Kagi Update: AI Image Filter for Search Results | Hacker News Reader

Kagi Update: AI Image Filter for Search Results

help.kagi.com

271 pointslkellar105 comments

For those who haven't jumped ship to Kagi, there's a uBlacklist feed which strips out most big sites dedicated to AI images, with an optional extra "nuclear" feed which also knocks out sites that aren't strictly dedicated to AI images but do have a very large proportion of them.

https://github.com/laylavish/uBlockOrigin-HUGE-AI-Blocklist

Still, give Kagi a chance. I don't work for them, I don't have friends who work for them, I'm just a guy who uses Kagi and will never look back. It isn't expensive and it's SO worth it.

Same, best digital product I’ve spent money on in a long time. It’s an improvement over Google, and well worth the price.

I tried it for two months and it was good, but not that much better that I'd spend $10/mo on it. The results were basically about the same as DDG. I can't really relate to all the praise I see about it.

However, I do have to say that, when I was looking for a very specific post, I spent around twenty minutes on Google and DDG and they came back with trash, whereas Kagi found it right away. In that one instance it was, indeed, fantastic.

I don't feel I have enough ground truth to know how good Kagi's search is in absolute terms (eg. there are plenty of searches where I still just "fail to find" and I don't know if it's my fault or theirs), I just know that I get less "junk" results with it than with DDG by far. With the additional ability to customize results and filters, I'd say it's a good product and it's worth a small subscription fee.

The quality of all free search is just bad. Kagi, even when it fails, is basically the difference between me looking for something or just giving up and deciding it's not worthwhile. Kagi's not magic-tier like early Google was - it's basically just the only modern web search engine.

“fail to find” is one of my favorite aspects of Kagi. When I get that message, I know I need to adjust my search terms. In contrast, Google wants to show me something even if there are no good results, but when it’s just junk, that wastes my time.

Been using it for a week and it is so nice to not have to your results cluttered by marketing bs. I'm faster at work as well because I find actually relevant informational.

Switched about ten months ago and it’s been so good I forget it’s even not Google sometimes.

If you forget it's not Google, I don't want to sign up. I want something that reminds me all the time that it's not Google, mainly by its ability to actually find the things I search for.

I suppose what they mean is that it's easy to forget that it isn't Google from ten years ago when Google was good?

Exactly. Going to real Google now is like ?? How did I use this before?

It's very much not Google and you find things faster. They have 100 free searches for you to try.

I'm not sure what they mean. I have mine tuned to the point that I can immediately tell when I accidentally google something because the results are so shockingly bad.

Kagi is REALLY good at giving you the ability to filter out the type of results you never want to see.

My main problem is that I don't so much have results I never want to see, as I want to be able to get results from blogs and small sites, the articles that are really interesting.

Google never shows you anything any more from sites that aren't in the top 100 sites by traffic, and many of my searches don't fit that pattern.

Kagi does better, but there isn't a "downrank all the low quality spam" option.

Small Web is for you: https://blog.kagi.com/small-web

Same. I’m so used to it now that when I see a google search page I think “woah, what is this? Oh. Yeah. Bleh”.

The only thing that reminds me it’s not Google is when I search for a place I don’t get a good map that then allows me to route to the place. Still use Google for that.

For me there are plusses and minuses. It doesn't push ads, but it doesn't seem quite so good at picking out phrases in the query. So I will search for something and then have to go back and quote the phrases

I don't work there or know anyone who does either, but I wish I did. Kagi is great.

It is kinda expensive, but the quality is very high and I search all day long. It's definitely worth it.

I'd love to still use it, alas they've partnered with Brave - which made my stop immediately after hearing that - partnering with a crypto+advertising company - no thanks. It might still be the best option of the bunch, which is sad.

Source? All I can find is them integrating Brave Search API as one of their multiple sources. I wouldn't call that a partnership.

It seems they've redacted pretty much everything related to that, and any search sources for that matter. Had to join the discord to even see someone talk about this;

https://layer8.space/@k/113350415680785878

I was a paid Kagi user.

It’s one of those tools where I have to say, “it doesn’t fit me but you’re doing something good in the world so keep going.”

I don’t do the kind of searches where Kagi is a lot better than Google (I bet folks here do).

On the searches I do want to do, i.e steaming movies, local business and map related, Kagi is not yet strong at. I kept having to !g.

So I’m back to Google which I find fits me better.

But I’m glad Kagi exists.

The strange thing is that kagi gets most of its search results from other providers like brave, bing and I believe even Google. It should be able to find those things.

I use SearXNG by the way. Kagi is better but I like the way I can configure SearXNG.

Try looking for your a local Thai restaurant in your area. Kagi doesn’t profile you so it doesn’t know where you are, so any local searches involve adding a keyword or doing the search a second time with !g.

