested.
The arXiv paper is a serious attempt to specify what would actually be required, in measurable terms, before a clinical AI is granted autonomous decision-making authority. It is also an implicit indictment of the practice that has tended to prevail in the field, in which AI tools are deployed with the language of “decision support” and then drift into operational use as decision-makers, on the back of confidence scores that have not been calibrated against any externally validated baseline. The drift is documented in domain after domain. There is no reason to think it would not occur in capacity assessment. There is every reason to think it would, because the surrounding economic and political pressures all point the same way: faster, cheaper, less litigable, more deniable.
The Two Errors
The categorical human stakes are easily stated. An AI capacity-assessor that wrongly grants approval to a person who lacks genuine capacity authorises a death that cannot be reversed. The reversal cannot be partial. There is no appeal procedure that returns the dead to their families. An AI that wrongly denies approval to a person who does have capacity denies them a legal right at the moment of maximum suffering. The denial is also categorical in its way: a person whose end-of-life decision has been refused does not, in any general sense, get to try again under different circumstances. They live the time they live, and they suffer what they suffer, with whatever options were available before the algorithmic refusal. Both errors are irreversible. The first is irreversible in the metaphysical sense. The second is irreversible in the human one.
This is the asymmetry that distinguishes assisted dying from almost every other domain in which clinical AI is being proposed. A misdiagnosis in radiology can, in most cases, be corrected by a second opinion or a subsequent test. A bad triage decision in an emergency department can be revisited as new information arrives. A wrong recommendation by a clinical decision support tool can be overridden by a clinician who notices something the system missed. The Zabolotnii, Holinko and Antonenko framework relies, throughout, on the existence of a human in the loop who can revise the system's output. Nitschke's proposal explicitly removes that human. The AI's answer is the answer. The pod, in his architecture, then enforces the answer without further review.
A defensible deployment of AI in capacity assessment, on the paper's logic, would begin as advisory only. It would generate an output. A trained clinician would review the output, would interview the candidate, would arrive at an independent finding, would compare. Only when the AI's outputs had been demonstrated to converge with skilled clinical judgement across a representative cohort, with measurable calibration and a documented evidence trail, would the question of expanded autonomy even arise. Even then, the irreversibility of the underlying procedure provides a principled reason to retain final human authority. The asymmetry of error makes the cost of one wrong call so high, and so unrecoverable, that no defensible trust metric is likely to be permissive enough to justify removing the human entirely.
The Regulatory Vacuum
The systems that have actually been built and deployed in clinical AI live within a regulatory regime that, with respect to autonomous life-ending decisions, does not yet exist. In the European Union, the AI Act entered into force on 1 August 2024, with the main applicability date for high-risk AI obligations set for August 2026. Medical devices that incorporate AI are classified as high-risk by default and required to comply with both the AI Act and the existing Medical Device Regulation. The Act mandates risk management, transparency, technical documentation, post-market monitoring, and meaningful human oversight. It does not, in its current form, contemplate the use of AI as the autonomous final adjudicator in an assisted-dying procedure. The category does not exist in the regulatory taxonomy. Whether such a system would be permitted at all, under the AI Act's prohibitions and high-risk provisions taken together, is an open question that has not been litigated because no one has yet tried.
The United States is more fragmented. The Food and Drug Administration regulates Software as a Medical Device through its Digital Health Center of Excellence, and has cleared hundreds of AI-enabled tools for clinical use. Almost all of them are deployed in a decision-support mode in which a clinician retains authority. The legal status of an autonomous AI that itself decides eligibility for medical-aid-in-dying in the states where the practice is permitted has never been adjudicated. The state statutes were written to govern the conduct of physicians, not algorithms. A model that produced an eligibility decision would not, on its face, be the kind of actor the statutes contemplate.
The United Kingdom is in the awkward position of having no current statute for assisted dying and a fragmented regulatory regime for clinical AI. The Medicines and Healthcare products Regulatory Agency has issued software-as-a-medical-device guidance and is developing the AI Airlock sandbox for testing of higher-risk AI applications. The Ada Lovelace Institute, in its May 2025 report on facial recognition governance and in subsequent publications on clinical AI, has argued that the UK lacks the statutory framework required to govern the deployment of high-risk biometric AI in any setting, let alone in life-ending decisions. There is no UK regulator with the authority, at present, to license or refuse the deployment of an AI capacity-assessor for an assisted-dying procedure if such a procedure were to be permitted by future legislation.
