When “Better” Microscopy Becomes Fiction
In light of the current controversy surrounding the use of AI in the Nikon Small World in Motion competition, I want to point out that AI enhancement of still microscopy images may be even harder to detect. With a movie, our eyes have temporal information that can sometimes clue us in to the AI-ness of what we are seeing; with a single image, that information is gone. Here is an example. The top image is the best reflected-light image I was able to acquire when I was young and had time to do such things. I never had the proper equipment for this type of imaging, which is essentially macrophotography through a microscope. I was simply holding a cheap digital camera up to the eyepiece by hand, and this was the best result I could get. The image below was generated by giving that original image to ChatGPT with the simple prompt, “Make this image of an ant that I acquired using a microscope better.”
The original image contains limitations: blown highlights, relatively low local contrast, noise and grain, limited depth of field, and regions where morphology simply isn’t resolved. The AI version doesn’t merely suppress those imperfections; it converts uncertainty into apparent information. The head acquires extremely convincing cuticular texture, the compound eye gains a beautifully resolved ommatidial lattice, individual hairs become sharply defined, thoracic surface structures become crisp, and the mandibles and antennae become more cool looking. Importantly, these additions are biologically plausible. Nothing immediately screams “AI.” That is what makes this fundamentally different from ordinary sharpening, denoising, deconvolution, or contrast adjustment. The resulting image invites the viewer to make biological observations about structures that may never have been recorded by the camera in the first place.