The image above is what an AI anatomy error actually looks like: chambers mirrored, a stray vessel, everything plausible — and wrong. Across the reviewed examples in this scientific-illustration collection, the same failure patterns recur. Here are twelve of them, what they look like, and prompt changes that can help correct each one.
Regenerate a plate correctly · Full workflow · Label test results
One principle before the list: never fix an error by rerolling and hoping. Name the error in the prompt — "the chambers are mirrored; the right atrium is on the viewer's left" — and regenerate. Targeted corrections keep everything that already works; blind rerolls gamble it away.
Anatomy errors
1. Mirrored left/right structures
Looks like: a heart with right and left sides swapped; a brain slice that flips its lobes. Everything else can be perfect, which is what makes it dangerous.
Fix: name the orientation explicitly — "right atrium on the viewer's left, as in a patient facing the examiner". For teaching figures where handedness doesn't matter, the fastest fix is often removing left/right from the labels entirely.
2. Invented structures
Looks like: an extra vessel, a fifth valve, an organelle that appears nowhere in any textbook. Models reward plausible detail — which is exactly the failure.
Fix: "each organelle shown once" / "only the listed labels appear on the plate". Constraining what exists is as important as requesting it.
3. Wrong proportions
Looks like: a mitochondrion the size of the nucleus; a sunflower disk three times wider than any real capitulum ratio.
Fix: name the relationship — "the nucleus is the largest organelle", "with a size scale bar". If proportions must be exact, plan to correct them in a vector editor (workflow here).
4. Misplaced arrows and flow
Looks like: blood flowing backward through the heart; electrons drifting the wrong way along a transport chain.
Fix: arrows obey the sentence that names their direction — "flow arrows from atrium to ventricle", "electron path PSII to PSI, left to right". One direction per prompt sentence.
Label errors
5. Misspelled text
Looks like: "MITOCHONRIUM". Rare on short quoted labels, common on long ones and anything with symbols.
Fix: quote labels, keep them to 1–2 words, all caps, and spell out symbols (ALPHA, not α). Our 20-plate label test covers the reliability curve by label length.
6. Labels pointing at the wrong thing
Looks like: "GOLGI" with its leader line ending on a lysosome. Spelling perfect, anatomy wrong.
Fix: name each label's home — "the Golgi label points at the stacked membranes near the nucleus". If it persists, move the label into the panel name instead: "a small grid of captions around the cell".
7. Too many labels colliding
Looks like: leader lines crossing like circuit traces, labels overlapping the drawing.
Fix: the six-label cap, then split the topic across two plates. There is a full rule for this: how many labels should a scientific diagram have.
8. Confident nonsense vocabulary
Looks like: the model inventing its own term — labeling a mitochondrion "POWERHOUSE" or a vessel by a word from a different organ.
Fix: the label list comes from your reference, quoted verbatim. The model's vocabulary is a suggestion; yours is the spec.
Composition errors
9. Decorative overload
Looks like: lens flares on a kidney, a gradient sunset behind a cell. Style colonized the figure.
Fix: pin the style slot — "clean textbook diagram, white background, flat colors". The word "textbook" is doing real work in that sentence.
10. Crowded layout
Looks like: eight panels squeezed until none is readable.
Fix: name a small panel set — "one large central cutaway, one magnified inset, nothing else". "Nothing else" is a load-bearing phrase.
11. Microtext
Looks like: labels rendered at 8-point equivalent, unreadable at use size.
Fix: "labels large enough to read at half size", or shorten the label. This is also why plates should be generated in 16:9 or 1:1 at high resolution — see raster vs vector for the size strategy.
12. Style drift across a series
Looks like: plate 1 in watercolor, plate 2 in flat vector, plate 3 in something between.
Fix: reuse the identical style clause in every prompt of a series, and regenerate the outlier rather than "matching later". Consistency comes from the prompt, not from luck. See one subject, five styles for how much the style slot changes.