I just tried it on Kagi and it suggested a Yelp link for restaurants in McKinney TX which is nowhere near where I am.

Google uses what it knows about you as context — which arguably many folks here are against— but it does get me the right results in one go.

Ah ok that explains a lot. I use a lot of privacy protection when I use any Google features (like running them in a separate Firefox container so it can't see other web activity). So I'm used to specifying that already.

> Google uses what it knows about you as context — which arguably many folks here are against— but it does get me the right results in one go.

Kagi does let you opt-in your location to improve searches. So there is a way and the user is in full control.

I just looked and I can't find any option to localize results more specifically than "United States." I'm not finding anything in a KB search either. If the option is there, it has really poor discoverability.

I love the service, but poor local results is definitely one of its weak points and basically the only reason I ever fall back to Google anymore.

I’m not sure it exists. You can opt in on Kagi maps but not search.

Cancelled my sub recently. Not because I don't like it, but because I'm taking a break before starting my next job and wondering if I can get by without it.. Guess only thing I don't like is their cheapest tier is just slightly under my usage and the next tier is unlimited (more than I need).

I had cancelled mine to see if I benefited from it as well. I re-subbed a few days ago. I was looking for a carpet repair shop in my area a week prior and Google was showing me one, which had turned out to be permanently closed when I went there.

I did the same search on Kagi. Not that Kagi has a strong maps feature or anything but turns out there are at least 4 carpet repair shops in my area, which Kagi showed as top results. Called one, looks good, went and dropped it off. It was 4 minutes away from my location.

I'm surprised it worked for something I have gotten so used taking Google's word for.

Thanks for this. Any other blocklists you can share?

This is a solid idea, but I wish they addressed the elephant in the room — image search is by far the weakest part of Kagi. For a considerable portion of queries, a large part of the results isn't relevant. If you use filters, they're often ignored or don't apply correctly. Many images are tangentially related at best. The list goes on. I've been paying for Kagi for a long while, yet I've seen nearly no improvements on this front. Image search is one part of their product where I often go to Google or other options because what Kagi does there just doesn't cut it.

I agree. I’ve often wondered, what could be the possible reason given Kagi is using google APIs behind the scenes?

Do you have any example to share? (we do not hear this feedback frequently so want to make sure we address it, thanks!)

It’s hard to put it into solid feedback and examples, because it is highly subjective. But I also find the image results from Kagi lacking, while I really enjoy the text based results. Especially for more specific queries, the image search just doesn’t hold up.

I just tried generating an example. Take the query “screenshot nero burning rom windows xp” - of the first 10 images, only 6 are screenshots of the program on Kagi. On Google, it’s a solid 10/10.

Of course it’s hard to take just one example, but it reflects the general feeling I have when using the image search quite well. The results aren’t necessarily terrible, it’s just that they aren’t as relevant as Google’s.

I see both Kagi and Google have 9/10 screenshots of Nero.

To make it easier for you to report any discrepancies I created a bug report with screenshots of what I see.

https://kagifeedback.org/d/5073-investigate-image-search-res...

Would you mind being specific and posting a search quality issue with a concrete example(s) to kagifeedback.org? We are keen to address the issues you are seeing.

I actually did post on your feedback site nearly a year ago, still no feedback or response on it so far [0]. It's been marked as under review, so it must be one thorough review process.

[0] https://kagifeedback.org/d/2565-image-search-doesnt-respect-...

Thanks, replied in the thread.

That's kind of you, but if you have to resort to HN comments as your feedback system, then your existing feedback system doesn't really work, does it?

I do not think this is a case of that happening like with other big tech names treating customers like shit until they make it onto HN. Vlad's super active (like, is he ok?) on the Orion browser bugs/feedback forums so I have to assume the same is true of Kagi, their actual money maker. Sometimes things fall through the cracks. One data point isn't a trend.

I'm a subscriber. But I don't feel like creating a signup for kagifeedback.

https://kagi.com/images?q=https%3A%2F%2Fs3.amazonaws.com%2Fp...

Lets use this as an example. I would personally like to have a list of exact results. Separately - a list of similar images would maybe be nice. But tbh, 99.99% of the time with reverse image search, I am trying to play detective, not find similar images. I am usually looking to see the first, original source of something. Or maybe other places I can find this image.

I will point out that detective stuff like this is crucial to try to prevent being catfished, or phished. I am not ignorant that it is also a concern for those who don't want to be found, for privacy or safety reasons; however some threat actor could just find a less public reverse image search I'm sure.

> I'm a subscriber. But I don't feel like creating a signup for kagifeedback.

It's a separate login? That's really weird, isn't it?

Not really. It's not that uncommon for a support/forum/feedback site to use a separate account. It just means that they didn't have time, didn't want to spend the money, or couldn't link the support software to their main user account system.

