Safeguarding nanomaterials science from AI-faked microscopy images

The rising danger of AI-generated images in nanomaterials science - and what to do now

It's now trivial to produce fake microscopy images that fool skilled reviewers. With a few prompts, models can emit TEM, SEM, AFM, and fluorescence images that look valid, complete with scale bars, plausible noise, and "good" particle size distributions.

This isn't a hypothetical. Side-by-side tests show that fakes guided by real data can be indistinguishable, and purely text-prompted images can look authentic enough to pass a quick check. Treat this as a present risk to data integrity, not a future concern.

Why these fakes are so convincing

• Models have strong priors for particle morphologies, lattice fringes, and contrast patterns common in nano-imaging.
• They can generate "believable" noise fields and depth cues that satisfy what reviewers expect to see.
• Scale bars, labels, and panel layouts are easy to synthesize and align.
• Even experts anchor on plausibility. Without provenance, authenticity is guesswork.

Immediate lab-level actions

• Commit to full image provenance: Store and share raw instrument files (e.g., .dm3/.ser/.tif), not just exports. Include beam conditions, magnification, detector settings, calibration files, and software versions.
• Hash at acquisition: Generate checksums for raw files when captured and keep them in versioned lab records. Recompute on submission to verify integrity.
• Lock processing pipelines: Save processing scripts, GUI logs, ROI masks, filters, and parameters. Export a step-by-step record from raw to final figure.
• Separate illustrative from evidentiary images: If synthetic images are used for communication, label them clearly and never mix with data.
• Two-person verification: Require an internal "image audit" before submission: scale bars, metadata, processing steps, and consistency with methods.

The MAIF storage principle

Organize every figure around four linked layers: metadata, acquisition files, intermediate processing, and the final figure. Keep them together, cross-referenced, and checksum-verified.

• Metadata: Instrument ID, session time, operator, sample prep, calibration files.
• Acquisition: Original microscope outputs and instrument logs.
• Intermediate: All transforms, masks, and scripts that touched the data.
• Final: Exported panels with scale bars and captions tied back to source files.

This makes audits fast, repeatable, and fair - without adding heavy overhead.

Editorial and funder policies that work

• Mandatory raw data deposition: Require original instrument files and processing records on submission and publication.
• Provenance statements: Short, structured description of acquisition, processing, and verification steps for each figure.
• Standard checklists for reviewers: Simple image integrity items: scale calibration, noise structure, metadata presence, processing disclosure.
• Randomized image audits: Journals or funders audit a subset of figures each issue/grant cycle.
• Clear labeling rules: Synthetic/AI-assisted images must be explicitly marked and excluded from evidence unless justified and approved.

Triage signals for spotting AI-generated microscopy

• Scale bars and labels: Crispness inconsistent with image resolution; mismatched bar length vs stated magnification; identical bars across unrelated panels.
• Textures and noise: Repeating motifs, isotropic "too clean" noise, or pattern duplication across particles or fields of view.
• Lattice and edge artifacts (HRTEM/AFM): Locally perfect periodicity without defects; fringe terminations that don't align with crystal symmetry.
• Metadata gaps: Missing instrument IDs, absent acquisition parameters, or EXIF tags inconsistent with instrument software.
• Panel consistency: Identical dust, scratches, or background features appearing in separate "independent" images.

No single indicator is decisive. Use them as prompts to request raw files and processing records.

What to automate

• Instrument-signed outputs: Work with vendors to embed cryptographic signatures in raw files at capture.
• Checksum pipelines: Automatic hashing on acquisition and before submission.
• Provenance manifests: Auto-generated summaries that bundle metadata, hashes, software versions, and processing steps alongside figures.
• AI screening as triage: Use detectors only to flag images for human audit. Avoid overreliance on classifier scores.

Community steps

• Open benchmarks: Curate sets of real and AI-generated microscopy for method testing and reviewer training.
• Shared SOPs: Publish lab SOPs for image provenance and audits; reuse what works.
• Clear norms for disclosure: Standard wording for any AI involvement in figure creation or enhancement.
• Training: Short, practical modules on image integrity, provenance, and AI risks for students, staff, and reviewers.

Practical resources

• NIST AI Risk Management Framework - useful structure for lab and journal risk controls.
• COPE guidance - policies and checklists to strengthen image integrity and editorial practice.

If you need to upskill your team

AI literacy reduces misuse and improves review quality. If you want structured options for different roles, see AI courses by job.

Bottom line

Assume highly convincing AI-generated microscopy exists - because it does. Raise the standard of provenance, automate audits where possible, and make disclosure non-negotiable.

These steps are simple, measurable, and enforceable. They preserve trust without slowing science.