← Back to blog

Blog

AI images in research and academia: why preprint figures need a different check

ai images research integrity

Short answer

An ai images research integrity problem looks nothing like a newsroom verification problem. There is no adversarial public audience, no viral distribution, and no urgency measured in hours. What exists instead is a figure inside a document, produced months earlier by an author with an incentive to make a result look cleaner than it was, reviewed by people with limited time and no forensic training.

Illustration: Short answer

Why the existing workflow does not transfer

A newsroom asks whether a picture depicts something that happened. A journal asks whether a figure accurately represents data that was collected. Those are different questions with different failure modes.

A fabricated news image fails by being too perfect: the wrong shadows, the wrong crowd, a detail that contradicts the record. A manipulated scientific figure often fails by being too consistent, since panels that should show natural variation turn out to be near-duplicate renders of the same underlying image.

Illustration: Why the existing workflow does not transfer

What actually gets checked, and by whom

Most journals run figures through a screening step before acceptance. In practice that step usually means a similarity search against previously published figures and a visual read by an editor. Neither reliably catches a panel generated from scratch, because a generated panel has no prior copy to match.

  • Duplication across papers: the most established check, and the one most often automated.
  • Internal consistency: whether brightness, background texture, and noise are consistent across panels that should share a capture session.
  • Splice detection: whether a figure is a composite of regions taken from different sources, visible as abrupt changes at panel boundaries.
  • Metadata: almost always absent by the time a figure reaches review, since figures are re-exported during manuscript preparation.
Illustration: What actually gets checked, and by whom

Where detection genuinely helps

A detector can flag panels that do not resemble a camera capture, which is useful for prioritising a queue when a journal receives more figures than it can inspect closely. It cannot establish that data were fabricated. A high score is not proof of misconduct, and a low score does not establish that a panel reflects real measurements. Where the score conflicts with a duplication finding, escalate to human review with both results attached rather than letting either one decide.

A workflow that holds up under appeal

  1. Preserve the figure as submitted, before any re-export during production.
  2. Record which check produced which result, including the checks that came back clean.
  3. Separate the finding from the inference: "this panel does not resemble a camera capture" is a finding, "this author fabricated data" is an inference that requires more.
  4. Route contested cases to a committee that includes someone with imaging expertise, not only editorial staff.

What this cannot do

It cannot confirm that data were collected as described. It cannot recover the original instrument output if the authors no longer hold it. And it cannot substitute for an institutional process, since misconduct findings carry consequences that require a standard of proof no automated tool can meet. Read the methodology for how the layers combine, or check sample reports to see how an inconclusive case is documented.

If you are reviewing a figure and need provenance, metadata, and detector output in one place before raising a concern, start a free review and record all three alongside the finding.

ImageVerity Verification notes

This article is written around current image-authenticity and AI-media verification topics. Brand references stay intentionally light so the article can remain news-led and useful first.

Review a real image with the workflow.

If this article was useful, the next step is to review a real file and inspect provenance, metadata, and detector evidence together.