Latest / Inside DuckDuckGo / Duck Tales: Why DuckDuckGo built a way to filter out AI-generated images (episode 2)
Transcript
- 0:00SPEAKER_01: Hello.
- 0:00SPEAKER_01: Hello, everybody.
- 0:01SPEAKER_01: Welcome to DuckTales.
- 0:03SPEAKER_01: This is our series where we go behind the scenes at DuckDuckGo
- 0:06SPEAKER_01: I'm the founder of DuckDuckGo
- 0:08SPEAKER_01: I have Rachel here who works at our company.
- 0:11SPEAKER_01: Hi, Rachel.
- 0:13SPEAKER_01: Hi.
- 0:14SPEAKER_01: So,
- 0:14SPEAKER_01: you know,
- 0:15SPEAKER_01: DuckTales is really about discussing,
- 0:17SPEAKER_01: you know,
- 0:17SPEAKER_01: the technology we use,
- 0:19SPEAKER_01: kind of the people working on it,
- 0:21SPEAKER_01: how we're building privacy tools.
- 0:24SPEAKER_01: We hope, you know, join us.
- 0:25SPEAKER_01: You'll hear about all aspects of DuckTales essentially.
- 0:29SPEAKER_01: But today we are talking about AI a bit,
- 0:33SPEAKER_01: and in particular,
- 0:34a feature that we recently released that allows you to filter out
- 0:39AI-generated
- 0:40SPEAKER_01: images on our search results.
- 0:41SPEAKER_01: So taking a step back,
- 0:45SPEAKER_01: you know,
- 0:45SPEAKER_01: our approach to AI is,
- 0:47SPEAKER_01: we wrote a whole post on this if you want to check it out,
- 0:49SPEAKER_01: maybe we can figure out how to put things in show notes,
- 0:51SPEAKER_01: but three things,
- 0:53SPEAKER_01: private,
- 0:55SPEAKER_01: useful,
- 0:56SPEAKER_01: and optional.
- 0:57SPEAKER_01: So private,
- 0:58SPEAKER_01: Kind of goes without saying,
- 0:59SPEAKER_01: uh,
- 1:00SPEAKER_01: it's what we do at DuckDuckGo,
- 1:02SPEAKER_01: but we're AI in particular,
- 1:04SPEAKER_01: you know,
- 1:04SPEAKER_01: we're focused on,
- 1:05SPEAKER_01: um,
- 1:07SPEAKER_01: you know,
- 1:07SPEAKER_01: Rachel,
- 1:07SPEAKER_01: you know this,
- 1:08SPEAKER_01: so I'm just saying this for everybody else listening,
- 1:10SPEAKER_01: but we're focused on keeping it anonymous and also,
- 1:16SPEAKER_01: um,
- 1:17SPEAKER_01: you know,
- 1:17SPEAKER_01: not trading on your data.
- 1:19SPEAKER_01: Um,
- 1:20SPEAKER_01: And then useful is we're really trying to make,
- 1:22SPEAKER_01: you know,
- 1:22SPEAKER_01: AI tools that are actually useful,
- 1:24SPEAKER_01: not just for the sake of using them,
- 1:26SPEAKER_01: but,
- 1:27SPEAKER_01: you know,
- 1:27SPEAKER_01: just sake of making them like some other companies,
- 1:29SPEAKER_01: but just so they're actually useful.
- 1:31SPEAKER_01: So on the search engine,
- 1:32SPEAKER_01: you know,
- 1:32we have answers now at the top that kind of help you figure out what you're
- 1:38trying
- 1:39SPEAKER_01: to look for without having to click through a lot of results.
- 1:42Although you're welcome to,
- 1:44of course,
- 1:45and the sources are clearly labeled there.
- 1:48SPEAKER_01: But the third pillar is optional.
- 1:52SPEAKER_01: So not everybody wants to use AI.
- 1:54SPEAKER_01: We realize that.
- 1:55SPEAKER_01: And so we've been approaching it by making everything optional.
- 2:00SPEAKER_01: And so you can turn off those answers at the top of the search engine.
