A trial schema like this — five labeled stages, response branches, publication-clean — is exactly the figure class where AI drafting belongs in research: schematic, vocabulary-driven, and cheaper to regenerate than to commission. This guide covers where AI illustrations genuinely fit in papers, where they don't, the verification bar, and how to disclose them so editors and reviewers stay on your side.
Try a draft · Verification protocol · Graphical abstracts with AI
Where AI illustrations fit in a paper
Good fits — schematic, semantic, regenerable:
- Trial and workflow schematics — like the figure above; boxes, arrows, stages, branches. The content is what happens in what order, which is prompt-stable.
- Mechanism cartoons — "antibody bridges T cell to tumor": the message is topological, not metrical.
- Graphical abstracts — the one-figure summary; we have a full guide.
- Review-article overview figures — synthesizing a field's consensus model.
Poor fits — keep them out of the methods section:
- Data figures. Blots, gels, microscopy, plots, anything evidentiary. These must be real — an AI-generated "result" is fabrication, not drafting, and is treated as such by every serious journal.
- Exact quantitative diagrams. Circuit boards, statistical plots, anything where geometry encodes numbers.
- Anything you'd have to caption "representative" without it being representative. If the figure claims to show the data, it can't be generated.
The dividing line is simple: AI drafts explanatory figures; it never produces evidentiary ones.
The workflow, research-grade
- Start from your protocol or methods text. The stages are the label list — pull the exact terms (apheresis, conditioning) straight from the document.
- Draft with the template — the workflow guide has the grammar; discipline prompts are in the examples gallery.
- Regenerate to spec, not to taste — figure dimensions, color space and font policy come from the journal (see below).
- Verify at Certified level — the six-point checklist plus a domain read: terminology current, pathway direction right, no structure your field would question. For a mechanism cartoon, you are the domain expert — act like it.
- Disclose. Below.
Journal specs are the real constraint
Journals constrain figures harder than any aesthetic consideration: column widths, dpi minimums, file formats, color-model policy, and graphical-abstract rules all vary by venue — and they change. The practical workflow is to check the current author guidelines for your target journal before final export, not after. What AI drafting changes for the better here: when a journal rejects your figure's aspect or resolution, regeneration to the new spec costs minutes, not a round-trip with an illustrator.
Technical format decisions (raster vs vector, dpi by use) are covered in raster vs vector scientific figures — the short version: schematics for print want vector paths in the end, which is why our AI-to-vector workflow matters for this use case.
Disclosure: what journals expect
Policies differ in wording but converge on the same three points:
- Describe the tool use in Methods or Acknowledgments — which tool, for what (figure drafting), and the extent (draft only vs. draft-and-edit).
- State human responsibility for accuracy — the authors verified all scientific content; the AI did not "conduct research".
- Never present generated content as data. This is the bright line: generative drafting of explanatory figures = disclosable; generation of anything evidentiary = fabrication.
A Methods sentence that works:
Schematic figures (Figs. 1 and 4) were drafted with an AI image-generation tool (glmimage.app) and subsequently edited and verified by the authors for scientific accuracy; all data figures derive from original experimental results.
Also check your target journal's AI policy and your institution's — some funders add their own disclosure layer. When in doubt, disclose more: the sentence costs nothing, and undiscovered fabrication costs everything.
The verification bar is higher here
Classroom-standard review (labels spelled, structures present) is the floor. For publication:
- Terminology current — field-standard terms, no textbook-legacy names your reviewers would flag.
- Topology right — arrows, activations, inhibitions match the mechanism you're asserting, because reviewers will read the arrows as claims.
- No decorative excess — the style clause should read
clean, publication-grade, white background, restrained color palette. - A colleague pass — someone in the field looks at the figure cold and reconstructs your sentence from it. If they can't, the figure isn't done. (This is the same "one message" test as graphical abstracts.)
Scientific review checklist
- [ ] Are all labels spelled correctly?
- [ ] Are all requested structures present?
- [ ] Are positions and proportions reasonable against the literature?
- [ ] Do arrows match the mechanism as claimed in the text?
