In the previous Megamachine papers we described the obsolescence of technical intelligence through individual examples — as the creeping disappearance of occupations whose knowledge and skill carried certain industrial structures for decades, and whose function is today being taken over by other structures, other actors, other tools. This paper deepens the diagnosis through a concrete case — the medical profession — and at the same time deepens it by a second layer that until now had only been hinted at: the question of responsibility. What happens when the profession loses its core competence but the responsibility for the consequences of that competence continues to rest with it? What happens when the substance migrates but the liability stays? Unlike the previous papers, the subject this time is not specifically German or European. The medical profession is a global one, the AI tested is American, the study comes from Boston, the operational reality affects doctors and patients worldwide. Where the essay turns to the German or European situation, it does so deliberately — in the case of the liability architecture, which is built differently in every legal system, and in the case of political steering, which can be pinned to concrete actors. Otherwise, however, the picture is an international one.
The occasion for this paper is a study published on 30 April 2026 — yesterday, the day before this paper — in the journal Science. It was presented by a team led by Adam Rodman, internist at Beth Israel Deaconess Medical Center, and Arjun Manrai, bioinformatician at Harvard Medical School, with the participation of several further authors — Peter Brodeur, Thomas Buckley, Zahir Kanjee, Ethan Goh, Jonathan Chen. The study tests the diagnostic ability of OpenAI's language model o1-preview on two-tier material — 143 unusual case histories from the New England Journal of Medicine and 76 patient records from the emergency department of Beth Israel Deaconess hospital in Boston. It arrives at a result that the authors themselves describe as "downright shockingly good", and one that has a historical reference — the task of having advanced computers make medical diagnoses through logical inference had been defined as a benchmark in a scientific publication in 1959. Sixty-seven years later, the authors say, the goal has been reached.
The study would deserve a bigger public debate than it is likely to get. Over the next few weeks it will appear in the science sections of the major newspapers — Der Spiegel published its report yesterday evening — and then attention will drift away. The next AI news item will follow, the next sensational study, the next reassuring formula from the researchers in charge. What remains is the operational reality that was there long before the study: according to the "Clinician of the Future" report, half of all doctors worldwide already use AI, and one in five regularly seeks a second opinion from it. Manrai puts it in his own words: "The genie is out of the bottle."
The paper has two points. The first is that the medical profession is caught up in a process which the Megamachine series calls a shift of substance — the core competence migrates, the shell remains. The second is that the legal architecture of responsibility that applies to this case in Germany and the EU produces a constellation we may call Kafkaesque: the doctor loses his competence and keeps his liability. The AI gains his competence and does not take over his liability. What stands between the two is a diffuse responsibility — which, structurally, is exactly what the Megamachine defines as its mode of operation.
I. What the Study Found — the Exact Figures
Methodologically, the study is built in three parts. Part one: 70 selected NEJM cases, on which language models, specialised medical software and human doctors with aids (search engines, specialist literature) are compared. Part two: 76 real emergency department records from Beth Israel Deaconess Medical Center, on which the model competes against two emergency physicians working there. Part three: five clinical case studies, on which the model competes against 46 human doctors in so-called "management reasoning" — that is, in developing treatment plans, for instance on administering antibiotics or planning end-of-life conversations.
The results. In part one — the NEJM cases — OpenAI o1-preview produced the correct or nearly correct diagnosis in 88.6 per cent of cases. Other language models and specialised software scored lower, human doctors considerably lower. On a subset of the cases the model achieved just under 80 per cent correct diagnosis in the narrowest sense — that is, as the top-ranked of the differential diagnoses it proposed.
In part two — the real emergency department — o1-preview achieved a hit rate of 67.1 per cent at triage, the first diagnostic step in the emergency department. The two experienced Beth Israel Deaconess doctors serving as the comparison baseline came in at 55.3 and 50.0 per cent. As more information became available, the model's accuracy rose to 82 per cent — the human doctors reached 70 to 79 per cent. With greater density of information, this difference was no longer statistically significant; at triage, with the least information, the difference was most pronounced.
