Nature Structural Molecular Biology AI Policy: ChatGPT and Generative AI Disclosure Rules for NSMB Authors

Nature Portfolio says generative-AI images are not permitted in publications, except narrow cases under the policy. For NSMB authors, treat that as a high bar. A structural-biology figure is usually not decoration; it is evidence.

The policy risk is highest when AI touches:

• cryo-EM density maps, model overlays, local-resolution panels, or model-validation plots;
• X-ray crystallography electron-density figures, ligand-fit panels, or refinement summaries;
• microscopy images, segmentation masks, colocalization panels, or representative fields;
• AlphaFold, RoseTTAFold, docking, or molecular-dynamics visuals that are shown as mechanistic support;
• graphical abstracts or schematics that imply a molecular pathway, binding model, or conformational state.

Recent NSMB examples show why image and model provenance matters: cryo-EM structure of SID-1 and dsRNA (10.1038/s41594-024-01277-8), gamma-TuRC cryo-EM structures (10.1038/s41594-024-01345-z), protein-coat assembly around cytosol-invasive bacteria (10.1038/s41594-024-01403-6), and RNA-guided RNA editing by a Cas13b-ADAR2 complex (10.1038/s41594-025-01529-1). These DOI examples are not AI-use claims. They show the evidence surfaces where undisclosed AI involvement would affect reader trust.

Use specific language. The editor should be able to see the tool, the use case, and the author verification path.

"During manuscript preparation, the authors used ChatGPT-4o to improve English-language clarity in the Introduction and Discussion. The tool was not used to generate data, select references, build structural models, alter figures, analyze cryo-EM maps, or write the final mechanistic interpretation. All authors reviewed and edited the text and take responsibility for the final manuscript."

For AI-assisted code, modeling, or structure-prediction work, make it a Methods detail rather than a generic writing note:

"Claude Sonnet 4 was used to draft an initial Python script for plotting model-validation statistics. The authors rewrote and validated the script, checked all outputs against the deposited coordinates and map files, and archived the final code with the analysis package. The tool was not used to choose reported structures or alter figure panels."

Weak declarations create avoidable risk:

• "AI tools were used in manuscript preparation." This does not name the tool or manuscript surface.
• "ChatGPT helped with writing." This does not say whether the tool touched Methods, figures, interpretation, or references.
• "AI was used for analysis." This is too vague for structure-prediction, model-building, or image-processing workflows.
• A cover-letter-only note. Nature Portfolio policy points authors toward Methods or a suitable manuscript section.

In our pre-submission review work for Nature Structural & Molecular Biology and adjacent molecular-biology journals, the AI-policy risk is usually not simple copy editing. The risky manuscripts are the ones where AI quietly touches the structure-to-mechanism bridge.

That bridge is the NSMB editorial problem. A paper may have a strong dataset, but if AI-assisted wording makes the mechanism sound cleaner than the validation supports, the disclosure is no longer a formality. It becomes part of whether the manuscript is honest about uncertainty.

In the NSMB manuscripts we review before submission, we check the AI statement against the manuscript parts that create reviewer reliance: Methods, figure legends, model-validation panels, PDB or EMDB deposition language, supplementary methods, code availability, benchmark tables, and response-to-reviewer drafts. The recurring gap is not that authors used a tool. It is that the declaration says "writing assistance" while the manuscript package shows AI influence on a structural or mechanistic surface the reviewer needs to audit.

Common NSMB AI-policy failure patterns

Structure-prediction language presented as experimental evidence. Authors use AI-assisted structure prediction or model-building language, then write the Results as if the model had the same status as an experimental structure. The declaration needs to separate prediction, validation, and experimental support.

Cryo-EM or crystallography figure provenance left unclear. A figure panel combines maps, models, overlays, or segmentation with polished labels, but the Methods do not tell the reader whether AI helped create, enhance, or interpret the visual. For NSMB, that is not cosmetic.

AI-assisted code hidden under a writing disclosure. The manuscript says AI helped with prose, while code for map statistics, binding-site plots, or structural comparisons was drafted with an AI tool. If code affects a plotted result, the Methods should name the tool role and author validation path.

Literature benchmarks drifting through AI summaries. NSMB papers often live or die on how they compare a molecular mechanism to prior structures. If AI helped summarize the comparison set, authors need to verify every citation and avoid letting a generated summary overstate novelty.

Mechanism claims strengthened during AI rewriting. We often see cautious author language become sharper after AI polishing: "may stabilize" becomes "stabilizes," or "consistent with" becomes "demonstrates." That can change reviewer perception even when the data have not changed.

A missing or vague AI disclosure can lead to an editorial query, rejection, correction, retraction, or publication-ethics review when the undisclosed use affects authorship, images, code, data, or interpretation. Nature Portfolio also asks peer reviewers not to upload manuscripts into generative-AI tools because submitted manuscripts are confidential.

The ambiguous cases are the ones authors should resolve before upload:

• AlphaFold or other prediction tools: describe the tool and validation when it supports a claim.
• AI-assisted code: disclose it when it shaped model fitting, image measurement, or figure quantification.
• AI-assisted figure drafting: avoid it for evidence figures; check the policy before any conceptual graphic.
• AI translation: disclose when the tool materially rewrote scientific claims, not when it only fixed surface grammar.
• AI citation discovery: verify every citation manually and disclose if the tool shaped the literature screen.