- 2:03SPEAKER_01: You can turn off our chat product,
- 2:04SPEAKER_01: Duck.ai,
- 2:06SPEAKER_01: which brings us to the AI image filter,
- 2:08SPEAKER_01: which is a little different type of feature.
- 2:10SPEAKER_01: because it's not something that we're doing generation of images,
- 2:14SPEAKER_01: although we're probably gonna end up doing that sometime.
- 2:17It's more in that optional camp,
- 2:19taking AI away if you really would like to.
- 2:22SPEAKER_01: You wanna explain it a little bit?
- 2:24SPEAKER_00: Yeah, yeah.
- 2:25SPEAKER_00: So the AI image filter,
- 2:27SPEAKER_00: actually,
- 2:28SPEAKER_00: let me just share my screen so I can quickly demo it as we talk about it.
- 2:34SPEAKER_00: So the AI image filter is a feature you can find on our image filter bar.
- 2:41SPEAKER_00: So up here, this is the filter here.
- 2:43SPEAKER_00: And you can also find this feature on our user settings page.
- 2:47SPEAKER_00: But yeah, it's literally what the name suggests.
- 2:51SPEAKER_00: It's a filter that filters out AI generated images.
- 2:57SPEAKER_00: At a high level,
- 2:59SPEAKER_00: you can put in a query and if we detect that you've enabled the feature,
- 3:06then our images service is going to do some basic pattern matching against a
- 3:12block
- 3:12SPEAKER_00: list that we have.
- 3:15SPEAKER_00: return the filtered results.
- 3:19SPEAKER_00: And the block list itself is actually from an open source project called
- 3:24SPEAKER_00: uBlockOrigin,
- 3:26SPEAKER_00: huge AI in all caps block list.
- 3:29And I did want to take a second to give a shout out to the project maintainer,
- 3:34SPEAKER_00: Lee Lavish and the 40 plus contributors to the project.
- 3:38I think they started this project like a year ago and they've been doing a
- 3:42great
- 3:42SPEAKER_00: job frequently updating
- 3:44SPEAKER_00: the list.
- 3:44SPEAKER_00: It's a huge props to them for helping make this feature possible.
- 3:49SPEAKER_01: So great query.
- 3:50SPEAKER_01: I have two cats.
- 3:53Is the filter list that we're currently using for this,
- 3:56is that at a domain level?
- 3:58SPEAKER_01: So essentially...
- 4:00SPEAKER_00: Domain,
- 4:01likethefileoftheimageandwherethatdomainorwherethefilelivesinthat
- 4:07SPEAKER_00: domain.
- 4:08SPEAKER_00: Yeah.
- 4:09SPEAKER_01: So it could be, it's kind of like the general uBlock rules.
- 4:12It's like a reg X type of thing,
- 4:13but it could be a similar URL pattern,
- 4:16whatever.
- 4:17SPEAKER_00: Exactly.
- 4:17SPEAKER_00: The URL pattern matching.
- 4:18SPEAKER_00: That's exactly it.
- 4:20SPEAKER_01: Got it.
- 4:20And so it works now where if it's on that list and you select this filter,
- 4:26SPEAKER_01: we'll just remove it client side from you.
- 4:29Exactly.
- 4:31SPEAKER_01: Cool.
- 4:31SPEAKER_01: And have we looked at other lists?
- 4:34SPEAKER_01: I mean, is that like the main list we found?
- 4:36SPEAKER_01: Did we kind of look at other things?
- 4:38SPEAKER_00: Yeah, this was the main list that we had found.
- 4:41There are definitely a handful of other lists that were maybe not as
- 4:46frequently updated.
- 4:47SPEAKER_00: And so I think this one was probably the most reliable one to use.
- 4:54SPEAKER_01: Got it.
- 4:55SPEAKER_01: And let's speak to why people might want to use this.
- 4:57SPEAKER_01: So it's gotten a,
- 4:59SPEAKER_01: well,
- 4:59SPEAKER_01: first of all,
- 5:00SPEAKER_01: let's say that it's gotten a lot of support when we put this out.
