On June 22, 2026, The Conversation ran a blunt warning by Nan Li: “Anyone can fake a scientific image with AI, tricking even academic journals — and undermining trust in science.” The claim lands at a bad time for research integrity. It now takes minutes, not months, to mint lab-looking pictures. That raises the stakes for editors, reviewers, and funders who rely on pictures as evidence.
What The Conversation found about AI fake scientific images
The Conversation’s piece, published on June 22, 2026, argues that visual proof in many fields has become a soft target. Microscopy panels, western blots, gel bands, and cell culture plates can be fabricated by text prompts or fine-tuned models. The article frames the risk in plain terms: journals are being fooled, and the cost of fabrication keeps falling. That combination puts the scientific record in harm’s way.
Fields from cell biology to materials science lean on figures to carry key claims. Reviewers judge whether a band is present, whether a colony grows, whether a lattice is uniform. If a model can synthesize that signal with convincing noise, older checks miss it. The Conversation’s authors argue that trust built on visual evidence alone no longer holds. The message is clear: treat imagery as data, not decoration.
Why peer review misses AI-generated figures
Peer review was designed to stress-test ideas, not authenticate pixels. Most journals do not provide reviewers with raw image stacks or instrument logs. Many rely on compressed PDFs, where artifacts can blur red flags. According to The Conversation, that gap lets fabricated figures slide through, because the review workflow assumes good faith and limited manipulation.
Editors are also fighting volume. Paper mills have raised output while getting better at pasting plausible methods around pictures. Nature reported on June 28, 2023 that publishers banded together to counter such mills, but the tactics keep shifting. Reviewers, working nights and weekends, often lack time and tools to run forensic checks on every panel. That’s not an excuse; it’s a workload reality that bad actors exploit.
Some journals do screen figures, but coverage is uneven. Guidance exists, yet adoption varies by title and field. The Committee on Publication Ethics (COPE) offers frameworks for handling image concerns, and publishers such as PLOS spell out strict figure rules and file requirements in their submission guidelines. Policies help, but they don’t catch problems when images are never checked at the right resolution or without access to originals.
Detection won’t be enough on its own
Many hope automated detectors can flag synthetic content at scale. They help, but they break when models change. Studies show that deepfake detectors often fail to generalize beyond the specific generators they were trained on. One analysis asked a simple question — do detectors generalize? — and the answer was usually no when faced with new methods (arXiv:2005.05432).
False positives also carry a cost. A legitimate blot called “fake” can harm a lab’s reputation and stall a student’s degree. False negatives let bad data into citation chains where it can steer real experiments astray. The Conversation’s warning lands here: AI fake scientific images shift the odds toward both errors at once. Review pipelines need redundancy that doesn’t rely on a single classifier.
Detectors improve when paired with access to source data. If reviewers can check raw microscopy stacks, acquisition timestamps, and instrument settings, many fabrications crumble. That’s verification, not just detection. It’s also slower work, which means journals must choose where to spend limited effort.
Provenance and signed media can help, with limits
One fix gaining traction is content provenance. The Coalition for Content Provenance and Authenticity (C2PA) standard defines a way to attach trustworthy creation and edit histories to media. In science, that could mean lab instruments and imaging software writing tamper-evident logs at capture time, then carrying those credentials into figures.
Provenance changes the default question from “is this fake?” to “can the origin be proven?” It won’t stop a bad actor from stripping metadata, but it gives journals a fast way to reward compliant labs while routing unsigned figures to deeper checks. The same playbook appears in newsrooms adopting content credentials to signal what a photo is and where it came from.
Provenance is not a silver bullet. It is opt-in and unevenly supported across devices and software. Legacy lab cameras, custom analysis pipelines, and exported TIFFs can break chains. Still, pairing provenance with targeted forensics and clear policy creates friction for fraud without drowning everyone in audits.
What labs and journals can do now
There’s no single fix, but several steps move the needle fast.
- Require raw data on submission for image-heavy claims. Give reviewers private access to original files, acquisition logs, and analysis scripts.
- Run triage checks before peer review. Basic duplication scans and contrast/edge analysis catch many low-effort fakes quickly.
- Adopt clear, public figure policies. Point authors to concrete rules on exposure, contrast, splicing, and quantitation, as in the PLOS guidance above.
- Invest in training. Short modules for editors and reviewers on common red flags help build shared instincts.
- Build a provenance lane. Encourage or require signed images where feasible; route unsigned, high-stakes figures to manual scrutiny.
External watchdogs already surface problems. Retractions and image concerns are documented by efforts such as Retraction Watch and independent analysts who examine figures at scale. Those signals should feed back into editorial checks and community norms, rather than turning into post-publication cleanups only.
None of this removes the value of human judgment. It refocuses it. Editors decide which submissions demand the strictest examination. Reviewers ask for raw data and say no when it doesn’t arrive. Funders and institutions back these demands with time and tools. The alternative is more corrections years after publication, when the damage is done.
Why this matters beyond one essay
The Conversation’s June 22, 2026 article names the threat. The deeper story is the posture shift it demands. AI fake scientific images aren’t a niche problem for one field or one publisher. They strike at how evidence gets trusted across journals, preprints, and grant reviews.
Science has faced figure fraud before. What’s new is scale and speed. Generative models make it easy for almost anyone to fabricate a plausible panel, then try again if a detector blocks the first attempt. That means the defense has to be layered: policy, provenance, selective automation, and a culture that treats figures as auditable data.
Readers don’t need to accept every processing step as magic. They should expect proof of origin when images carry the main claim. Journals that move first will set the norm. Those that wait will host the next wave of retractions. The warning from The Conversation is less about panic than priorities: spend scarce review time where evidence can be faked fastest, and make AI fake scientific images the first stop for stronger checks. For more on this, see bloomberg.com and nytimes.com.