On a scientist's desk, an image glows — a computer simulation of two planets crashing into each other. You want to tweak the shadows a bit, remove some noise, so the viewer can see the main event. But where does simple enhancement end and fabrication begin? Researchers showed 95 specialists in computer vision, or image analysis—15 different ways to retouch scientific images, from innocuous camera rotations to adding non-existent details. Half the edits were done by humans, the other half by a program using Machine Learning (trained on examples).

The ratings formed a clear ladder. Rotating, adjusting lighting, and smoothing noise scored high (around 3 out of 5). But moving an object or changing its shape were nearly forbidden (below 2). The most striking finding: when the exact same edits were carried out by software, experts gave lower scores. Why? Humans are credited with understanding intent and exercising caution, while a program seems untrustworthy: it might “hallucinate” details that don't exist. Even if the program tries to be Explainable AI (disclosing its steps), trust doesn't increase.

These findings matter for developers of Digital Image Processing tools. Modern neural networks (including Diffusion Models, Generative Adversarial Networks, and transformers) are increasingly used for retouching. But, as it turns out, any algorithm that changes shapes should warn the user. Moreover, all edits ought to be logged, like browser history. Because in science, the key is not a perfect picture, but trust in what it shows. With honest retouching, that very image of planetary collision could hang on the pages of a journal. Not every touch-up is deception. Transparency: that's the real currency.