How to Verify an AI-Generated Scientific Illustration (Checklist)

Split figure showing a researcher comparing an AI-generated cell diagram against a reference with a checklist card

Verification is the step that turns an AI image into a scientific illustration draft that is ready for expert review. This page provides a six-point checklist, a reference-anchored pass in the right order, and disclosure guidance for reviewed plates.

Generate a draft to verify · Common errors and fixes · Label test

The six-point checklist

For every plate:

[ ] Are all labels spelled correctly? Read each one character by character — your brain autocorrects; the figure doesn't.
[ ] Are all requested structures present? Against your prompt's label list, item by item.
[ ] Are positions and proportions reasonable? Against the reference, not against memory.
[ ] Do arrows show the correct direction? Trace each arrow's tail and head deliberately.
[ ] Are stages shown in the correct order? Especially life cycles and pathways — see the life-cycle guide.
[ ] Has it been checked against a trusted source? A textbook, atlas, journal figure, or official reference page — named, dated, and kept.

The order matters less than the discipline of doing all six. What does not work: glancing at the plate and confirming it "looks right". AI errors are specifically the kind that look right.

The verification pass, in order

1. Labels first, science second

Spell-check labels before checking anatomy. It's the cheapest check and it changes how you read everything after: if "LEFT VENTRICLE" came out "VENTRCLE", you know the plate needs a regeneration before it's worth five minutes of anatomical review.

2. Structure against the reference, side by side

Open the reference next to the plate — same screen if possible. Work through the label list top to bottom, checking each term against the reference's depiction. This is where mirrored anatomy (error #1) gets caught: the right atrium on the wrong side reads as fine until you compare.

QA comparison of two heart cutaways — one correct, one flawed with circled errors

For anatomy specifically, use a real atlas — our medical illustration guide lists the standard references per region.

3. Arrows and sequences traced

Physically trace flow arrows with your cursor or a pencil tip. Check stage order in cycles and processes against the reference's own ordering. Sequence errors survive casual review because a wrong-order cycle still looks like a cycle.

4. The "what's missing" pass

Check for what the model should have drawn but didn't — the absent organelle, the missing legend, the scale bar the prompt requested. Omissions are the least visible error class; the checklist's second item exists for them.

5. The "what's extra" pass

Look for structures that shouldn't be there — the invented vessel, the decorative sixth panel. Anything present that your reference doesn't show and your subject doesn't have must go. Fix by naming it in the prompt, not by hoping the next roll omits it.

What "verified" lets you claim

Verification has levels, and mixing them up is how people get in trouble:

Level	What you did	What you may claim
Draft	Generated, glanced at	"AI-generated draft" — nothing more
Reviewed	Six-point checklist against a reference	"Human-reviewed for labels and structure"
Corrected	Errors found and fixed (regeneration or edit), re-checked	"AI-drafted, human-corrected"
Certified	Expert checked the science itself (yourself, if you're the domain expert)	Ready for teaching or publication, with disclosure

For classroom slides, Reviewed is usually enough. For published curricula, exams, papers, or anything with your institution's name on it, Corrected or Certified plus disclosure is the honest bar. Our research-paper workflow covers what journals currently expect to see disclosed.

Why disclosure helps rather than hurts

Saying "AI-drafted, human-verified" does two things: it keeps you on the right side of the growing disclosure norms in education and publishing, and it raises the perceived reliability of the figure — someone checked it. What damages credibility is the midpoint: publishing an unreviewed AI figure with no note, and having the mirrored heart found by someone else's class.

Verification is cheaper than it looks

The whole protocol is five minutes for a six-label plate — less time than redrawing any single panel by hand, which is the entire economics of AI drafting done honestly. The workflow guide puts this step fourth for a reason: everything before it is cheap, and everything after it is trust.

Scientific review checklist

(Yes, again. This is the page you send people to.)

[ ] Are all labels spelled correctly?
[ ] Are all requested structures present?
[ ] 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?

FAQ

Who counts as a verifier?

Anyone who can check the plate against the reference and would notice the difference — which for most classroom and blog figures is you, with the reference open. For publication claims, "verifier" means someone who could defend the figure to a reviewer: you if you're the domain expert, or a colleague who is.

How long does verification take?

Five minutes for a six-label plate once you have the reference beside you; two minutes for line-art plates with short labels. Label spell-check is the fastest pass; the "what's missing" and "what's extra" passes take the longest because they fight your brain's tendency to see what it expects.

What if I'm not a domain expert?

Then the figure's claims can't outrun your source. Use references that carry their own authority (textbook, atlas, field guide, standards body), check each label against the reference's own depiction, and disclose the level honestly: "generated, checked against [source]" rather than "verified." The disclosure table above exists precisely so you don't overclaim.

Do I need to keep records?

Keep three things with each published figure: the prompt, the reference link, and the date of your check. It's two minutes of filing that turns every future question — from a student, editor, or your own memory — into a lookup instead of an archaeology project.

Does verification get faster with practice?

Sharply — the checklist internalizes in about a week of regular use, and you start seeing categories (mirroring, invented, direction) instead of checking items. What never speeds up past its floor: the reference pass. Comparing against a source is irreducible reading time, and that floor is exactly why verification stays meaningful — it's the part that can't be waved through.

The verification artifact

The habit that scales: every verified plate gets a one-line record — prompt used · reference checked · date · verifier initials. Teams that adopt this stop relitigating figures ("was this one checked?") and start building a trusted library. Ours lives next to the asset index that powers this site's galleries; yours can be a spreadsheet. The medium doesn't matter; the trail does.

The five-minute pass, minute by minute

Minute one: label spell-check, character by character, cheapest first. Minute two: structures present — the prompt's label list against the plate, item by item. Minute three: placement against the open reference, left-right and orientation hardest. Minute four: arrows and stage order traced physically. Minute five: the missing-and-extra sweep, then the record — prompt, reference, date (the artifact). Five minutes is also the honest ceiling for a six-label plate; if verification is taking twenty, the figure has too many labels (the rule) or too little specification — and the fix is upstream, at the prompt, not downstream at the checklist.

Teaching verification to others

If you hand figures to students, TAs, or junior team members, teach the protocol explicitly rather than hoping it's obvious: give them the six-point checklist, one reference, and three figures — one clean, one with a planted spelling error, one with a planted anatomical error — and watch them find two of three. That miss rate is the lesson: verification is a skill with a known miss rate, improvable by method (character-by-character reading, reference open, arrows traced), not by care alone. Teams that train it stop shipping the plausible-wrong figure; teams that don't, eventually ship one publicly. Our label test doubles as ready-made practice material.

Draft on the scientific illustration presets, run this protocol, and keep the reference link with the plate. Then read: 12 errors and their fixes · label reliability test · journal figure specs.