Medical Illustration in the Age of AI: A Practitioner's View
There is an image circulating in a clinical training module that looks, at first glance, entirely convincing. The anatomy is plausible, the shading is confident, and the proportions are broadly correct. It is also, to anyone who knows what they are looking at, clinically useless. That tension sits at the heart of what artificial intelligence is doing to medical illustration right now.
As an NHS microbiologist who works alongside visual communication in diagnostic and educational contexts, I want to offer a grounded view of what is actually changing and what is not. The conversation around AI-generated medical imagery tends toward two unhelpful extremes: either panic about replacement or uncritical enthusiasm. Neither captures the real shift, which is subtler and more interesting.
This post works through the evidence carefully. From the specific failure modes of AI-generated anatomy, to the question of what separates a surgical diagram from a patient-facing guide, to where quality medical art prints and professional illustration still carry irreplaceable weight, the argument builds toward a conclusion that AI has made the illustrator's core skill more visible, not redundant. That skill was never draughtsmanship. It was judgment.
The Image That Looked Right But Wasn't
Picture a trainee surgeon reviewing an AI-generated illustration of Calot's triangle before a laparoscopic cholecystectomy. The proportions are correct. The cystic duct, hepatic artery, and common bile duct are all present. What it lacks is any sense of operative context: the tissue planes that distort under traction, the anatomical variation that makes this landmark genuinely hazardous, the visual hierarchy that tells a trainee where to look first. The image looked right. For its purpose, it wasn't.
This is the tension AI has made impossible to sidestep: anatomical fluency is not the same as clinical utility. A generated image can be trained on accurate sources and still communicate nothing useful to the person who needs it most.
Evaluating imagery inside an NHS workflow is a different exercise from evaluating it in a design studio. The question is never simply whether the anatomy is defensible; it is whether the image does its job for a specific person at a specific point in their clinical development. That judgement requires proximity to clinical practice, not just access to a prompt field.
The argument running through this piece is straightforward: the genuine value of medical illustration has always been editorial judgement, the decision about what to show, what to omit, and who the image is speaking to. AI accelerates production and has made that distinction more visible, not more optional.
What AI Is Actually Doing to Medical Imagery
The mechanics are not subtle. Generative tools have compressed what once took a specialist illustrator several days, reference gathering, concept iteration, schematic drafts, into minutes. That compression is real, and it is not going away.
What has changed alongside it is access. Clinicians, educators, and publishers now generate medically themed imagery without commissioning a specialist, and the volume of that imagery in circulation has increased sharply as a result. The barrier to producing an image that looks medical has essentially disappeared.
That is where the problem surfaces. The acceleration does not create the gap; it exposes one that was already there. Speed and anatomical correctness were never the hard part of medical illustration. They were the visible part, the part that looked like the job. The actual work, deciding what an image needs to communicate, to whom, and in what clinical context, was always the harder and less visible function.
NHS settings feel this particularly acutely. Resource constraints already pushed teams toward stock imagery and pre-made schematics long before generative AI arrived; the threshold for producing imagery has now dropped further, while the accuracy stakes have risen. The implications for NHS medicine right now are not theoretical.
The "will AI replace illustrators" framing is, frankly, the wrong question. It focuses on execution. The more useful question is whether anyone designing these AI-assisted workflows has paused to ask what the illustrator was actually doing before the tool arrived. Most have not.
Anatomically Correct, Clinically Meaningless
That gap between what AI produces and what clinical work needs has a name: the difference between anatomical correctness and clinical meaning.
Consider a rendered illustration of a renal artery. The calibre is plausible, the branching pattern recognised, the surrounding structures present. A vascular radiologist without context will likely find nothing technically wrong. But hand it to an interventional nephrologist planning a percutaneous procedure and the image is useless, because it shows the structure without showing what matters: the ostial angulation, the relationship to the aortic wall, the collateral anatomy relevant to that specific clinical decision. Correct and useless are not mutually exclusive.
AI-generated images pass superficial anatomical review because they are trained on anatomically accurate sources. That is precisely the problem. Training on correct data teaches an image to look right; it does not teach it what job it has to do.
Clinical meaning is relational. A pathology slide annotation prepared for a final-year student is a fundamentally different object from the same slide presented in an MDT. One is building a mental model; the other is supporting a decision already in progress. The relevant details, the framing, the level of certainty signalled visually all shift depending on who is looking and why.
