In this episode, Gabriel (Founder & CEO) introduces our vision for AI and speaks with Rachel (front-end team) about our newly released AI image filter.

Disclaimers: (1) The audio, video (above), and transcript (below) are unedited and may contain minor inaccuracies or transcription errors. (2) This website is operated by Substack. This is their privacy policy.

Show notes: See the full post on our approach to AI — private, useful, and optional.

GabrielHello, hello everybody. Welcome to DuckTales. This is our series where we go behind the scenes at DuckDuckGo. I'm the founder of DuckDuckGo. I have Rachel here who works at our company. Hi, Rachel.

RachelHi!

GabrielSo, DuckTales is really about discussing the technology we use, the people working on it, how we're building privacy tools. We hope you'll join us. You'll hear about all aspects of DuckDuckGo essentially.

Today, we are talking about AI a bit and in particular a feature that we recently released that allows you to filter out AI generated images on our search results. Taking a step back, our approach to AI is, we wrote a whole post on this if you want to check it out, maybe we can figure out how to put things in show notes.

But three things: private, useful, and optional. So private kind of goes without saying, that's what we do at DuckDuckGo. But with AI in particular. You know, we're focused on, you know, Rachel, you know this, so just saying this for everybody else listening. We're focused on keeping it anonymous and also not training on your data. And then useful is we're really trying to make AI tools that are actually useful, not just for the sake of using them, but...so they're actually useful. So on the search engine, you know, we have answers now at the top that kind of help you figure out what you're trying to look for without having to click through a lot of results. Although you're welcome to, of course, and the sources are clearly labeled there. But the third pillar is optional. So not everybody wants to use AI. We realize that.

And so we've been approaching it by making everything optional. And so you can turn off those answers at the top of the search engine. You can turn off our chat product, Duck.ai, which brings us to the AI image filter, which is a little different type of feature because it's not something that we're doing generation of images, although we're probably going to end up doing that sometime. It's more in that optional camp, taking AI away if you really would like to. You want to explain it a little bit?

RachelYeah, so the AI Image Filter, actually let me just share my screen so I can quickly demo it as we talk about it. So the AI Image Filter is a feature you can find on our image filter bar. So up here, this is the filter here. And you can also find this feature on our user settings page. But yeah, it's literally what the name suggests.

A filter that filters out AI-generated images. At a high level, you can put in a query. And if we detect that you've enabled the feature, then our images service is going to do some basic pattern matching against a block list that we have and return the filtered results.

The list itself is actually from an open source project called Ublock Origin. And I did want to take a second to give a shout out to the project maintainer, Laylavish and the 40 plus contributors to the project. I think they started this project like a year ago and they've been doing a great job frequently updating the list. So huge props to them for helping make this feature possible.

GabrielSo great query. I have two cats. Is the filter list that we're currently using for this, is that at a domain level?

RachelDomain, like the file of the image and where that domain or where the file lives in that domain.

GabrielSo it could be, it's kind of like the general UBlock rules. It's like a reg X type of thing, but it could be at the URL pattern, whatever people want.

RachelExactly, URL pattern matching. That's exactly it.

GabrielGot it, and so it works now where if it's on that list and you select this filter, we'll just remove it client side from you.

RachelExactly.

GabrielCool. And have we looked at other lists? I mean, is that like the main list we found? Did we kind of look at other things? Are we looking at other things?

RachelYeah, this was the main list that we had found. There are definitely a handful of other lists that were maybe not as frequently updated. And so I think this one was probably the most reliable one to use.

GabrielGot it. Let's speak to why people might want to use this. It's gotten a, well, first of all, let's say that it's gotten a lot of support when we put this out. It was high on like Reddit technology. Saw all sorts of articles about it and it's being reported in our subreddit a lot of people really appreciate the feature. And kind of to think about why, you know, there's clearly value in AI generated images because you can get things that didn't exist. And so there, I think there's definitely value in seeing AI images, but then there's also value in when you want to use something in real life that it wasn't AI. You want to know that it wasn't AI generated.

RachelRight, yeah. I mean, one of the big reasons why we built it was because we were seeing a huge spike from users about not wanting to see AI. And it's understandable. So many things on the internet now are generated by AI. And it can be really mentally exhausting to have to process. Like, is this AI or is this not? I just want to find the thing that I'm looking for. Yeah, the users were really vocal about just not wanting to deal with it.

GabrielInteresting. And then, so I understand that we're, you know, certainly after the support, it's a feature we're definitely going to keep. And so now we're also trying to improve it if we can. What areas of improvement are you looking towards that we're exploring?

RachelMm-hmm. Yeah, so as soon as we launched, we got immediate feedback from users again on little things that we could do to improve the feature. I think one of the most frequently requested feature was the ability to flag images that they believe were AI generated. So I'm actually going to roll this out right after we finish recording this. So we're going to have that ability for flagging images.

I think once we start getting reports in, there's a lot of different directions we can go from there. If we get enough reports on a specific domain, we'll probably do some sort of vetting process to improve the block list itself. I think we could also potentially allow for users to create their own custom block list, sort of like how we allow users to exclude certain domains for organics.

And I think more like in the long term, we do have people on the team exploring other open source solutions and classifiers and also like the image metadata, like how we can leverage that to better determine if an image is AI generated. But it's still like a matter of figuring out what the right balance is between those options, just because like image metadata, for example, anyone can tamper with the image. They can strip the metadata, can edit it or crop it or whatever. So it's still a pretty fragile source of truth and even the classifiers, like best ones maybe will get 80-90% of accuracy but it's not going to be bulletproof. So yeah, we're trying to figure out what combination is the right one for getting enough accuracy and also keeping cost and latency in mind.

GabrielYeah, so a couple of follow-ups on that. So again, it is important for us to be clear that it's definitely not 100% accurate in either direction. It could let things through that are AI generated, and it can also theoretically flag some things as AI generated that are not AI generated. And so it's an effort here, and we're going to try to improve it, but it probably will never be 100% accurate.

RachelYeah, yeah. No, I mean, it's sort of, yeah, the problem is really interesting. And in some ways, I'm like wondering if it's probably a longer discussion. But like, if we don't have some sort of change at, like, I don't know if this is the right term, but like at a policy or an infrastructure level on how we actually maintain, like we need a consistent way to be able to really know if the data on the image is reliable.

And It's sort of like playing whack-a-mole right now, but if we can contain the mole, then it would be maybe a little bit easier. But right now, if we don't have that, I think the issue will unfortunately become more prevalent and even harder to solve.

GabrielI have heard some rumors at least in some reports of some standards being worked on in that regard. It possibly would help. The other thing you mentioned was, you know, we're using the block list, which is obviously kind of on or off of it. So it's very fast. Whereas if we were