I don't agree with that take.
We rely on publicly available lists to filter out AI-generated content, like those provided by uBlockOrigin & uBlacklist Huge AI Blocklist, an open-source blocklist, manually curated by project contributors.
This is an open source block list, which is manually put together by volunteers. It's not even reading metadata, presumably because that's too resource intensive (or proprietary/unreliable).
Gemini was able to tell that was AI because it was able to read the image ID water mark and associated it with how Adobe sequences AI generated images. The image ID is in the file name, but if we wanted to catch it that way the filter would need to work on logic, not just keywords or site lists. Programming a filter to search the file names to identify if it belonged to the sequence would be more complex.
Alternatively, if you clicked though to the image you are immediately told if it is AI. Whenever I search for microbe photos I always click the source to determine if it's a render or a stylized electron microscope capture, and I figure it's always been good practice to get the image from the source to verify it's not totally bullshit either in terms of the the search pulling up something unrelated or the image itself being misleading or even a humorous misrepresentation/meme.
I think it's important to understand how the tools we rely on work.