Illustrators embedded in NHS environments learn this through proximity: sitting with the renal team, attending the radiology meeting, understanding the workflow before picking up a tool. That contextual literacy has to actually mean something in the finished image; it cannot be recovered at the prompt stage.
The real risk is not the hallucinated extra vertebra. It is the image that is entirely defensible on inspection, quiet in its wrongness, and confidently used anyway.
The Surgeon, the Student, and the Patient: One Structure, Three Different Images
The biliary tree makes the case precisely because it looks like a single subject but isn't.
Normal biliary anatomy is present in only around 58–60% of the population. That single statistic already tells you why three different images are needed.
For a surgeon, that variation is the entire point. The image must foreground right posterior hepatic duct anomalies, flag Rouvière's sulcus as an intraoperative landmark, and suppress everything that does not bear on dissection of Calot's triangle. It is a decision-support tool built around operative anatomy, not descriptive anatomy. Structures that orient the eye in a textbook actively clutter a surgical schematic.
For a medical student, the priority inverts. Classical confluence comes first; relationships between the common hepatic duct, cystic duct, and common bile duct need to be spatially clear before variants are introduced. Variation taught too early collapses the mental model before it has formed. The image's job is sequential construction, not completeness.
For a patient awaiting cholecystectomy, clinical specificity is counterproductive. The image communicates process and reassurance. It shows that a thing happens, not how the surgeon navigates it. Any labelled anatomical variant at that stage is more likely to generate anxiety than understanding.
Now enter a text prompt: "diagram of the biliary tree." The output is one image. It will be anatomically defensible. It will serve none of these three purposes well, because serving any of them requires knowing which version you are making before you begin.
That is not a rendering problem. It is an editorial one. The illustrator's core competency is the decision that precedes the image, not the execution that follows it.
Editorial Judgment: The Skill That Was Always the Point
That distinction between knowing which image to make and knowing how to render it points directly to the skill that has always defined the discipline.
Editorial judgment means deciding what a medical image must show, what it must omit, how it frames clinical information, and who it is addressing. It is not a finishing touch applied after the drawing is done. It is the decision that makes the drawing possible.
This predates AI entirely. When Gray and Carter produced Gray's Anatomy in 1858, they were making arguments: which structures merited emphasis, which relationships warranted clarification, which details would impede understanding. The plates are editorial positions rendered in ink. Go further back, to Vesalius commissioning artists trained in the Titian tradition for De Humani Corporis Fabrica, and the same logic holds: aesthetic and structural choices encoded a specific view of what mattered. The illustrations were never neutral.
AI has changed one thing meaningfully: it has democratised draughtsmanship. A technically competent rendering is no longer a barrier. What it has not democratised is clinical knowledge, or the judgment that accumulates from working inside healthcare environments, attending MDTs, watching trainees misread a schematic, or understanding what a procedural guide actually needs to do under pressure.
In NHS practice, the stakes are not abstract. A poorly framed illustration in a procedural guide does not just confuse; it can embed a misunderstanding at exactly the point where accuracy matters most. That places medical illustration within clinical governance, not outside it.
Practitioners who will use AI well are those who arrive at the tool with the editorial decision already made, using it to accelerate execution rather than to generate judgment they have not yet formed.
How AI Fits Into a Sensible Workflow
So, accepting that editorial judgment is the irreplaceable element, what does a functional AI-assisted workflow actually look like in practice?
AI earns its place at the drafting stage. It generates initial concept iterations quickly, produces orientation-level reference imagery, and handles schematic visuals where clinical nuance is deliberately low. A structural draft that once took several hours can be in front of a reviewer within minutes. That compression is real and worth using.
The boundaries are equally real. AI is poorly suited to any image where clinical context determines what is salient, where audience-specific framing is the entire point, or where regulatory-grade accuracy is required. For patient-facing materials or anything feeding into clinical decision support, AI output alone is not an acceptable endpoint.
A practical NHS workflow reflects this division: AI produces a first draft of an educational schematic, a clinician or trained illustrator reviews it editorially, and the final version reflects that human judgment. The AI saves time; the reviewer delivers fitness for purpose.
Governance is the outstanding problem. When an AI-assisted procedure guide contains a subtle anatomical error, responsibility is not clearly allocated under most current NHS frameworks. The MHRA's ongoing work on AI regulation in healthcare signals that clarity is coming, but it is not here yet. In the interim, institutions should explicitly assign sign-off responsibility for clinical accuracy in any AI-assisted illustration before it reaches trainees or patients.