- 5:04SPEAKER_00: I'm going to stop sharing the screen now.
- 5:08SPEAKER_01: It was high on like Reddit technology.
- 5:10SPEAKER_01: I saw
- 5:11SPEAKER_01: all sorts of articles about it.
- 5:13And it's being reported in our subreddit a lot that people really appreciate
- 5:16the future.
- 5:17SPEAKER_01: And kind of to think about why,
- 5:20SPEAKER_01: you know,
- 5:21SPEAKER_01: there's clearly,
- 5:21there's clearly value in AI generated images because you can get things that
- 5:27didn't
- 5:28SPEAKER_01: exist.
- 5:29SPEAKER_01: And so there,
- 5:30SPEAKER_01: I think there's,
- 5:30SPEAKER_01: there's definitely value in seeing AI images,
- 5:33SPEAKER_01: but then there's also value in
- 5:34SPEAKER_01: when you want to use something in real life that it wasn't,
- 5:39SPEAKER_01: you want to know that it wasn't AI generated.
- 5:41SPEAKER_00: Right.
- 5:42SPEAKER_00: Yeah.
- 5:43SPEAKER_00: I mean,
- 5:43SPEAKER_00: I think,
- 5:43SPEAKER_00: I mean,
- 5:44one of the big reasons why we built it was because we were seeing a huge spike
- 5:50from
- 5:51SPEAKER_00: users about not wanting to see AI and it's understandable.
- 5:55SPEAKER_00: Like so many things on the internet now is generated by AI and
- 6:00it can be really mentally exhausting to have to process,
- 6:04like,
- 6:04is this AI or is this not?
- 6:06SPEAKER_00: Like, I just want to find the thing that I'm looking for.
- 6:08So yeah,
- 6:10the users were really vocal about just not wanting to deal with it.
- 6:14SPEAKER_01: Interesting.
- 6:16SPEAKER_01: And then,
- 6:17SPEAKER_01: so I understand that we're,
- 6:19SPEAKER_01: you know,
- 6:20SPEAKER_01: certainly after the support,
- 6:22SPEAKER_01: it's a feature we're definitely going to keep.
- 6:24SPEAKER_01: And so now we're also trying to improve it if we can.
- 6:28SPEAKER_01: Where,
- 6:29kind of what areas of improvement are you looking towards that we're
- 6:32exploring
- 6:34SPEAKER_00: Yeah,
- 6:35SPEAKER_00: so as soon as we launched,
- 6:37we got immediate feedback from users again on like little things that we
- 6:41coulddo
- 6:42SPEAKER_00: to improve the feature.
- 6:43I think one of the most frequently requested feature was the ability to flag
- 6:51images
- 6:51SPEAKER_00: that they believe were AI generated.
- 6:53SPEAKER_00: So I'm actually gonna roll this out right after we finish recording this.
- 6:58SPEAKER_00: So we're gonna have that ability for flagging images
- 7:03And I think once we start getting reports and there's a lot of different
- 7:07directions
- 7:07SPEAKER_00: we can go from there,
- 7:09SPEAKER_00: like if we get enough reports on a specific domain,
- 7:12we'll probably do some sort of vetting process to improve the block list
- 7:16itself.
- 7:17SPEAKER_00: I think
- 7:18We could also potentially allow for users to create their own custom block
- 7:23list,
- 7:23sort of like how we allow users to exclude certain domain domains for
- 7:27organics.
- 7:30SPEAKER_00: And I think more like in the long term,
- 7:32wedohavepeopleontheteamexploringotheropensourcesolutionsandlike
- 7:37SPEAKER_00: classifiers.
- 7:40SPEAKER_00: and also like the image metadata,
- 7:43like how we can leverage that to better determine if an image is AI generated.
- 7:49Butit'sstilllikeamatteroffiguringoutwhattherightbalanceisbetween
- 7:53SPEAKER_00: those options,
- 7:53SPEAKER_00: just because like,
- 7:55Image metadata,
- 7:56like for example,
- 7:57anyone can tamper with the image,
- 7:59right?
