The Robodoc

Diagnostic AI

The Patient Who Came Prepared

The story began, as so many diagnostic encounters now do, not in a clinic but in front of a screen. The patient had uploaded raw data from a consumer ancestry test to an AI assistant and asked a straightforward question: what does this tell me about my health? The response arrived in seconds. Among a range of findings was a specific flag - two variants in a gene called SLCO1B1, identified by their SNP designations, with a risk figure attached: a 17-fold elevated likelihood of myopathy - muscle damage - from statin use. The source cited was a landmark 2008 study from the SEARCH Collaborative Group, published in the New England Journal of Medicine. That finding had been in the scientific literature for seventeen years. The patient's genomic data had contained the relevant variants for just as long. What changed was that an AI could now read both and connect them in the time it takes to compose a message.

The GP, presented with this information, made the right call. He prescribed Ezetimibe - a cholesterol-lowering drug that does not carry the same myopathy risk for this genetic profile - rather than the statins he had intended. No harm was done. The appointment concluded as normal. And the reason no harm was done was that the patient had conducted a pharmacogenomic analysis before walking through the door - the kind of analysis that, had the standard system been left to itself, would have required a specialist referral and weeks of waiting, if it had been ordered at all.

A separate encounter around the same time illustrated a different facet of the same shift. A patient photographed a printed ECG and fed the image to a consumer AI assistant. The AI's response was rapid and specific: the electrodes had been placed incorrectly, inverting the waveform. This was a procedural error, not a cardiac finding - the kind of mistake that a time-pressured reading could easily misinterpret as an arrhythmia, or that might generate false reassurance when an actual problem existed. The AI caught it before any clinical interpretation had taken place.

These are small stories in the sense that no one died, no emergency was averted in dramatic fashion, and nothing went on any official record. They are large stories in the sense that they describe something structurally new: patients arriving at clinical encounters already in possession of analyses that the appointment system was not designed to generate. The asymmetry that has defined medicine for centuries - the doctor holds the diagnostic knowledge, the patient presents symptoms - is beginning, quietly and unevenly, to shift.

Pattern Recognition

Seeing What Human Eyes Miss

The evidence base for AI diagnostic performance has grown from promising to substantial over the past decade. In 2016, a Google research team published a study in JAMA describing the training of a deep learning algorithm on 128,000 retinal fundus photographs, each evaluated by between three and seven ophthalmologists from a panel of 54. When tested on clinical validation sets of approximately 12,000 images, the algorithm achieved an F-score of 0.95 - compared to a median of 0.91 for the ophthalmologist panel. The task was detecting diabetic retinopathy, the leading cause of blindness in working-age adults. The AI was not approaching specialist accuracy. It was matching it, and on some metrics lightly exceeding it.

Four years later, a Google Health team published results in Nature from an international evaluation of an AI breast cancer screening system. Tested against six experienced radiologists, the AI achieved a higher area under the receiver operating characteristic curve than any of the human readers. In UK screening data, the system reduced missed cancers by 2.7% and unnecessary callbacks by 1.2%. In US data, the margins were larger: missed cancers reduced by 9.4%, false positives by 5.7%. The AI was not just matching expert performance. It was catching cancers that experienced radiologists had not flagged.

In radiology more broadly, deep learning models now routinely exceed human accuracy on specific narrow tasks. One widely cited system achieved 94% accuracy in detecting lung nodules, compared to 65% for the radiologist cohort it was benchmarked against. A 2024 meta-analysis in npj Digital Medicine synthesised results across diagnostic specialties and found AI algorithms achieved overall sensitivity of 87.0% and specificity of 77.1%, versus 79.78% and 73.6% for clinicians. For ECG interpretation, a 2025 study published in Nature found that an AI model identified structural heart disease with 77.3% accuracy - compared to 64.0% for cardiologists working without AI assistance. When those same cardiologists were given the AI's assessment as a second opinion, their accuracy improved to 69.2%. The most capable diagnostic entity was neither the AI alone nor the physician alone. It was the physician informed by the AI.

The pharmacogenomics picture is different in character but no less significant. The SLCO1B1 finding does not involve image recognition at all. It involves reading a genomic file, matching specific variants against published drug-gene interaction data, and surfacing the relevant risk figures in plain language. The SEARCH Collaborative Group's 2008 NEJM paper established that patients who are CC homozygotes at rs4363657 face a 17.4-fold elevated risk of myopathy from 80mg simvastatin. That finding has been replicated and extended in the subsequent pharmacogenomics literature, and is now embedded in clinical prescribing guidelines in multiple jurisdictions. It is exactly the kind of information a GP might know exists in general terms but would rarely retrieve and apply to an individual patient unless a formal pharmacogenomics referral had been made - a step most health systems do not include as routine before initiating statin therapy.