Switzerland, where Nitschke's pod first operated, is in a stranger position again. The country has long permitted assisted suicide under the relatively permissive provisions of Article 115 of the Penal Code, which criminalises assisting suicide only when done for selfish motives. There is no specific Swiss statute that governs the eligibility assessment for assisted dying, which is in practice carried out by clinicians within the right-to-die associations. After the September 2024 use of the Sarco pod, the Swiss minister for health, Elisabeth Baume-Schneider, said in parliament that the device did not meet the requirements of product safety law and that the use of nitrogen was not legally compliant. The prosecution then collapsed when the homicide charge against Willet was withdrawn. The pod has not been used since, but the absence of a clear regulatory determination means that no court has authoritatively decided whether a future capacity-assessment AI integrated into such a device would be permissible. The vacuum is real. It is the vacuum into which Nitschke's January 2026 announcement was made.
Who Is Accountable When the System Is Wrong
If a Spanish psychiatrist working under Organic Law 3/2021 wrongly assesses capacity, the responsibility chain runs through professional regulation, civil liability, and, in serious cases, criminal investigation. The clinician is identifiable. Their training is documented. Their professional indemnity insurer is on the hook for compensable harm. The Guarantee and Evaluation Commission is the procedural oversight body. The system has its critics, but it has actors who can be named and held to account.
The chain is not the same for an AI assessor. A model is, in any meaningful legal sense, not a person. It cannot hold a professional registration. It cannot be deposed. It cannot be struck off. The candidate liable parties are the developer who built and trained the model, the operator who deployed it, the clinician (if any) who reviewed its output, the regulator who licensed its use, and the procedural body that integrated it into the assessment workflow. The history of liability in clinical AI, such as it is, suggests that none of these is currently a satisfactory locus. Developers point to terms of service that disclaim responsibility for clinical decisions. Operators argue that they followed the manufacturer's instructions. Clinicians, where present, often defer to the algorithmic output and treat it as authoritative. Regulators license tools at the level of the device rather than the deployment.
This pattern of distributed and diluted accountability has been documented in domains as varied as algorithmic hiring, predictive policing, child-welfare screening and welfare fraud detection. The pattern arises not by accident but by design. The procurement structures of public administration favour the procurement of tools whose vendors carry the technical expertise and the legal liability disclaimers, and where the deploying institution can present the algorithmic output as merely advisory while in practice treating it as binding. The drift is consistent with the structural pressures that make a capacity-assessment AI attractive in the first place: it is cheaper than a psychiatric consultation, it is faster than a panel review, it is more deniable than a human judgement, and the responsibility for its errors can be allocated across a chain of actors none of whom carries the whole weight.
A defensible accountability regime for an AI capacity-assessor would have to invert most of those incentives. It would have to require named clinical responsibility for every deployment. It would have to mandate publication of model cards, training data composition, demographic performance, and calibration curves. It would have to provide the candidate with a meaningful right of contest before, not after, the procedure is enacted. It would have to assign liability for catastrophic error to a party that has the resources and the legal exposure to take the design choices seriously. None of these requirements is technically infeasible. None of them is currently in place.
What the Standard Would Have to Be
What standard of evidence, accountability and explainability should be required before AI is permitted to substitute for clinical human judgement in assisted-dying eligibility, and who bears responsibility when the system errs? The components of an honest answer can be sketched.
The first component is independent validation on the population to which the system would be applied. Not on a generic clinical cohort, not on the records the model was trained on, but on a representative sample of candidates with their own demographic, linguistic and diagnostic particularities. The validation has to include stratified performance reporting: by age, sex, ethnicity, diagnosis, language of assessment, socioeconomic background. The Buolamwini and Gebru paradigm applies here as elsewhere. An AI that performs well in aggregate while performing badly on identifiable subgroups is, for the purposes of an irreversible decision affecting members of those subgroups, an unsafe instrument.
The second component is calibrated and explainable confidence. The Zabolotnii, Holinko and Antonenko framework offers a vocabulary for this. The system must report not only its decision but the calibration of that decision against external evidence. It must articulate the reasoning chain in a form that a human reviewer can audit. The contemporary literature on explainable AI in clinical decision support is unsparing on the limits of post-hoc explanation: saliency maps and attention visualisations are widely accepted within the machine-learning community to be unreliable as faithful accounts of model behaviour. A capacity-assessment AI that cannot produce a contemporaneous, auditable reasoning chain that a clinician can independently verify is not a candidate for autonomous deployment.