In part three — management reasoning — the model achieved 89 per cent. The 46 human doctors, who worked with the help of conventional tools such as search engines and specialist databases, achieved 34 per cent. That is the sharpest difference in the entire study. It concerns not just diagnosis but the next step — the treatment plan.
These results have a few important properties that ought to shape how the study is perceived. First, the model tested here is o1-preview — the predecessor of the o1 model released in December 2024. Both models have a pre-training cutoff of October 2023. That means: the models on the market today, in May 2026 — o1, o3, GPT-5, Claude 4 and so on — are considerably superior to the models tested. The study measures a lower bound, not an upper bound.
Second, the comparison figure on the doctors' side is the performance of experienced internists or emergency physicians with the aids customary today. It is not the performance of beginners in the profession, not the performance of general practitioners without specialisation, not the performance of doctors in under-served regions, in developing countries, in overloaded emergency departments. On the overall distribution of world medicine, the performance of the comparison group ranks very high. What the model beats here is the upper half of the professional distribution.
Third, the study is retrospective. It tests the model on records already documented, with known diagnostic outcomes. It does not test the model on real patients in real time. As the study's authors themselves stress, the next step would be a prospective clinical trial — one that would test the model under real conditions to the usual standards of medical research. That will happen over the next few years. Its results are not known today, but they will in all likelihood confirm the direction of the results presented here, possibly sharpen it.
II. The Historical Benchmark — What Was Promised in 1959
The study has a historical reference which Der Spiegel's coverage mentions briefly but does not elaborate. In 1959 — when computers still worked with punched cards — two American scientists, Robert S. Ledley and Lee B. Lusted, had published an article in the journal Science entitled "Reasoning Foundations of Medical Diagnosis". They had put forward the thesis that medical diagnosis is at its core a process of logical inference — and that advanced computers, as soon as they became available, ought to be able to carry out these inferences just as reliably as a human doctor.
Ledley and Lusted did not formulate this as a threat to the profession but as a promise of relief. Computers would free the doctor from the routine part of his work and give him time for what goes beyond diagnosis — the relationship with the patient, informing the patient, treatment planning, individual adjustment. In 1959 this thesis was utopian. It remained so over the following six decades. Various attempts to develop medical expert systems — MYCIN in the 1970s, INTERNIST-1 in the 1980s, many others — failed to deliver on the thesis. The systems were too rigid, too narrowly confined to particular clinical pictures, too difficult to integrate into everyday clinical practice.
What has changed since 2022 — with GPT-3.5, with ChatGPT, with GPT-4, with reasoning models such as o1 — is a different kind of AI. It is not rule-based. It has no explicit medical knowledge base. It is trained on vast quantities of text, part of which is medical literature, but the greater part everything else — books, internet texts, forums, code, translations. It has no special medical training. And precisely because it is nothing special, it can do something the specialised expert systems never could: it can cope with the disorder of the real patient record — with typing errors, with missing information, with context-dependent clues, with the testicular pain that was mentioned in the initial consultation and then forgotten.
Adam Rodman, the study author, describes in the article and in his interviews a concrete case that stuck in his memory in particular. A patient comes to the emergency department with a cough and severe breathing difficulties. In the initial consultation he also mentions testicular pain. The doctors concentrate on the breathing difficulties — the obvious main symptoms. The testicular pain is forgotten. Only 24 hours later, when necrotising fasciitis develops in the patient's genital area — an aggressive, life-threatening inflammation of the tissue that can only be treated with the scalpel — is the connection recognised. The model, Rodman says, would not have made that mistake. It would have linked the testicular pain with the information that the patient had undergone an organ transplant and was taking immunosuppressive drugs — and would therefore have taken the pain more seriously.
The point of this case is not that the doctors were bad. The point is that in an overloaded emergency department, with shift changes, with handovers, with a hundred other cases at the same time, human memory and human attention have limits that a language model does not have. What was promised in 1959 was relief for the doctor from routine. What is being delivered in 2026 is a relief that at the same time relieves the doctor of his diagnostic core competence.