> It just means that they didn't have time,

That's it!

Yandex image search is really good.

Probably the best tbh

Well, your anecdote is completely contrary to mine. Image search has always worked great for me and it's easier to save the image because Kagi doesn't play games with the source of the image like Google Images does.

It's Google Images that I find unusable.

Image search is the one area in Kagi that I've seen the most improvement over the past year. When I first subscribed, I'd often switch to Google to find what I was looking for. Nowadays it works exactly as I intend.

Interesting, another battlefront for AI vs. AI.

Probably will be well-liked overall by Kagi customers, I'm sure. I'd be a bit concerned about false positives, but I suppose the stakes are pretty low compared to other similar situations (e.g. using AI to detect AI-generated essays in University), so there's not much of a concern.

Is there any mechanism to provide feedback for false positives?

> Is there any mechanism to provide feedback for false positives?

I think this could possibly be nice reward program, provided guardrails are in place to prevent abuse.

Just don't search for "baby peacock". The AI filter is no match for the famous baby peacock, I say that because the AI pictures are actually now shown because they are part of articles discussing AI generated photos.

First thing I tried and was disappoint.

I tried the the linked example search for "baby cat" and it returned the same three AI cats you can see in their Google search comparison screenshot on the first page. None of them labeled as AI generated.

Edit: When I explicitly choose to "Include" AI images from the toolbar option, they disappear. When I choose to "Exclude" them, they reappear. Still seems a bit buggy.

Been using Kagi for a good year or more now and am very happy, but this feels like a real level-up.

This is exactly the kind of thing I want to be paying for. It doesn't even matter if it's not 100% accurate (I don't think it ever could be without some serious processing), the commitment to down-ranking sites that have low quality content is the whole ball game for me.

Kagi has a lot of features that seem “neat”. In fact, their pitch and feature set was enough to convince me to sign up. However, I haven’t been wowed by the service. I do not find there to be a noticeable difference above or below any other search engine I have used (all of which I would rank about the same).

I’m curious what other Kagi users are doing or finding that makes their experience so overwhelmingly positive. Is there some workflow I’m missing that’s required to get stupendous results? Are other search needs better met by Kagi that are missed by other search engines? Have others also felt a lack of “wow”?

A clean page of results that the search engine deems best [i.e., no antagonistic commercial interest driving priorities] for the query and the ability to block domains - that was enough for me. ‘Summarize this page’ is useful as well.

Do you find that this actually saves you time, and that you get to your desired search result faster? These sound like great features to me, but I haven’t found them translating to less time spent searching.

Yes

I search a lot. Heavy heavy searching. Kagi's no-BS approach to search gets me to relevant results much faster than any other search engine. And the ability to down-rank or block unreliable sources only sharpens that experience.

I have used, and continue to try on occasion to keep my opinions updated, all the big search engines. DDG, Bing, and Google. The search experience is really not there for me in those 3. Bing consistently produces poor results, DDG is a bit better but I still have to manually sift through poor quality results to get what I want, and Google makes me scroll past a lot of bloat and ads first to get to quality results, with more ads interspersed throughout.

I like being able to search, get my quality results, and move on with my life. That's worth paying for to me.

Same experience here. I used to try Kagi from time to time back before they cut down the free tier and the results seemed fine but not particularly noteworthy. The lack of ads is of course nice, but having all of my searches tracked under a single login gives me a bad feeling, especially with a company with such a manic and AI focused product strategy. They might be trustworthy now, but we've seen so many companies start out ethical and then change course once the money gets tight or they get more greedy.

Semi related +1 for mojeek powering their results. Kagi gets a lot of favour on HN, worth a nod to Mojeek for powering their organic results to an extent.

wow just compare these two results:

https://www.mojeek.com/search?q=baby+peacock&fmt=images

https://www.google.com/search?q=baby+peacock&sclient=img&udm...

hmm, I am not certain that I prefer the option without the source listed. For example, I might prefer an image from Wikipedia over another image. Or if I am searching for a map and the URL has the word 'historical' I would not choose that one. etc.

It depends on what I am searching for and why I want to use it.

If you look closer, Kagi uses Mojeek for organic results, not images

I guess to me anyway, images/news etc is a sideshow to the wider web, as in a crawler/indexer and not a pretendy meta search.

I use Kagi. I love Kagi.

(Please, don’t support (financially, by “donating” your attention) the surveillance, psychological addiction, and manipulation economy.)

And this is a great feature!

I do wonder how long the distinction between AI & non-AI, for any type of content, will be algorithmically or manually detectable.

Another nice Kagi feature that I eventually won't want to do without. I already can't imagine going back.