- 8:00SPEAKER_00: They can strip the metadata, they can edit it or crop it or whatever.
- 8:03SPEAKER_00: So it's not like, you know, it's still a pretty fragile source of truth.
- 8:07SPEAKER_00: And even the classifiers, like best ones, maybe we'll get 80, 90%.
- 8:13SPEAKER_00: of accuracy, but it's not going to be bulletproof.
- 8:17SPEAKER_00: So yeah,
- 8:18we're trying to figure out what combination is the right one for getting
- 8:23enough
- 8:23SPEAKER_00: accuracy and also keeping cost and latency in mind.
- 8:29SPEAKER_01: Yeah.
- 8:29SPEAKER_01: So a couple of follow-ups on that.
- 8:31SoIthinkitisimportantforustobeclearthatit'sdefinitelynotahundred
- 8:35SPEAKER_01: percent accurate in either direction.
- 8:37SPEAKER_01: We could let things through that.
- 8:40are AI generated and that can also theoretically flag some things that are
- 8:44agender
- 8:44SPEAKER_01: that are not AI generated.
- 8:46SPEAKER_01: And so that's effort here and we're going to try to improve it,
- 8:49SPEAKER_01: but it's,
- 8:49SPEAKER_01: it probably will never be a hundred percent accurate.
- 8:52SPEAKER_00: No, I, I mean, it's sort of, um, yeah, the problem is really interesting.
- 8:57SPEAKER_00: And in some ways I'm like wondering if it's probably a longer discussion,
- 9:02SPEAKER_00: um,
- 9:02SPEAKER_00: but like if we don't have some sort of change at like,
- 9:06SPEAKER_00: uh,
- 9:07SPEAKER_00: I don't know if this is the right term,
- 9:08but like at a policy or an infrastructure level on how we actually maintain.
- 9:14SPEAKER_00: Right, right, right.
- 9:15Like we need a consistent way to be able to really know if the data on the image
- 9:21is reliable.
- 9:22SPEAKER_00: And it's sort of like playing whack-a-mole right now, but if we can like,
- 9:27SPEAKER_00: contain the mole, then it would be maybe a little bit easier.
- 9:32SPEAKER_00: But right now,
- 9:33SPEAKER_00: if,
- 9:33SPEAKER_00: you know,
- 9:33SPEAKER_00: we don't have that,
- 9:34and I think the issue will unfortunately become more prevalent and even
- 9:39harderto
- 9:39SPEAKER_00: solve.
- 9:40SPEAKER_01: I have heard some rumors,
- 9:42at least in some reports of some standards being worked on in that regard,
- 9:46SPEAKER_01: it possibly would help.
- 9:48SPEAKER_01: The other thing you mentioned was,
- 9:49SPEAKER_01: you know,
- 9:49SPEAKER_01: we're using the block list,
- 9:50SPEAKER_01: which is obviously kind of on or off of it.
- 9:54SPEAKER_01: So it's very fast.
- 9:56SPEAKER_01: Whereas if we were starting to use classifiers,
- 9:59SPEAKER_01: that might be too slow in real time to like do it on the search results.
- 10:05SPEAKER_01: And then it's too expensive to offer.
- 10:09SPEAKER_01: And so I think that's the trade off there with the classifier.
- 10:12SPEAKER_01: It's also interesting to note of using AI to
- 10:16SPEAKER_01: stop seeing AI generated images.
- 10:19SPEAKER_00: The irony there.
- 10:21SPEAKER_01: Yeah, exactly.
- 10:22SPEAKER_01: But I mean, if it works, I think that is an interesting avenue.
- 10:25SPEAKER_01: But yeah, I think it might be, there's a speed trade off there.
- 10:30SPEAKER_00: Yeah,
- 10:30especially because we're dealing with like hundreds of millions of images a
- 10:34day,
- 10:35SPEAKER_00: like even with a good caching layer,
- 10:37SPEAKER_00: I think it'll still be a difficult problem to solve there.
- 10:41SPEAKER_01: Cool.