III. The Profession and Its Pillars
The medical profession is one of the oldest in documented history. Hippocrates lived around 460 BC. The collection of Hippocratic writings — the "Hippocratic Oath", the Aphorisms, the Prognostics — forms the basis of a professional ethic that, in modified form, still holds today. The medieval universities of Bologna, Padua and Salerno institutionalised medical teaching. The Paris clinic of the early modern period, the Berlin Charité system of the 19th century, the Johns Hopkins reform of American medical education around 1900, the Flexner reform of 1910 — these shaped the medical profession in its present form.
The profession rests on four pillars. First pillar: specialist knowledge — anatomy, physiology, pathology, pharmacology, diagnostics, therapy, acquired over some twelve years of study and specialist training. Second pillar: diagnostic competence — the ability to infer from the findings gathered the right disease, which is decisive for treatment. Third pillar: treatment competence — the ability to choose the right therapy on the basis of the diagnosis, to adapt it individually, to observe its effect, to modify it where necessary. Fourth pillar: professional responsibility — the legally and morally defined duty to answer for one's own work, to be liable with one's own licence to practise, to appear in court in the event of harm.
The Harvard study shows that the second pillar — diagnostic competence — is today not only matched but surpassed by an AI. It further shows that the third pillar too — treatment competence in the sense of management reasoning — is surpassed by the AI, in the specific test even dramatically (89 to 34 per cent). What remains is the first pillar (specialist knowledge) and the fourth pillar (responsibility). These two pillars, however, stand in a difficult relationship to the world of AI. The specialist knowledge acquired in twelve years of study and specialist training is today largely available in the language model — retrievable faster, more up to date and more interconnected than in human memory. Responsibility — which we will deal with in the following chapters — is the pillar that, under today's legal constructions, remains with the doctor even when the diagnosis was made by the AI.
We can now ask more precisely what is happening to the profession here. It is not losing its function — patients still go to doctors, they are still treated, the hospital keeps running. It is losing its justification — the professional lead over the layperson that was justified by the long years of training and the institutional calling. What the profession was — a group that administered scarce and hard-to-access knowledge and received social recognition, high pay and institutional autonomy in return — is today a group that administers increasingly accessible knowledge and must find its justification anew. The pillar of specialist knowledge has shifted. The pillars of diagnostic and treatment competence are eroding empirically. What remains is the pillar of responsibility — and it becomes the load-bearing pillar of a profession whose other pillars no longer bear the load.
IV. The Figures from Practice — the Genie Is Out of the Bottle
The Harvard study measures what AI can do today. A second question is what AI is already doing today. Here the study does not supply the central statistic, but in their interviews Rodman and Manrai cite another finding — the "Clinician of the Future" report by Elsevier Health, the latest edition of which was produced for 2025. According to this report, half of all doctors worldwide now use artificial intelligence in some form. One doctor in five regularly seeks a second opinion from an AI — typically from ChatGPT, Claude, Gemini or a specialised medical assistant. More than half of all doctors say they intend to use AI-supported second opinions for complex cases.
These figures are the operational reality into which the Harvard study is being written. They make the study at once superfluous and urgent. Superfluous, because what the study validates scientifically has long been happening in practice — doctors ask the AI. They do so informally, often without documentation, often without informing the patient, often without clear standards. Urgent, because what is happening in practice is happening without scientific validation, without clinical trials, without standards of application. The operational reality runs ahead of institutional clarification — a pattern we know from the previous Megamachine papers, but which appears here with particular sharpness.
Der Spiegel's coverage carries Manrai's own statement: "The genie is out of the bottle." That is a remarkable formulation. It concedes that the process can no longer be stopped. It concedes at the same time that the doctors who already use AI today are doing so without the conditions under which they ought to be doing it having been clarified. What needs to be done now, according to Manrai, is clinical trials — that is, retrospective comparisons of AI-supported and non-AI-supported treatment. That will take years. Until the trials are available, AI-supported medicine carries on. Until the trials are available, more doctors will delegate their diagnostic practice to AI. Until the trials are available, more patients will be treated in a constellation whose conditions have not been clarified.