The third component is meaningful human authority. The staged-autonomy framework is, on its own terms, a framework for graduated reduction of human oversight as the system earns the right. In the highest-stakes application, an irreversible procedure with categorically asymmetric error costs, the principled reading of the framework is that the highest stage is not reached. The human stays in the loop, with final authority, throughout the system's operational life. The AI's role is to enrich the clinical judgement, to flag inconsistencies, to surface the patterns that a tired clinician might miss. The role is not to displace the judgement.
The fourth component is real contestability. The candidate, before the decision is acted upon, must have the right to know that AI was used, what it concluded, what the underlying evidence was, and to obtain a substantive review of the decision by a different clinician or panel that is not bound by the system's output. The review has to be funded. Legal aid for capacity disputes in assisted-dying cases has, in most jurisdictions, never been adequately resourced even for human-only decisions. It would have to be restored as a precondition of any AI deployment.
The fifth component is the accountability regime described above: named clinical responsibility, mandated transparency, clear liability allocation, and an independent regulator with audit powers. The European Union's AI Act is the closest existing instrument to the kind of framework this implies, and even the AI Act does not yet contemplate the specific case. The work of writing the regime is, at the moment, work that has not been done.
Against this five-part standard, Nitschke's January 2026 proposal does not even rise to a starting position. There is no independent validation. There is no published calibration. The human authority has been explicitly removed. There is no contestability mechanism. There is no accountability regime, because there is no statute, no regulator, and no jurisdiction that has agreed to host the system. What there is, instead, is a press conference, an underlying ideology that locates the right to die in the autonomy of the individual to the exclusion of every other social good, and a 3D-printed capsule sitting in a workshop somewhere in continental Europe.
The Castillo Ramos case in Spain illuminates the alternative. Her capacity was assessed, contested, re-assessed, litigated through five levels of courts, and finally confirmed not because the system was efficient but because the system included multiple human decision-makers, each accountable to a professional regime and a public, who could be made to defend their conclusions in open court. The proceedings were slow, painful, and at moments inhumane. They were also the proceedings the law specifies, and the proceedings whose existence makes the eventual finding of capacity legible as a finding rather than as a verdict from inside a sealed box. To replace that process with a conversation between a vulnerable person and an avatar, with no appeal and no accountability and no audit trail beyond what the developer chooses to disclose, is not a refinement of the existing system. It is a different proposition. It belongs to a different jurisprudence.
The choice the next few years will pose is not a choice between human fallibility and machine reliability. It is a choice between two different kinds of fallibility, in a domain where both kinds are categorical, and where one kind comes attached to a chain of accountable persons and the other kind does not. The Zabolotnii, Holinko and Antonenko framework, by insisting that trust is something to be measured rather than asserted, offers the beginning of an answer to the question of when the substitution might be defensible. That answer, applied honestly to assisted dying, is: not yet, possibly not ever in the autonomous form, and only under a regime of staged authority and human supervision that nobody has yet built. The room described at the opening of this article, with its chair and its screen and its avatar, is not a future the law currently authorises in any jurisdiction on earth. The interesting question is whether the law will continue to refuse to authorise it once the technology is sold to states as an efficiency. The Sarco pod sits in a workshop. The avatar exists in beta. The case for the standard, against the case for the procurement, is what the next legislative cycle will decide.
