Medical accuracy in AI video: anatomy, devices and the frame that fails review

You make AI-generated medical video accurate by treating accuracy as a production stage with an owner, a reference pack and a sign-off, and by checking the finished cut frame by frame against that pack. Generators produce anatomy that looks right and is wrong, devices that resemble your product without matching it, and procedures in an order no surgeon would follow. A named medical reviewer, working from locked references before generation and on the final cut after it, is the control that catches them.

The uncomfortable part for a CMO is how good the wrong frames look. A brand team signs off on what it can judge: light, pace, performance, logo. A misplaced tendon or a catheter entering from the wrong side passes that review routinely, because nobody in the room was checking for it. The frame that fails is the one a cardiologist, a KOL or a regulator's reviewer notices, and by then it is live.

This piece covers visual accuracy only. Advertising law for healthcare brands sits in "What Healthcare Brands Need From Creative in a Regulated Market", and shot grammar in "AI Filmmaking: What Separates Cinematic From Generic".

Where generated anatomy goes wrong

The published evidence is consistent, and it is worse than most marketing teams assume.

A Cureus study published on 21 November 2024 generated 240 photorealistic images of people across three text-to-image models and scored every anatomical error by type and severity. Severe errors made up a large share of the total for all three models. Configuration errors, meaning structures in the wrong arrangement, were the most common type. Hands failed most often, followed by faces and limbs, and scenes with several people scored far worse than single subjects (Source: Muhr et al., Cureus, 21 November 2024).

When the task moves from ordinary bodies to clinical anatomy, the numbers get harder to read. A study in the Journal of Hand Surgery Global Online, October 2025, asked six generators for labelled, anatomically accurate images of common hand procedures, 1,500 images in total. It found that 99.8 percent contained at least some fabricated anatomy: structures that do not exist, drawn with total confidence (Source: Journal of Hand Surgery Global Online, October 2025).

And the errors are already reaching print. A review in Plastic and Reconstructive Surgery Global Open, April 2026, screened 4,734 manuscripts in aesthetic medicine and found AI-generated facial anatomy in 4 of 37 relevant articles published in 2025. Every one of those images carried gross anatomical inaccuracies. The same review found erroneous AI imagery on nearly half of the online anatomy course sites it checked (Source: Plastic and Reconstructive Surgery Global Open, April 2026). If peer review in aesthetic medicine let these through, a brand approval chain built for taste will too.

Two caveats matter. All three studies tested still images, and newer models may score better. Video still multiplies the problem, because a structure that is right in frame 40 can drift by frame 90. The generator works from a statistical impression of what bodies tend to look like, and it redraws that impression every frame.

The pattern a reviewer learns to hunt for is predictable:

1. Wrong count or wrong arrangement: extra fingers, merged vessels, a rib cage with the wrong curve.
2. Invented structures: a ligament that does not exist, a label pointing at nothing.
3. Layer confusion: muscle tucked beneath the bone it should cover, organs sharing space.
4. Side and orientation: a heart with its apex on the right, a scar on the wrong knee.
5. Temporal drift: anatomy that is correct at the start of a shot and quietly changes by the end.

The device render that does not match the box

For a medical device brand, the anatomy problem has a sibling that costs more. Ask a generator for a stent, an insulin pen, a glucose monitor or a knee implant and it returns the average of every such device it has seen. The result looks plausible and matches no product on the market, including yours.

The failures are specific. Port positions move. Button count changes between shots. The colour of a cap, which on some devices signals a dose or a size, comes out wrong. An implant gains a thread pattern it does not have. On a regulated product, a render that shows a feature the device lacks may amount to a product claim nobody approved, made in a frame nobody checked. Whether it does is a question to take to your regulatory counsel before the film ships.

The fix is to keep the product real. Build a locked reference pack from the device maker's own CAD files, product photography and instructions for use. Generate the environment, the light and the hands, and composite the real object into the frame wherever it has to be exact. Where a generated device appears at all, the product manager signs it off against the pack, shot by shot. This matters most in KOL-led campaigns, where a known doctor is seen holding or discussing the device. A doctor appearing in that creative needs explicit written consent and a compliance check before production starts, which "Faces You Do Not Own" covers in full.

Procedures shown out of sequence

The third failure is narrative. A generated procedure film can get every individual frame right and still show the steps in an order no clinician would follow: the incision after the instrument is inside, the dressing before closure, a scan result displayed before the scan. A model generates shots with no knowledge of clinical protocol.

This is a script problem before it is a generation problem. The procedure sequence goes into the shot list as numbered clinical steps, checked by a clinician before any shot is generated, and the edit follows that list. If the story needs compression, the reviewer decides which steps can be implied and which must be shown, because a skipped sterile step reads very differently to a surgeon than to a brand manager.

The medical reviewer's pass, as a named stage

Every healthcare brand already has a medical or legal review somewhere near the end. That placement is the problem. A reviewer who sees only the finished cut can reject it but cannot prevent the errors, so the brand pays for generation twice. The review has to move forward and happen at three stages.

Before generation: the reference lock. The reviewer approves the script, the numbered procedure steps and the reference pack: anatomical references from a named atlas or the client's own clinical imagery, device CAD and photography, and the list of what must never appear. Every generation prompt draws from that pack.

During generation: the shot check. Any shot carrying anatomy, a device or a clinical step gets reviewed as a still sequence before it goes into the edit. It is cheaper to regenerate one shot than to re-cut a film.

After the edit: the frame review. The reviewer watches the final cut at reduced speed, frame-stepping through every clinical moment, and signs a log that records the version, the shots checked, the issues raised and what changed. That log is what you show when a KOL, a hospital's medical director or a regulator asks how the film was checked.

Who the reviewer is depends on the content. For patient education and hospital brand film, a practising clinician in the relevant specialty. For device content, a clinician plus the device maker's product or regulatory lead. The reviewer's name and qualification go on the log. A brand manager with a medical dictionary does not qualify, however careful.

Reviewing at each stage adds calendar time and cost. The alternative is the reshoot after a KOL refuses to share the film, or the takedown after a specialist posts a screenshot of your logo beside the wrong anatomy.

What to ask whoever makes your video

The anatomy belongs to the reviewer. The CMO's job is to know the check exists and who owns it. Four questions settle it:

1. Who is the medical reviewer on this film, what is their qualification, and at which stages do they review?
2. What reference pack are the generations built from, and who approved it?
3. How is the real device kept accurate: composited from real assets, or generated?
4. Can we see the signed review log for the final cut?

A team that answers all four with names and documents has built accuracy into the pipeline. A team that answers with "our AI is very good at medical content" has not.

Nextdot's Creative Intelligence Pod produces healthcare creative for hospital and healthcare brands, and every healthcare video it makes gets a medical accuracy review before it ships. These four questions are the ones we expect a CMO to put to us. Any reference pack built for a client's films belongs to the client, along with the rest of its creative memory, and leaves with the client in usable form if the engagement ends.

Frequently asked questions