Anyone can fake a scientific image with AI, tricking even academic journals - and undermining trust in science

They are asking the wrong question. The headline says AI-generated images are “undermining trust in science.” As if trust were a vase on a shelf that a clumsy machine might knock over. The trust was already shattered. The image is only the latest symptom. For decades, the structure of scientific publishing has been a plantation system of its own: the labor of researchers, often publicly funded, harvested for profit by a handful of corporate journals, their work locked behind paywalls, their careers held hostage to the whims of a peer-review process that is as opaque as it is arbitrary. The trust was not in the image; it was in the process. And that process has long been compromised.

Now the forger’s tools are democratized. Anyone, they say, can fake an image. But who has been faking the narrative? Who has been presenting corporate-funded research as disinterested truth? Who has been airbrushing out negative results to maintain the facade of seamless progress? The AI did not create this crisis of credibility. It has merely exposed the fault lines that were always there, the gap between the professed ideals of open inquiry and the closed, proprietary reality. The periphery - the early-career researcher, the scholar from the global south without institutional access - has always seen this logic most clearly. Now the machine makes it visible to everyone.

The solution is not better detection software. It is a different political economy of knowledge. When the enslaved took Saint-Domingue, they did not ask for better whips. They seized the means of production. Science must do the same. Open archives, communal peer review, a dismantling of the prestige economy that rewards splashy, easily-faked images over slow, reproducible work. The fake image is not the problem. It is the alarm. And the alarm is ringing because the house was built on sand.

June 23, 2026.

The reports of synthetic imagery infiltrating the journals do not surprise me, but they necessitate a rigorous definition of terms. What do we mean by “scientific image”? If we mean a record of an observation, it is a measurement. If we mean a decorative representation of a theory, it is an illustration. The crisis today arises because the journals have conflated the two. An illustration requires only aesthetic plausibility; a measurement requires a chain of provenance.

This crisis rests on two assumptions. First, it assumes that the prestige of the journal - the institution - is a proxy for the truth of the claim. Second, it assumes that visual “evidence” is self-authenticating. Both assumptions are fictions. The institution is a stone portico; it can be occupied by anyone. The image is merely a shadow on the wall of the cave.

Let us apply a geometric proof to the problem of trust. Axiom: Science is not the belief in results, but the verification of the process. Step one: An AI generates a result without a process. Step two: A reviewer accepts the result because it looks familiar. Conclusion: The reviewer has abandoned the method for the sake of the institution’s schedule.

When the image is faked, the institution suffers a wound, but the method remains untouched. If every journal burned tomorrow, the Pythagorean theorem would remain true. The solution is not better software to detect fakes; it is a return to the portable method. Demand the raw data. Demand the proof. If a claim cannot be reconstructed from its first principles by a skeptical mind, it is not science; it is merely a story told with light. We must protect the method of inquiry, for it is the only thing that survives the fire.

The alarm over AI-forged images in scientific journals is understandable, but misplaced. The problem is not new; it merely wears a digital mask. Fraudulent data has always existed - fabricated tables, manipulated photographs, selective reporting. The tool changes, not the temptation.

What interests me is the disproportionate reaction. If a researcher alters a gel electrophoresis image by hand, we call it misconduct. If an AI does it, we call it a crisis for science itself. The standard of evidence should not fluctuate with the method of deception. A lie is a lie, whether penned or pixelated.

The deeper issue is the over-reliance on images as proof. Visual evidence has always been persuasive, but it was never sufficient. A photograph of a specimen does not replace the specimen; a graph does not substitute for raw data. The error lies in treating images as conclusions rather than illustrations. If journals now fear AI’s capacity to deceive, perhaps they should revisit their own assumptions about what constitutes proof.

Science has survived worse. It will survive this, provided it remembers that trust is not built on the absence of fraud, but on the presence of verification.