An observation in passing: what Der Spiegel's coverage reports as the experience with writing doctors' letters — "Expectations were enormous: two or even three hours a day, it was thought, could be saved by doctors thanks to AI and devoted to their patients instead. Reality, however, looks different. The time saved is astonishingly small" — is the standard experience with productivity gains through IT over the last 30 years. It goes back to what the Nobel laureate in economics Robert Solow formulated in 1987: "You can see the computer age everywhere but in the productivity statistics." AI in medicine follows this pattern. It does not replace hours of work but shifts them — from diagnosis to validation, from interpreting findings to correcting, from direct experience to administering the AI output. What doctors gain is not time for patients. What they gain is a changed structure of activity in which the AI does not relieve but transforms.
V. The Legal Architecture — § 630a BGB and the EU AI Act
At this point we change perspective — from the medical reality to the legal architecture in which this reality is embedded in Germany. Here lies the paper's second point.
The legal situation in Germany is not as open as the Spiegel article suggests with its rhetorical question "Who takes responsibility when the AI errs?" It is clearly regulated — and in such a way that liability stays with the doctor.
First, the German Patients' Rights Act of 2013 anchored the treatment contract in §§ 630a to 630h of the German Civil Code (Bürgerliches Gesetzbuch, BGB). Under § 630a BGB the doctor owes treatment "in accordance with the generally recognised professional standards existing at the time of treatment". That is the general basis of liability. It applies to every doctor, in every treatment, with or without AI.
Second, the EU AI Act of 1 August 2024 — Regulation 2024/1689 — defines most medical AI systems as "high-risk AI" under Annex III or via the Medical Device Regulation, Rule 11. Whoever uses such a system becomes, under Art. 3(4) of the EU AI Act, the "deployer" of the AI. That has consequences. Under Art. 26 of the EU AI Act the deployer must fulfil six obligations: AI literacy (that is, a sound understanding of the system), human oversight, checking the relevance of input data, continuous monitoring with a duty to report, logging for at least six months, and informing staff and patients. These obligations apply in full from 2 August 2026 — that is, in exactly three months. The AI literacy obligation (Art. 4) has already applied since 2 February 2025.
Third, and this is where it becomes Kafkaesque: under the existing construction, a doctor who uses AI is doubly liable in the event of a treatment error. First under § 630a BGB as the patient's contractual partner. Second under the deployer obligations of the EU AI Act. If he has not exercised sufficient human oversight — if, that is, he has adopted the AI diagnosis without critical examination — this counts under German patients' rights law as a "gross treatment error". In the case of a gross treatment error, § 630h BGB applies: the reversal of the burden of proof. It is not the patient who must prove that the error caused the harm. The doctor must prove that the error did not cause it. Legally, that is a very strong position for the patient — and a very weak one for the doctor.
Fourth, the manufacturer of the AI — that is, OpenAI, Anthropic, Google or others — is liable for product defects under product liability law. But the main burden remains with the doctor. The manufacturer can invoke the fact that the AI system was sold as a tool whose competent use is the responsibility of the user. The user is the doctor. The doctor is the one responsible.
Fifth — and this is the decisive finding —: in 2022 the EU Commission had drafted its own AI liability directive, which was meant to close precisely these gaps. It provided for distributing the liability risks between manufacturer, deployer and user more precisely, with stricter disclosure obligations for manufacturers and with eased rules of evidence for injured parties. This directive was withdrawn on 12 February 2025 by the new EU Commission under Ursula von der Leyen — the previous one had already been under her. The official justification: the aim was to "avoid bureaucratic over-regulation" and not to hinder innovation. That was a political decision in the interests of the AI industry. Its consequence is that the legal gap is not being closed. Liability stays with the doctor.
VI. The Kafkaesque Constellation — Competence Gone, Liability Remains
At this point the constellation we may call Kafkaesque becomes tangible. It has five layers which together form a picture that in its logic resembles the picture of the waiting patient in the EPS system — structurally, not in detail.
