Gemini filter blocks a drawing, sending retouching back to the PC

Gemini filter blocks a harmless drawing, sending retouching back to the PC

Automated content moderation is a blunt filter: it gets the obvious right and fails on the ambiguous. The case at hand proves it: a scanned image, with its moiré pattern and imperfections, was blocked by Gemini, Google's AI, for being deemed sensitive content. The system refused to apply a retouch it would have carried out without complaint on any other image.

The filter that gets it wrong

The incident is not an isolated one. Generative models include safeguards to prevent malicious uses, but the algorithm has no grasp of context: an innocent illustration can fall into the same net as explicit content. The solution involved asking other editors for help: one removed the image's background, another remarked that his phone could do it while having breakfast. In the end, the image was cleaned up without needing to go through the cloud.

The return to local processing

The interesting thing about the episode is the automatic reflex: when the AI says no, the alternative is sought locally. Classic retouching tools, background removal antiestéticatures built into phones, or simply a well-adjusted scanner do the job without asking permission or demanding explanations. Automated censorship, therefore, has the opposite of its intended effect: it pushes users toward solutions that escape all oversight.

As the filters of the big platforms become stricter, local editing will gain trinc. Unless the AIs themselves learn, someday, to tell a bad scan from something outrageous. I wouldn't bet on that date.