References
- Euronews. “The inventor of the 'suicide pod' says AI should decide who can end their life.” 22 January 2026. https://www.euronews.com/next/2026/01/22/the-inventor-of-the-suicide-pod-says-ai-should-decide-who-can-end-their-life
- UNILAD Tech. “Inventor of controversial 'suicide pod' says AI will judge if a person is fit to use the machine.” 27 January 2026. https://www.uniladtech.com/news/ai/sarco-pod-inventor-ai-judge-person-fit-use-machine-554632-20260127
- Exit International. “AI, Sarco & the Right to Die.” https://www.exitinternational.net/ai-sarco-the-right-to-die/
- Serhii Zabolotnii, Viktoriia Holinko and Olha Antonenko. “From Black-Box Confidence to Measurable Trust in Clinical AI: A Framework for Evidence, Supervision, and Staged Autonomy.” arXiv:2604.26671, 29 April 2026. https://arxiv.org/abs/2604.26671
- CNN. “Noelia Castillo spent 20 months battling to die under a euthanasia law. On Thursday, Spain let her.” 26 March 2026. https://www.cnn.com/2026/03/26/europe/euthanasia-spain-noelia-castillo-latam-intl
- SWI swissinfo.ch. “After the first 'Sarco pod' death, will Switzerland introduce stricter rules for assisted suicide?” https://www.swissinfo.ch/eng/assisted-suicide/after-the-first-sarco-pod-death-will-switzerland-introduce-stricter-rules-for-assisted-suicide/88824081
- CNN. “Sarco suicide pod: Arrests after American woman dies in Switzerland in controversial machine.” 24 September 2024. https://edition.cnn.com/2024/09/24/europe/switzerland-arrests-sarco-suicide-capsule-intl-hnk
- Wikipedia. “Sarco pod.” https://en.wikipedia.org/wiki/Sarco_pod
- The Last Resort. “Prosecutors investigate after suicide of former Sarco boss Willet.” https://www.thelastresort.ch/3140-2/
- Exit International. “Dr Philip Nitschke.” https://www.exitinternational.net/about-exit/dr-philip-nitschke/
- Spanish Government. Organic Law 3/2021, of 24 March, regulating euthanasia. Boletín Oficial del Estado, 25 March 2021. https://www.boe.es/buscar/act.php?id=BOE-A-2021-4628
- Thomas Grisso and Paul S. Appelbaum. “The MacCAT-T: a clinical tool to assess patients' capacities to make treatment decisions.” Psychiatric Services, vol. 48, no. 11, 1997. https://pubmed.ncbi.nlm.nih.gov/9355168/
- Scott Y. H. Kim, Raymond G. De Vries and John R. Peteet. “Euthanasia and Assisted Suicide of Patients With Psychiatric Disorders in the Netherlands 2011 to 2014.” JAMA Psychiatry, vol. 73, no. 4, 2016.
- Frontiers in Psychiatry. “Physician Assisted Death for Psychiatric Suffering: Experiences in the Netherlands.” 2022. https://www.frontiersin.org/journals/psychiatry/articles/10.3389/fpsyt.2022.895387/full
- PubMed Central. “Exploring the interplay of clinical, ethical and societal dynamics: two decades of Medical Assistance in Dying (MAID) on psychiatric grounds in the Netherlands and Belgium.” https://pmc.ncbi.nlm.nih.gov/articles/PMC11422721/
- Government of Canada. “Final Report of the Expert Panel on MAiD and Mental Illness.” https://www.canada.ca/en/health-canada/corporate/about-health-canada/public-engagement/external-advisory-bodies/expert-panel-maid-mental-illness/final-report-expert-panel-maid-mental-illness.html
- UK Parliament. Terminally Ill Adults (End of Life) Bill 2024-25. https://bills.parliament.uk/bills/3774
- House of Commons Library. “The Terminally Ill Adults (End of Life) Bill 2024-25.” https://commonslibrary.parliament.uk/research-briefings/cbp-10123/
- Joy Buolamwini and Timnit Gebru. “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification.” Proceedings of Machine Learning Research, vol. 81, 2018. https://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf
- Oxford Open Digital Health. “Gender and racial bias unveiled: clinical artificial intelligence (AI) and machine learning (ML) algorithms are fanning the flames of inequity.” https://academic.oup.com/oodh/article/doi/10.1093/oodh/oqaf027/8279897
- npj Digital Medicine. “Racial bias in AI-mediated psychiatric diagnosis and treatment: a qualitative comparison of four large language models.” https://www.nature.com/articles/s41746-025-01746-4
- European Union. Regulation (EU) 2024/1689 (the AI Act). Entered into force 1 August 2024. https://artificialintelligenceact.eu/
- Ada Lovelace Institute. “An eye on the future: Examining the UK's approach to facial recognition technology governance.” May 2025. https://www.adalovelaceinstitute.org/report/an-eye-on-the-future/
- PMC. “Explainable AI in Clinical Decision Support Systems.” https://pmc.ncbi.nlm.nih.gov/articles/PMC12427955/
- Swiss Federal Council. Article 115, Swiss Penal Code. https://www.fedlex.admin.ch/eli/cc/54/757_781_799/en
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