First layer: according to the Harvard study and to lived practice, AI diagnosis is better than the doctor's. Whoever fails to take note of this risks worse treatment. Whoever takes note of it and does not act risks treatment that falls short of the professional standard. That is direct pressure on the doctor to use AI.
Second layer: whoever uses AI takes on liability as deployer under the EU AI Act and German law. He is liable not only for his own work but also for the application of the AI — for the choice of system, for checking the plausibility of the output, for informing the patient, for documentation. If he does not exercise sufficient oversight of the AI, the gross treatment error with reversal of the burden of proof applies.
Third layer: the AI itself cannot be liable. It has no legal personality. Its manufacturer — OpenAI, currently valued at around 500 billion US dollars — is liable only for product defects in the narrow sense. For the diagnostic misadvice the model systematically produces (hallucinations, false correlations, recommendations out of context), the manufacturer is not liable — as long as the product was used within the scope of its intended application. What is "intended" is defined by the manufacturer in its terms of use.
Fourth layer: the patient is entitled to the professional standard. If AI today exceeds the professional standard, the patient may be entitled to have the AI used. This is a construction already being contemplated in legal articles, but not yet tested in practice. Whoever tests it might come to the conclusion that a doctor who does not use AI is likewise committing a treatment error.
Fifth layer: the doctor is thus in a constellation in which he faces two legal risks at the same time — if he uses the AI and it is wrong, he is liable as deployer; if he does not use it and a case ends badly, he may be liable for breaching the professional standard. That is a constellation of a duty of proof in both directions, in which the only means of reducing risk is very elaborate documentation. This documentation costs time. The time that AI was supposed to save is used up by the documentation that is needed to safeguard the use of AI in the first place. The reality of Der Spiegel's coverage — "The time saved is astonishingly small" — is no accident. It is the structural consequence of the liability architecture.
Whoever thinks this picture through to the end sees a constellation in which the doctor is caught in a legal vice he did not build himself. He is caught in it because the Patients' Rights Act of 2013 defined professional responsibility in a world in which there was as yet no medical AI. He is caught in it because the EU AI Act extends deployer responsibility for AI users without tightening manufacturer liability. He is caught in it because the AI liability directive that could have changed this was withdrawn in February 2025. He is caught in it without any single person, any single authority, any single decision having wanted the constellation. He is caught in it because so many decisions have meshed together that in the end nobody is responsible for the result any more.
VII. The Link to the Colombian Patient
At this point the link to the diagnosis of the Colombia paper becomes visible. There we described a patient in the EPS waiting room who loses a day because her membership has been suspended and nobody can explain why. The system administers her without treating her. Responsibility is diffuse — between the EPS, the hospital, the central office, the Constitutional Court, the failed reform.
What does that have to do with Boston, with OpenAI o1, with the Harvard study? A great deal. On the surface the two worlds are incomparable — the Colombian patient is poor, the Boston clinic is rich, the one is emergency administration, the other is cutting-edge research. At the structural level they are identical. In both cases diagnosis is separated from treatment, treatment from responsibility, responsibility from the person. In both cases the subject — the patient — disappears between aggregates and structures that do not exist for the patient but in which the patient merely occurs. In both cases the only person who could be held directly responsible — the counter clerk in Bogotá, the doctor in Boston — is forced into a position in which they bear the responsibility without being able to exercise it.
In Bogotá the counter clerk does not know why the membership has been suspended. She cannot change it. She can only refer it on to the central office, which is somewhere else. The responsibility keeps travelling without finding a destination. In Boston the doctor uses the AI because otherwise he would fall short of the professional standard. He checks the output because otherwise he would breach human oversight. He documents the check because otherwise he risks a gross treatment error. He cannot hold the AI to account because it has no legal personality. He cannot hold the manufacturer to account because the product was used within the scope of its intended application. He bears the responsibility alone.
