Generation and verification record
Distinguish generation from editing
A fully synthetic illustration, generative expansion of a photograph, object removal and colour correction are different interventions. Describe the function actually used. “Made with AI” can be too broad where only a background changed; “author's figure” can conceal a substantial generated contribution.
If an external image was the starting point, cite and assess it separately. A prompt naming an artist or protected character is not rights clearance. If your own photograph was altered, depicted people, places and confidential details remain relevant. Preserve source and final versions so the intervention can be inspected.
Caption patterns
Synthetic illustration: “Figure 9. Symbolic view of the workflow. Generated with [service, displayed model], 4 September 2026, prompt P-14; selected and labelled by the author. Not an empirical data display.”
Edited photograph: “Figure 10. Experimental setup. Author's photograph; background generatively extended with [service]. Measuring devices and labels unchanged; extension marked in grey.” Data-led graphic: “Figure 11. Category visualisation. Layout proposed with [service]; values from dataset X, version 2, verified and inserted manually by the author.”
Follow any mandatory wording, tools register or appendix required by the assessment. The caption should still make material generative involvement visible at the object itself.
Check facts, not visual plausibility
Generated images can contain anatomical, technical, spatial and textual errors. List every visible proposition that could be read as factual. For laboratory equipment, compare devices, connections and safety features with real documentation. Maps, historical scenes and scientific diagrams need sources for place, period and relationships. Decorative imagery must not masquerade as empirical evidence.
- Inventory every visible factual assertion.
- Map each one to independent evidence.
- Check labels and units manually.
- Remove invented or unresolved elements.
- Mark symbolic and non-scale components.
- Compare the final figure with methods and sources.
Review accessibility. Alternative text communicates the evidence relevant to the argument and can identify the synthetic character where useful. Put important text in the document, not only inside pixels, and never rely on colour alone.
Rights, privacy and scholarly limits
Do not upload unpublished research images, personal data or confidential drafts to an unapproved system. Synthetic faces may resemble real people; trade marks and protected characters may appear unintentionally. Record your review and replace risky elements. Embedded metadata can support the record but does not replace a visible caption.
Disclosure cannot rescue a false scientific diagram. Nor does a prompt containing references support the image's claims. Cite data, models and facts at their originals. The media citations hub covers other formats. The AI use declaration supports the overall account. The citing sources guide explains evidence references.
Submission and archive audit
- Is AI involvement visible beside the figure?
- Are system, date and intervention specific?
- Do source material, prompt ID, variants and final output remain connected?
- Are depicted facts independently supported?
- Are human selection and editing distinguished?
- Were rights and sensitive content reviewed?
- Are symbolism and limitations explicit?
- Do caption, figure list, methods and declaration agree?
Assign a figure ID and keep original input, raw output, editing file and published version together. Record why this variant was selected. A later edit creates a new version; do not overwrite the earlier evidence. This converts a label into a traceable production record.
Check whether the image is necessary. A conventional diagram built from verified shapes and text may communicate a factual process more reliably than a photorealistic generated scene. Choosing not to use a persuasive but unverifiable output is itself an important scholarly decision and can be recorded briefly.
Worked case: a synthetic process image
A generated laboratory image adds a valve and reverses two steps. The researcher checks every element against the real setup, rejects the first output and builds a simpler diagram. The caption still discloses generation; experiment records support the actual sequence.
Retain rejected variants with fault categories: wrong label, invented object, risky reference, misleading perspective or inaccessible design. This avoids showing only the persuasive result. Record manual replacement of image text and verify terminology outside the generator. In a series, connect each version to its prompt and displayed model.
Compare caption and alternative text. Neither should borrow certainty from photorealistic styling. If a conventional diagram communicates the process more reliably, rejecting the generated picture is a defensible quality decision worth recording. Illustration can support explanation but cannot replace measured data, real documentation or a source.