Retouching by the rules: where science ends and deception begins ✛ ✎
On the screen is the result of a complex simulation: two planets collided, a shockwave ripples outward. A couple of clicks in an editor, and harsh shadows soften, noise smooths out, the background darkens. This kind of digital retouching (image editing to enhance appearance) has long been part of the scientist’s toolkit. But if you hand the 'enhance' command not to a person, but to Machine Learning (say, a generative neural network trained on millions of someone else’s images) the line between aesthetics and deception becomes blurry. A research team set out to find where exactly that line lies, and whether it depends on who does the retouching.
To tease apart the effect of the change type from its source, the authors prepared two sets of visualizations. The first was an abstract 3D scene featuring a teapot, into which 15 edits were manually introduced: from presentational (camera angle shift, lighting adjustment) to deep, altering the Digital Image Processing of data (noise removal, gap filling, shape distortion). Each alteration was shown as an animation, alternating between the original and edited images. Then 95 experts in computer vision and visualization design rated the acceptability of each edit. The scale ranged from 1 to 6. First, the author was a human, then an AI: Generative Adversarial Networks and Diffusion Models based on transformers.
The numbers formed a clear hierarchy. At the top were camera angle change (average score 3,38 out of 5), smoothing (3,27), and lighting adjustment (3,19). At the very bottom were moving objects (1,71), removing important details (1,62), adding nonexistent features (1,61), and shape distortion (1,40). The gap between the extremes was nearly threefold. Interestingly, Explainable AI didn’t help: even when the algorithm declares its steps, its edits are judged more strictly. neural networks systematically received lower scores, especially for previously acceptable changes. In open-ended responses, experts explained: humans are credited with intent, caution, and contextual understanding, while algorithms are seen as prone to hallucinations and lacking accountability. The hierarchy, however, barely broke down: both humans and machines equally fail when meaning is distorted.
For developers of AI visualization assistants, this means built-in safeguards are essential. Operations that alter shape or add details should come with warnings or be blocked. Survey participants emphasized that transparency of edits (which ones and why) matters more than explaining the neural network’s inner workings. Perhaps we’ll see dashboards where every change is documented and can be undone with a single click. And maybe 'verified by human' certificates for key publications, since only 14% of experts are categorically against AI tools, and two already use them in their work. Retouching is moving along an adoption curve. We’ve seen this before. Once, color correction was considered dishonest; today, smartphone filters are the norm. This study captures the early stage of that shift.
That very image of a planetary collision, having undergone honest retouching, could grace the pages of a journal. Not every embellishment is a sin. Transparency and honesty are the only currency of trust. And in five years, AI filters on scientific images will be treated as matter-of-factly as smartphone filters are today.
The same visual outcome (say, an object shifted from the center) was judged differently: if it resulted from a camera angle change (acceptable), it was accepted, but if from moving the object itself (unacceptable), it was rejected. Experts judged by intent, not by the image.