The Bogotá constellation and the Boston constellation are two poles of the same Megamachine. In Bogotá it is visible because it is everyday. In Boston it is invisible because it is staged as cutting-edge technology. But the structural logic is the same. A formal architecture (Ley 100, the EU AI Act) lays down a distribution of responsibility that does not work in lived reality. Operational practice (the cancelled appointment, the AI-supported diagnosis) generates consequences for which nobody fully answers. The only person who can be held directly responsible (counter clerk, doctor) carries a burden they cannot control. The subject (the patient in Bogotá, the patient in Boston) loses their day, their treatment, their security.
What the Megamachine produces at these two poles is not injustice in the classical sense — no person is persecuting the patient, no person is persecuting the doctor. What it produces is a distribution of consequences without a distribution of responsibility. That is the definition of the Kafkaesque that we developed in the Colombia paper. It holds in Bogotá. It holds in Boston. It already holds today, and it deepens with every month.
VIII. Three Possible Paths — What May Come
We can sketch three scenarios into which the constellation may lead over the next ten years.
Path A — covert substitution. AI becomes the invisible co-diagnostician. Doctors use it, document their use minimally, maintain responsibility formally. Patients do not learn that their diagnosis was made with the help of AI. The supervisory authorities tolerate it because the alternative — across-the-board non-use — would, according to the study results, mean a worse standard of care. The profession shrinks slowly — fewer doctors can serve the same number of patients because the AI takes over the diagnostic routine. Medical training is adapted; the emphasis shifts from diagnosis to patient communication and treatment management. In twenty years the profession in its present sense will no longer exist — but nobody will be able to name a date on which that happened.
Path B — regulatory clarification. The EU AI Act is tightened in several steps over the coming years. The AI liability directive is tabled again in modified form. The manufacturers of medical AI systems are drawn into a liability structure of their own. Hospitals receive clear rules for integrating AI into everyday clinical practice. The profession is redefined — as a triad of doctor, patient and AI, in which each role has its own responsibilities and its own competences. That is the path the study's authors hint at in their interviews as the "triad". It is possible, but it requires a feat of political steering that the EU is not discernibly achieving today. The withdrawal of the AI liability directive in February 2025 is a bad omen.
Path C — legal escalation. The first major cases of harm come to court. A patient dies because a doctor trusted an AI diagnosis too much or too little. The family sues. The legal reckoning takes years. Insurers withdraw from AI medicine or make professional liability insurance considerably more expensive. Doctors become more reluctant to use AI — not for medical but for legal reasons. Patients are treated worse because legal uncertainty blocks the operational reality. Politics responds with emergency laws that do not clarify the constellation but merely shift it. The system hangs between the demands — a professional standard at AI level, legal responsibility at pre-AI level — and nobody can untie the knot.
Which path prevails will largely be decided in the next two to three years. The full application of the EU AI Act in August 2026 is one date. The first major cases of harm, which will probably reach the courts in 2027 or 2028, are a second. The EU's next wave of regulation, which will probably not come before 2028, is a third.
IX. What the Profession Could Do Today — and Does Not
At one point the diagnosis of this paper must ask what the profession itself is doing. It is not a passive victim of the Megamachine. It is an actor that can take a stance.
What it does today comes down mainly to two things. First, it uses AI in individual practice, often without documentation, often without reflecting on the situation regarding responsibility. Second, it speaks out through the major professional associations — in Germany, for instance, the German Medical Association (Bundesärztekammer) and the National Association of Statutory Health Insurance Physicians (Kassenärztliche Bundesvereinigung), in the USA the American Medical Association, in Britain the British Medical Association, with a similar tone in most other countries — in a tone that wavers between reassurance and alarmism. The reassurance says: "Nobody should now be saying we no longer need doctors" — Adam Rodman in his own words, identical in the German professional associations, identical in the American ones, identical in the British ones. The alarmism says: "AI endangers the medical profession" — typically in ceremonial Sunday speeches, without consequences.
What it does not do: it does not conduct the strategic discussion that would be necessary. It does not negotiate with the EU about an appropriate liability architecture. It does not develop its own standards for the use of AI in hospital and practice. It does not invest in research structures of its own that would force the AI manufacturers to become more transparent. It does not organise its own substance — training and continuing education — so as to adapt it to the new situation. It does what many established professions have done in comparable situations — it hopes the wave will pass.