- 10:41SPEAKER_01: Well, is there anything else from working on this so far?
- 10:46you want to flag anything interesting or,
- 10:49um,
- 10:49I don't know,
- 10:50that we didn't talk about?
- 10:53SPEAKER_00: Um,
- 10:54SPEAKER_00: yeah,
- 10:54SPEAKER_00: I mean,
- 10:54I think the biggest thing is like the problem itself is super interesting for
- 10:58me
- 10:59SPEAKER_00: personally,
- 10:59SPEAKER_00: though,
- 11:00SPEAKER_00: if I could share,
- 11:01SPEAKER_00: um,
- 11:02SPEAKER_00: like when I started this project,
- 11:04SPEAKER_00: uh,
- 11:04SPEAKER_00: I was only four months in at DuckDuckGo.
- 11:07SPEAKER_00: So like,
- 11:07SPEAKER_00: everything was surprising.
- 11:09SPEAKER_00: And then I was still getting to know the product and the users.
- 11:12SPEAKER_00: But it's kind of a funny story how the project even kind of came to be.
- 11:19SPEAKER_00: like we knew this problem was happening.
- 11:24I think like over 50% of user complaints about images was about not wanting to
- 11:29see AI.
- 11:30SPEAKER_00: So we knew we had to do something.
- 11:31SPEAKER_00: It's just that the problem was huge.
- 11:33SPEAKER_00: And then like we talked about it for a while, kind of punted on it.
- 11:37SPEAKER_00: And then summer came around with the hack days.
- 11:41SPEAKER_00: And I kind of decided to revisit the problem
- 11:45And I think the interesting thing about a hackathon is that it's just two
- 11:48days.
- 11:48So like,
- 11:49you're really forced to think like,
- 11:50what am I actually gonna get done here?
- 11:54And coincidentally,
- 11:55at the time I was like looking into a bug that users were reporting.
- 12:00So there were some users who already had this block list and the uBlock
- 12:06browser
- 12:06SPEAKER_00: extension installed and they were dealing with some bugs.
- 12:10And I was looking into,
- 12:10and that's actually how I came across the block list.
- 12:15SPEAKER_00: And I figured,
- 12:16SPEAKER_00: you know,
- 12:16if there are users who are going out of their way to install this extension,
- 12:22SPEAKER_00: put this list on,
- 12:23and this is the solution that's sort of working for them for the time being,
- 12:27SPEAKER_00: like,
- 12:27SPEAKER_00: why not make this more widely and easily accessible for our users?
- 12:32SPEAKER_00: So the solution actually was really inspired by our users.
- 12:36SPEAKER_00: And I really love how...
- 12:38SPEAKER_00: this feature was end-to-end, just purely user-driven.
- 12:41SPEAKER_00: It was built based on user feedback.
- 12:44SPEAKER_00: The solution itself was something that users came up with, essentially.
- 12:49SPEAKER_00: And then even now,
- 12:50SPEAKER_00: after we've launched,
- 12:51we're constantly listening to what users are suggesting to help move the
- 12:56feature
- 12:56SPEAKER_00: further along.
- 12:58SPEAKER_01: That's awesome.
- 12:59Yeah,
- 13:00future for these episodes should definitely ask where this project came
- 13:03from.
- 13:05SPEAKER_01: The list of questions we need to ask.
- 13:08SPEAKER_01: But yeah, it's great that you took the initiative to do it.
- 13:10It's also great to hear that it came out of user feedback in that channel and
- 13:14the
- 13:14SPEAKER_01: hack days are working too.
- 13:17SPEAKER_00: Yeah.
- 13:18SPEAKER_01: Like, yeah, that's a great story all around.
- 13:20SPEAKER_01: So, okay.
- 13:20SPEAKER_01: So thanks for coming on.
- 13:22SPEAKER_01: Whoever is watching this,
- 13:24SPEAKER_01: thanks for joining our episode number two,
- 13:26SPEAKER_01: I guess,
- 13:27SPEAKER_01: of DuckTales,
- 13:27SPEAKER_01: but we'll do many more.