Kagi Update: AI Image Filter for Search Results

this is a dumb feature, you should judge an image by your perception of it, not how it was created (ie machine or human made).

the anti ai-generated image crowd is a loud minority, they won't matter in the long term and spending dev time on this is questionable decision making at best.

now if you're a forensics company or that is the angle, then yeah it could be an interesting tool to have, might be even more profitable than this custom search as a service thing (obsoleted already by llm tech).

I agree that the source of an image often doesn't matter. But while it's completely _possible_ to make high quality images with AI that match almost any style you want, the current _reality_ is that most AI generated images are slop with an obvious "AI" feel to them, that most people are often not looking for. If I can get rid of those in an automated way, that saves me a _bunch_ of manual decisions, and makes finding what I'm looking for easier and faster.

A tool to help filter out AI content is useful if you are looking for real content. Anyway, who's being anti-AI in this case?

It is very cumbersome to look at individual images in the search for e.g. reference images just to sort out the garbage. Using AI images for that would defeat the purpose. Not just because of the fact that training AI with AI output degrades the model, indicating it's a net negative on average.

It's not a dumb feature, this is what I wished for less than a week ago when using Google. I don't want my time wasted from judging AI images based on perception, I don't even want to perceive them.

When deepfake images of MicolashKyoka start clogging up image searches you may wish for this feature.

It depends on what you're looking for. If I want a photo of a place I'm thinking of visiting, or a wild animal I'm trying to identify- I want to make sure it's a photo of the actual thing, and not a photorealistic AI artwork tagged with that name that may or may not have anything to do with the real thing.

I'm not anti AI but usually when I do an image search, I'm looking for photos of the real world, not artwork (from humans or AI)- and AI is getting so good I can't visually tell them apart.

This reminds me a bit of the XKCD about filtering chat comments comic. If you have an "AI slop" filter that hits false positives on poorly designed real images and has false negatives on high quality AI images, isn't that overall not just a positive, but potentially a better positive than a filter that perfectly filters AI with no false negatives or false positives?

> you should judge an image by your perception of it, not how it was created

Maybe when it comes to art, but not when it comes to anything else. If I want to know what something looks like in reality, AI results won't be of use.

Should I? If I want to see what something looks like, I want a photograph of it, not some half-confabulated garbage. Sure photos can be over-the-top edited and retouched, but at least they have a reasonable starting point. AI images don't; they have a tenuous connection to reality at best, especially if I care about little details.

Similarly, there is a definite qualitative difference between some actual hand-drawn art and something entirely generated by a model. It's a pretty obvious distinction and it's more than reasonable for people to care about it.

Not to mention how much AI-generated imagery is absolutely tasteless slop. That certainly describes the obvious AI examples in the article! If all the filtering feature does is block those—and, unfortunately, it probably can't do more than that—it would still be really great. Even without AI we were already beset by visual garbage; AI has only made it easier to generate it; having some way to even partially filter it out is the least we should aim for.

Not if you care about either the human effort that went into something—which, even if you don't care about anything "fuzzy", is still a costly signal in the economic sense!—or if you care about finding images that are representative of reality. Having a magical oracle that can filter out even really "good" AI imagery would be useful and, critically, would let us do something that is otherwise difficult.

>the human effort that went into something

Do people ever pay for the effort or only the result? Results that cost more effort are something they are willing to pay more for, but the effort itself is not something I've seen directly valued. Some will say they do, but I don't really remember seeing people actually paying more for same or less quality/results just because they took more effort.