That is the pattern we know from the other Megamachine papers. The machine-tool industry did it for twenty years, until DMG Mori became a Japanese subsidiary. Medium-sized industrial firms in many countries are doing it today, while they are sold to Chinese, American or other investors. Established political structures of the centre in several European democracies are doing it, while right-wing populist forces overtake them. The medical profession is in no different a situation from other established structures that are losing their substance. It does not react more wisely. It does not react earlier. It does not react more strategically.
What it could do would be, first, to address the legal situation openly — not in Sunday speeches but in a public debate in which citizens understand that the architecture of responsibility is Kafkaesque and that the burden remains with the doctor alone. Second, to develop its own standards for integrating AI, standards not imposed by the manufacturers or the regulators but devised by the profession itself. Third, to adapt training — less factual knowledge, more critical engagement with AI, more patient communication, more ethical judgement. Fourth, to organise politically in order to push EU regulation in a direction that holds the manufacturers responsible and relieves the users.
Today, in May 2026, there is no sign of any of these four paths.
X. The Lesson of the Megamachine
The Megamachine, as we have described it in the previous papers, is a socio-technical system that, through entanglement, produces consequences for which nobody personally answers. It works not through commands but through structures. It works not through manipulation but through things taken for granted. It works not through conspiracy but through diffuseness.
The case of medical AI is a particularly clear variant of this machine. Nobody wanted the constellation. In 1959 Ledley and Lusted formulated a scientific programme in which no shift of power was intended. The OpenAI researchers developed a language model that finds many applications, of which medicine is only one. In 2013 the German federal government passed a Patients' Rights Act that was meant to secure professional responsibility. In 2024 the EU enacted an AI Act that was meant to regulate high-risk AI. In 2025 the EU Commission withdrew a liability directive because it did not want to hinder innovation. Doctors use AI because they must, in order to offer their patients the professional standard. Each of these decisions was justifiable. Together they produce a constellation in which the profession loses its substance and keeps its liability.
That is the Megamachine. It is not a villain. It is not a conspiracy. It is the accumulation of small decisions, each rational in itself, which taken together produce a picture that nobody explicitly wanted. Whoever lives in the picture cannot change it, because the picture does not come from any one place where it could be changed. It comes from the entanglement.
What the medical profession is going through today is structurally similar to what established structures in many other fields have gone through or are going through — from the machine-tool industry, whose German example we described at length in the DMG Mori case, to law firms, architectural practices, publishing houses, programmers' workshops and teaching staffs. Other occupations will follow. The order will be determined not by the importance of these occupations but by the speed at which the respective activities become substitutable by AI. Medicine is, paradoxically, one of the early cases, because its diagnostics are particularly well represented in the training data of the large language models — the medical literature is denser, more structured, more accessible than the literature of many other fields. That holds for diagnostics in Boston as in Berlin, in Bombay as in Bogotá. What varies between locations is not the capabilities of the AI but the legal, institutional and economic conditions under which it is deployed — and with them the distribution of the burden its introduction creates.
The profession loses its competence and keeps its liability. That is the definition of a Megamachine that traps its user: it changes what he can do without changing what he must answer for.
XI. What the Study Does Not Say — Three Omissions
A final observation concerns the study itself. It is methodologically sound, its results are plausible, its self-criticism is open. But it has three omissions that are important for understanding the logic of the Megamachine.
First omission: it says nothing about the ownership structure of the AI tested. OpenAI is an American company currently valued at around 500 billion US dollars, with Microsoft as a major shareholder (49 per cent), with a board that is predominantly of Silicon Valley profile. The study tests a product of this company and validates it scientifically. That is not necessarily ideology — the researchers took the best model available. But it means that the scientific infrastructure — Harvard, Beth Israel Deaconess, the most renowned science magazine in the world — is today becoming the validation infrastructure of a commercial product. Who served whom here? The study is silent on this.