The general pattern
Notice that every fix above is a constraint added to the prompt, not a plea. The mental model that works: you are not asking the model to draw; you are specifying a figure and rejecting nonconforming output. Errors aren't random noise — each one has a grammatical fix.
Scientific review checklist (the last line of defense after all fixes):
- [ ] Are all labels spelled correctly?
- [ ] Are all requested structures present — and nothing mirrored or invented?
- [ ] Are positions and proportions reasonable against your reference?
- [ ] Do arrows show the correct direction?
- [ ] Are stages shown in the correct order?
- [ ] Has it been checked against a trusted source?
Fix it in one pass
Paste any prompt from our examples gallery into the scientific illustration presets, apply the six-label rule, and review with the checklist. When a plate needs more than three targeted corrections, rewrite the prompt from scratch instead of patching — at that point the spec, not the rendering, is the problem.
FAQ
Are these errors random or systematic?
Systematic — which is the good news. Each error has a grammatical cause (an unnamed constraint), so each has a prompt fix. Random noise would make drafting a gamble; systematic bias makes it an engineering problem you can actually solve.
How many retries should I budget per plate?
One to two targeted corrections for a well-specified prompt; more if the spec is vague. When a plate needs a third fix, stop patching and rewrite the prompt from the four-slot anatomy — that's the refinement guide's patch-or-rewrite rule, and it holds.
Do negative instructions work ("no gradients")?
Weakly. Models honor positive constraints far better than prohibitions. Instead of "no decorative elements", specify what fills the space: "clean textbook diagram, white background, flat colors." Name what you want; absence is implied.
Which errors are unrecoverable?
Invented data — a fabricated structure presented as real anatomy — is unrecoverable in a scientific context: delete and regenerate. Mirrored layout, wrong proportions, and label errors are all recoverable by named correction. The category matters: presentation errors fix; epistemic errors delete.
Do the errors differ between disciplines?
The twelve are universal, but their frequency shifts: anatomy skews toward mirroring and invented structures; botany toward proportion drift; mechanism figures toward arrow direction; multi-panel layouts toward crowding and microtext. Run the checklist in your discipline's failure order — it front-loads the checks most likely to fire.
When a flawed draft is still useful
A plate with a circled error is not waste — it's teaching material. The heart-QA comparison at the top of this page is a product of error analysis: two versions, the flaws annotated, the corrections noted. If you maintain teaching materials, keep your failures; "spot the error" exercises built from real generation failures teach verification instinct better than any correct figure can.
Error triage in 60 seconds
When a plate comes out wrong, the decision tree: Is anything invented or fabricated? Yes → delete and regenerate; presentation errors can be fixed, epistemic ones can't. Is the layout right but details wrong? → patch with a named constraint; this is the majority case. Is the layout itself wrong? → rewrite the panel list from the four-slot anatomy (refinement guide); patching composition is thrashing. Are three or more fixes queuing? → rewrite, don't stack patches. Is it nearly right twice in a row? → narrow the scope instead of polishing; a smaller, correct figure beats a comprehensive, flawed one — the same verdict the label-count rule reaches from a different direction.
Errors as a quality signal
A team that can name its error patterns generates better figures than one that just "checks them" — pattern knowledge moves fixes upstream into prompts, where they're cheap. Keep a running tally for your own work: which of the twelve fire most, on which subjects, after how many labels. Two months of tallying and your first drafts stop containing your personal top three errors at all, because the constraints that prevent them have migrated into your default template. That's the compounding loop this site's own prompt packs encode — every verify note in them is somebody's tally, written down so you don't have to rediscover it.
<!-- scientific-illustration-cluster-links:start -->
Continue through the scientific illustration collection
Create a scientific illustration · Browse the scientific illustration resource hub
Related: Can AI Render Scientific Labels Correctly? A 20-Plate Test · GLMImage vs BioRender for Scientific Figures: Which Fits Your Work?
<!-- scientific-illustration-cluster-links:end -->