The blunt summary: AI makes the routine parts of illustration faster and cheaper. It makes the consequential part, clinical editorial judgment, more visible and more critical than it has ever been.
Why Medical Art Still Carries Weight
That clinical editorial judgment does not stay in the hospital. It shows up elsewhere, and that is worth noticing.
Medical art prints, anatomical diagrams, histology illustrations, procedural schematics rendered with genuine care, continue to resonate with doctors and medical students well beyond any practical utility. They appear in offices, on tote bags, on mugs. The reason is not nostalgia.
The best medical art prints encode a specific kind of attention. They show that someone understood what mattered in a structure or a process, decided what to show, and made that argument in visual form. That is editorial judgment, not decoration. A well-made anatomical illustration reads differently from a generic graphic because it carries the trace of clinical knowledge translated into image; the competence is visible in the choices.
UK audiences working in medicine are sensitive to this distinction, often without articulating it. There is real appetite for imagery that reflects the actual texture of clinical thinking rather than sanitised stock iconography. Images that feel made by people who know the material communicate something about shared experience and cultivated identity that generic medical graphics cannot.
For a brand like Clerked, whose designs originate inside NHS practice, this is structural rather than incidental. The editorial judgment is present in the work from the outset, which is precisely why we made these gifts the way we did. It is also why apparel makes a better gift than you might expect: a design built from clinical knowledge carries meaning a generic alternative cannot replicate, regardless of how polished the rendering.
What This Means for Anyone Commissioning or Creating Medical Imagery
That principle extends directly into practice. The right question to ask of any medical image, whether AI-generated or not, is not "is this anatomically correct" but "does this image know what it is trying to do, and who is it talking to." Anatomical correctness is a minimum standard, not a quality marker.
For clinicians and educators commissioning imagery, use AI tools freely for speed and iteration. Generate the first draft, compress the early cycle, explore options quickly. But apply clinical editorial judgment at the review stage, as a deliberate step, not a final formality. The review is where the actual work happens.
For NHS teams, the governance question deserves explicit attention. When an AI-assisted illustration contains a subtle error in a procedural guide or training resource, responsibility needs to sit somewhere named and accountable. Build that into the sign-off process now, before an incident makes it urgent. Establish clearly who in the workflow holds clinical accuracy responsibility for image outputs, and document it.
For medical students and junior doctors encountering AI-generated educational imagery, develop a working habit: interrogate the image. What has it chosen to show? What has it left out? Critically, does that omission serve your learning or obscure something you actually need to understand? Absence of detail is always a decision, and it is not always the right one.
The field is moving forward, not backwards. AI is a legitimate tool used well by practitioners who understand that it handles execution while judgment remains theirs.
Well-designed medical art prints, apparel, and products built from genuine clinical knowledge demonstrate exactly this: editorial judgment operating outside the hospital. The same rigour that makes an illustration clinically trustworthy makes a design worth owning.
The Argument AI Has Not Changed
All of this was true before the first generative model was trained, and it remains true now.
AI has accelerated the production of medical imagery considerably. It has not changed what makes a medical image good. That has always been editorial judgment: knowing which image to make, for whom, and what to leave out. Clinical context and audience awareness are not features that can be added at the prompt stage. They are the work.
The practical position is straightforward. Use AI tools freely for what they do well: rapid iteration, structural drafts, schematic-level visuals where nuance is intentionally low. Do not mistake anatomical fluency for clinical utility. An image that passes a superficial accuracy check can still be quietly wrong for its purpose, and in healthcare that distinction carries weight.
The best medical illustration has always been an argument about what matters. AI produces more images. It does not make that argument for you.
That principle extends beyond clinical settings. Clerked's medical-themed designs, developed from inside NHS practice, carry the same logic: the editorial judgment is present in the work from the start, not applied afterwards. That is what makes any medical image, print, or design worth taking seriously.
Conclusion
Medical illustration has never been about producing images. It has always been about making arguments: which structure matters, for which audience, at which level of detail.
AI generates imagery faster than any human illustrator. It does not replace the clinical judgment that determines whether an image is actually useful. Anatomical fluency and clinical utility are different things, and confusing them carries real consequences.
The principles that make a medical image trustworthy, selectivity, audience awareness, and editorial purpose, apply equally to clinical documentation, education, and design.
If you commission or create medical imagery, start with the question AI cannot answer: what does this image need to do, and for whom? Get that right, and the tools you use to execute it matter far less.
The argument was never about the technology. It was always about the judgment behind it.