- [ ] Are stages shown in the correct order?
- [ ] Has it been checked against a trusted source — and by a domain reader?
- [ ] Is the AI tool use disclosed at the right place?
FAQ
Do journals ban AI-generated figures?
Some ban generative AI in specific roles — most prominently in peer review (uploading a colleague's manuscript to a third-party tool), and a few restrict generated images generally. For your own explanatory figures, the dominant pattern is "allowed with disclosure", not banned — but this is exactly why you check the current policy of your target venue rather than trusting a summary, including this one.
What do reviewers actually object to?
In practice: undisclosed AI use, overclaimed precision, and generated content that drifts from the text's claims. A schematic that honestly matches the Methods, with a one-line disclosure, draws no objection anywhere we've seen. The objection is never "a machine drew the arrow" — it's "the arrow wasn't checked."
Who owns the copyright in a generated figure?
The honest answer is jurisdiction-dependent and still settling; venues differ and some assign rights questions to their standard author agreement. What is settled: you are responsible for the figure's accuracy and its compliance with the journal's terms regardless. For commercial reuse beyond the paper, check the tool's terms — ours are stated plainly — and when in doubt, ask the venue.
Preprint or journal first — does disclosure differ?
Disclose in both. Preprints are publications; the same Methods sentence travels to the journal version, and consistency between the two is free credibility. The failure mode is a preprint with generated figures and no note, indexed forever, discovered during peer review.
Co-author consent — do I need it for AI-drafted figures?
Practically, yes: disclose to co-authors before submission, not after acceptance. The Methods sentence covers the journal; a one-line note in the collaboration channel covers the humans whose names share the figure's credibility. Most friction in this area comes from surprise, not from the drafting itself.
Working with a professional illustrator
If your venue or budget includes a medical/scientific illustrator, AI drafts change that collaboration rather than replacing it: bring the drafts as the brief. Illustrators report that a generated draft communicates intent faster than any written description — the draft fixes composition and vocabulary, the illustrator fixes craft and correctness, and the result beats either path alone. The hybrid-lane idea from our BioRender comparison is the same principle with icon libraries.
A submission-week checklist
The morning you submit: (1) every generated figure disclosed in Methods with tool and extent named; (2) the venue's current figure policy re-checked against today's author guidelines, not last month's memory; (3) every schematic run through the six-point checklist with the reference link saved — reviewers ask, and "here is what it was checked against" ends the conversation; (4) exports in the venue's format, text still live text in vectors (format rules); (5) the prompt log archived with the manuscript source, because the revision letter will ask for figure changes and regeneration beats archaeology. Ten minutes, all five, and the AI-drafted figures become the least contestable part of the submission — which is the entire goal.
When the figure outlives the paper
Good schematic figures outlive their papers — reused in reviews, lectures, theses, and the next grant's introduction. That longevity argues for one extra investment at drafting time: make the figure self-contained. Complete labels, a caption that works out of context, no dependence on the paper's internal figure numbering, and a style that won't read as dated in three years (styles guide). The vector master matters here too — the figure that can be re-exported for any future format is the figure that keeps traveling (format rules). Draft for the paper; architect for the afterlife.
Next
Draft your schematic on the scientific illustration presets. Then: graphical abstracts with AI · verification protocol · AI-to-vector finishing · presentation and poster figures.
<!-- scientific-illustration-cluster-links:start -->
Continue through the scientific illustration collection
Create a scientific illustration · Browse the scientific illustration resource hub
Related: Scientific Figures for Presentations and Posters: AI Workflow Guide · How to Create a Graphical Abstract with AI (With Real Prompts)
<!-- scientific-illustration-cluster-links:end -->
<!-- scientific-authority-bridge:start -->
Start the figure in the right workflow
For a mechanism, process, or experimental setup, begin in the scientific diagram generator. For a specimen-led plate or a broader visual concept, use the scientific illustration generator. Keep the AI output as a disclosed draft until every label, relationship, unit, and claim has been checked against the paper and authoritative references.
<!-- scientific-authority-bridge:end -->