Should go ahead and add Pinterest and Adobe stock images to the list of "mostly AI" image providers

So that's nice for people that don't want AI images. In the other hand, I actually want the opposite. I want AI images ranked higher. They're more likely to be usable (CC-.., PD, etc)

Would you want to search for them though? Or ability to generate them yourself easily?

searching for the often leads to the prompts I need to get a similar image from which I can iterate from. So yes, I want to search for them

>searching for the often leads to the prompts I need to get a similar image from which I can iterate from.

how crazy are these images that you can't start with a prompt of your own?

at least in the image generators that I've used it takes quite a bit of experimentation to get both positive and negative prompts that produce both the result you want and a quality one as well.

Part of it, if I understand correctly, is that the data has no actual concept of grammar and meaning. So if you say "a table with 6 legs" it doesn't see that as a full concept, it sees words or segments of words, so "legs" can end up being legs of people (it might make a table with human or animal legs) and similarly it just might insert random legs in the scene because it has no understanding of the description.

So, people find ways to coerce / influence it to get to the right place.

It's also important to know what models/loras/etc were used as not all data sets generate good images for whatever your topic is.

Can't we demand (+) that AI generated images get a watermark that designates it as such?

(+) By frowning really hard at people who don't follow this rule?

Because, don't defecate where you eat.

It would be actually useful to mandate AI-generated images to include identifying metadata.

Is there any downside with introducing this demand to legislation? I feel that there's no problem with AI generated imagery being labelled as such, and on the other hand, we have a lot to gain by establishing a reliable signal that media was not generated by "AI". Users can always remove the metadata if they must exercise that freedom. But the vast majority of users won't so it would work in 99% of cases.

>this feature relies on the website's reputation rather than analyzing individual images.

Okay, this would not work for Reddit, where many of the AI-generated images come from, or any other site that allows user-generated content (unless the site is strictly AI-related).

> By default, Kagi Image Search downranks images from websites with a high proportion of AI-generated content.

Is this incentive for sites to avoid/discourage AI-generated images, to avoid that hurting search rankings?

That's definitely the incentive it produces but given Kagi's market share, I imagine it will have pretty much zero effect on sites' behavior. It's much more (and quite valuable) a browser-side improvement.

True, but not only Kagi might find this a good tactic. The idea of many people doing negative reinforcement against generative AI "content" is interesting.

As an aside, i would love “stickied” preferences for Kagi e.g. results from the last year, with forums selected. since i’m mostly doing reddit searches anyway

In the example in the docs, they show someone marking an image that has slipped through the filter as one that's AI generated. But, it comes from Adobe Stock - and I think this really highlights the biggest weakness with how they've gone about implementing this. It's not looking at the images at all, it's looking at where they come from.

The problem is that now that Pandora's box has been opened, all sorts of sites (incl. anywhere like social media that accepts user content) are going to have comingled AI-and-legitimate images that they host.

This is a hard problem to solve.

It is a hard problem to solve and we just started solving it a week ago. There is lot more to be done on this front, but when there is will there is way.

Dunno if vlad can see this, but

https://kagi.com/images?q=baby+peacock

...shows that infamous AI-generated peacock image multiple times on the first row of results.

Merely filtering out websites that tend to have lots of AI images does not prevent this failure case, since (for example):

https://birdfact.com/articles/baby-peacocks

has the fake image in there, as an example of "What does a baby peacock not look like?"

As Emily Bender has correctly pointed out, AI images are like an oil spill, and the cleanup (if such a thing is even feasible) will be challenging:

https://medium.com/@emilymenonbender/cleaning-up-a-baby-peac...

(edited to add that I'm a paying Kagi customer, and this failure case isn't a ding against my overall impression of what Kagi is, and I'll continue using it)

What's that about a fake peacock? This is the first I'm hearing of it.

> What’s that

https://news.ycombinator.com/item?id=41767648

I see that a few AI images get through. That is OK - we still filter almost 30 AI images correctly (scroll to the bottom of search results to see them all). Also overall the results seem to be of higher quality than on other search engines.

For something that we just started working on a week ago and knowing this is just the first iteration of the feature - I think we are doing good overall. When there is will, there will be a way. And there is plenty of will on our end to stop this thing.

This feature may single-handedly make me pay for Kagi rather than using the free trial searches. Looking for drawing references is infuriating these days with the amount of AI.

For Google Images I've had success with filtering by date to exclude everything posted since about 2023.

That said this feature might still be useful, especially if it's extended to normal search results for sites with AI-generated articles etc.

This works wonderfully, but is obviously not sustainable. It's not just that you miss out on newer content, but content rot progresses pretty quickly. Old Reddit accounts are deleted or blocked, Flickr users stop paying their subscription fees, etc.

There are so many photobucket.com URLs buried in old forum posts that no longer work...

Great! Now let's apply this to ads and notifications!

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.

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?

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.

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.

> 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.

Reply on news.ycombinator.com