Standard of Care

When Not Using AI Becomes Negligence

There is a legal argument forming at the edges of this evidence, and it is one that medical institutions have begun to take seriously. The argument goes like this: the standard of care in medicine is defined by what a reasonably competent practitioner would do under similar circumstances. As AI diagnostic tools become validated, widely available, and embedded in peer clinical workflows, that definition expands. The failure to use an accessible, validated, indicated tool begins to look structurally like the failure to order a test that was appropriate and available - an established basis for negligence.

The Milbank Quarterly published analysis in 2024 noting that AI-assisted diagnosis and predictive analytics can measurably improve patient outcomes, and that failure to consider them "may soon be viewed as falling short of the standard of care." Medical Economics framed the same argument from the malpractice attorney's perspective: if a physician's peers are routinely using a validated AI tool and a physician does not, their omission may constitute evidence of sub-standard practice. A 2025 legal analysis observes that "courts are beginning to consider whether a reasonable provider in today's tech-integrated environment should have used an AI system - and whether failing to do so could itself be a form of negligence."

No successful malpractice case on these specific grounds has yet been decided in common law jurisdictions. The legal framework is still forming, and the standard of care for AI use varies enormously by specialty, institution, country, and the specific tool in question. An FDA-cleared AI diagnostic system used routinely by a physician's peers carries different legal weight than a consumer chatbot. But the trajectory is clear, and legal scholars tracking this territory are not describing a distant hypothetical. They are describing an argument already being assembled in the literature and in law review journals across several jurisdictions simultaneously.

Medicine has been here before. When CT imaging became available, failing to order a scan where it was clinically indicated became a basis for malpractice. When evidence-based treatment protocols were established, departing from them without documented clinical rationale became problematic. AI is the next technology in this sequence. The question is not whether validated AI tools will become part of the standard of care. It is how quickly that transition happens, and which specialties reach it first.

The structural shift this creates is historically significant. Medical liability has always been primarily about acts - drugs prescribed, procedures performed, tests ordered. The emerging argument is about omissions: the cancer not caught because the AI second reader was not deployed, the drug interaction not flagged because the pharmacogenomics check was not run, the ECG error not identified because the image was not submitted for AI review.

The Inversion

The New Patient Equation

There is a quieter dynamic driving this shift that the legal arguments do not fully capture. In the cases described at the opening of this article, the patient did not receive AI-generated analyses from their doctor. They sought them out independently, processed the results before the appointment, and arrived prepared to have a more substantive clinical conversation than the appointment structure was designed to accommodate. The doctor adapted. The prescription changed. A harm was avoided. In this particular encounter, the system was the patient.

This is a structural inversion of how diagnostic knowledge has historically moved. The physician has always held the tools - the training, the equipment, the test results, the time to interpret them. The patient brought symptoms and deferred to clinical judgment. That asymmetry shaped every aspect of the consultation: its duration, its language, its power dynamics. The appointment existed to allow expertise to flow in one direction.

What consumer AI has done - partially, unevenly, and with significant caveats about accuracy and the limits of clinical context - is put diagnostic-adjacent capability into the patient's hands. Not diagnostic authority. The AI did not diagnose the patient with statin sensitivity. It identified genetic variants, cited a published risk coefficient from peer-reviewed literature, and recommended the patient take the finding to their doctor. The distinction matters enormously in terms of clinical and legal responsibility. But the practical effect was clinical: a better-informed decision, a drug interaction avoided, a harm that did not occur.

The time-pressured appointment - the GP with seven minutes and a waiting room of fifteen - is not a new observation. It is a chronic structural feature of healthcare systems under sustained demographic and financial pressure. What is new is that patients are beginning to arrive at those appointments having already run analyses the appointment itself could not generate in that time frame. A pharmacogenomics workup that would require a specialist referral and several weeks of waiting can now be approximated in the time it takes to upload a file and compose a question.

The 2024 npj Digital Medicine meta-analysis on AI versus clinicians across diagnostic specialties found that AI used as a decision support tool alongside a human - rather than as a replacement for one - produced better outcomes than either AI or clinician operating alone. The best diagnostic entity was the augmented one. The patient who arrived with a genomic analysis understood this intuitively, even without framing it in those terms. They were not trying to replace their doctor. They were trying to give their doctor better raw material to work with.

The question medicine now faces is whether it will build systems that make this augmentation routine - embedding pharmacogenomics screening into prescribing workflows, deploying validated AI second readers into cancer screening programs, making AI-assisted ECG interpretation the standard rather than the exception. Or whether it will continue to treat the informed, AI-assisted patient as an outlier, and the doctor who ignores validated AI findings as simply exercising clinical discretion. The evidence is accumulating on one side of that question. The legal framework is slowly aligning to match it. The patients, in the meantime, are not waiting for the institution to catch up.