Second omission: it says nothing about geographical distribution. It is a Boston study, at a teaching hospital of Harvard Medical School, with patients from the Greater Boston area. That is a particular population — predominantly well insured, predominantly with access to highly developed diagnostics, predominantly native English speakers, predominantly with clinical pictures that are well documented in North American medicine. What happens when this model is used in an Indian district hospital, in a rural hospital in Colombia, in an overloaded emergency department in Berlin's Neukölln district? The training data come predominantly from the West, the diagnostic patterns are Western, the clinical pictures are Western. Whoever deploys diagnostic software exports a world view. The study does not address this.
Third omission: it says nothing about the strategic consequences for the profession. The study says: the AI is good. It says: we need further clinical trials. It says: nobody should now be saying we no longer need doctors. It does not say: what will happen to medical training over the next ten years? What to staffing plans for doctors? What to hospital structures? What to the remuneration architecture, which today rests on the assumption that the doctor is the diagnostic bottleneck? These questions cannot be answered in a single study. But they are the questions that ought to be asked now — and that nobody is asking systematically.
What the study does is deliver the main scientific headline and leave the strategic follow-up questions open. What the coverage does is recite the main scientific headline and quote the researchers' reassuring formula. What the profession does is follow the coverage and believe the reassuring formula. What patients do is walk into a world in which their diagnosis has long ceased to be made by a single doctor, but by a constellation in which the doctor is one voice among several — and the only one that is liable.
XII. Conclusion — Where the Triad Ends
Adam Rodman closes his interview with the image of the triad — doctor, patient, AI in a new model of cooperation. "The profession of doctor has existed for thousands of years," he says. "Medicine is a very fulfilling field, and it will remain so, despite this new technology." That is a hopeful image. It is also an image that does not ask whether the triad is stable as a balance of the three actors, or whether, like most triads, it is reduced over time to a dyad.
The history of technological substitution teaches that triads are often transitional forms. If in ten years AI can reliably take over diagnostic and management work, the question will be what the doctor still contributes to the triad. If the answer is "responsibility", then the triad is not a form of cooperation but a form of risk distribution — the doctor as an insurance vehicle for the AI, which cannot itself be liable. On this reading, the triad is not a balance but a transition to a new constellation in which the AI makes the diagnosis and the plan, the doctor bears the liability, and the patient stands between the two — hoping that the AI is right, because the doctor cannot check the output anyway.
The Megamachine lesson of this case can be put in a single sentence. The substance migrates, the shell remains — we know the sentence from the DMG Mori paper. In the Index–Makino paper we saw that the shell can be a foundation. In the present paper we see that the shell can also be a profession — a 2,400-year-old profession that over the next ten to twenty years will lose its substance to a machine whose liability, under today's legal constructions, it nevertheless has to bear.
What remains to be done? It remains to describe the constellation before it becomes invisible. It remains to redistribute responsibility before it finally diffuses. It remains to put the profession, at some point, in a position to steer its own transformation before it is merely an object. Today, on 1 May 2026, little of such steering is to be seen. The study's authors formulate their results cautiously. The professional associations are silent. Politics regulates on the basis of outdated assumptions. Patients do not learn what is happening. The Megamachine works quietly — as always.
There will come a day when a court — in Germany, in the USA, in Britain, somewhere in a country with highly developed medicine and a well-developed body of patients' rights case law — hears the first major case of harm in AI medicine. On that day the legal constellation will become public. By then it ought already to have been recognised — not by the court, but by the profession, by politics, by patients. Whoever turns to it only once the harm has occurred will no longer be able to change the constellation. Whoever turns to it today perhaps can.
That is the only answer the present paper can give to the question "What is to be done?" Engage. Observe. Differentiate. Do not reassure. Do not alarm. Do not follow the reassuring formulas of the study's authors, nor the alarmist formulas of the critics. Instead, endure the constellation in its duality — the AI really is diagnostically better, and the doctor really does bear the liability. Both sentences are true. From their combination arises the Kafkaesque of our time.