Flock is Tracking People, Not Just Cars

[00:00:00]They pulled her over in this black Dodge Durango that allegedly hit another car and caused three people to die and it had not a scratch on it, but they still arrested her. This woman spent 13 days in jail because she was falsely arrested.

[00:00:16]Hello and welcome to the 404 Media podcast where we bring you unparalleled access to hidden worlds both online and IRL. [music] 404 Media is a journalist founded company and needs your support.

[00:00:27]To subscribe, go to 404media.co [music] as well as bonus content every single week. Subscribers also get access to additional episodes where we respond [music] to the best comments. Gain access to that content. That's 404media.co.

[00:00:42]Also remember to subscribe to our YouTube channel where you can watch all of our episodes including this one.

[00:00:48]Subscribe at youtube.com/404 media co. I'm your host Joseph and with me are the other 404 media co-founders.

[00:00:57]The first being Sam Cole.

[00:00:58]>> Hey, >> Emanuel Mayberg. Hello. And Jason Kebler.

[00:01:02]>> Hi.

[00:01:03]>> All right. So, we are having well, not just one event, two events to celebrate our third anniversary um coming up. Sam, what are these two events? And of course, there will be a link directly to get tickets in the descriptions in the show notes, but what are we talking about, Sam? What are we doing? Yeah, we are throwing down in NYC in September.

[00:01:27]So, we're turning three, [gasps] end of August, early September. So, we are going to throw two events. Uh, one of them is a live taping of the podcast.

[00:01:35]It's going to be talks. We're going to have uh Becky and Matthew G from the abstract and 44 come and speak. So, you get to meet them and hear them talk about their coverage areas, which are like science and data centers and all that kind of good stuff. So that'll be on September 3rd at the Green Space, which is a WMYC event space. It's in like the Chelsea area if you're interested. That'll be September 3rd.

[00:01:59]And then on September 4th, next night, which is a Friday, we're going to go to Thre's Brewing and have a big party. So, uh, come get a drink on us, come hang out. That's just going to be like a more casual hangout to celebrate three years.

[00:02:12]So, subscribers get half off their ticket to For Media Live. If you're a subscriber at the supporter level, you can check your uh membership at formed.co/membership.

[00:02:22]50% off the live taping for the party.

[00:02:25]You get into the party for free if you're a supporter subscriber, but there's various ways you can get into this party. You can pay full price if you want to also. [laughter] Um, but all of that information is at formedia.co/3-year.

[00:02:41]Uh, and you also just go to the website and find it. But yeah, we'll be in the show notes all the info that you need.

[00:02:46]But we we love doing these. They're it's a good chance to like meet you guys in person and to see what you're what's going on with you and uh what you're interested in and and get some good facetime. Even though we're constantly in cyerspace, we do miss seeing you in real life. And we'll all be there. The whole gang will be there, which is a very rare treat. We're never all in the same city. So >> yeah, we're usually only in the same place ever for these parties. Basically, um, let me just stress, you said, come have a drink on us. You can probably have more than one. It's an open bar for several hours.

[00:03:21]>> Yeah, please have more than one drink.

[00:03:23][laughter] >> Yes, it's >> at a brewery. Drink the beer, please.

[00:03:27]>> Yeah, it's an open bar party for paid subscribers of 404 media who get in for free. I don't know like how much more attractive you can make it. And then if you're not a subscriber, you can also just pay for the ticket and you get access to uh an open bar. But yes, link to more details below. You should probably get tickets as soon as you can because I mean I imagine especially as we get close to it, they're probably going to fill up.

[00:03:51]But yes, thank you Sam and we'll uh we'll definitely keep keep people updated on that. For this first story, Jason, do you want to take us through it? Yeah. So, uh we're continuing to talk about Flock, uh which over the last few months has really, um had quite a lot of opposition to it. Uh we we're starting to see a lot of um people like cut down flock cameras. I guess there there's like a lot of discussion about what flock cameras can and can't do, which is what this story is about. Um so, this story is called How Cops Use Flock to Track People, Not Cars. Um, Joseph, most people think of Flock as license plate reader cameras, uh, but they do more than that. So, what else do they do?

[00:04:36]>> Yeah. So, I mean, they absolutely do track people as well. And it's not so much specific people like there isn't facial recognition. And Flock will always stress that. There's no facial recognition on our cameras and that sort of thing. Um, you can't type in a name and bring up somebody, but flock cameras that police have access to can be searched as like a search engine. I think that's how you pull it um in your edits to my article, Jason, which I think is a really good way to pull it.

[00:05:06]So they can just type in something like male with tattoos, um, Caucasian woman with Ugg boots or something like that or any sort of permutation or combination of these words and they can instantaneously search a bunch of these cameras sometimes which stretches, you know, into the hundreds and hundreds of cameras at once to look for a specific person. Obviously, even if they're not typing in a name, they're clearly still looking for a specific person because they have this description, right? And I just think that's something that's been really really overlooked in the flock um debate and the flock conversation. And I feel like there's enough evidence to write about it now and like enough dialogue to bring it up as well. But yeah, they are constantly tracking not just license plates that go by, but all of the people uh as well. And it does depend on the specific sort of camera and we'll get into that.

[00:06:06]>> Yeah. Uh I mean these are called free form searches. What is a free form search?

[00:06:13]>> Yeah. So a a normal search for flock is like when you just enter a license plate, right? There's a little dialogue box where the cop can enter a license plate and that's it. Free Form is a product that Flock released I think in early 2025.

[00:06:28]Uh and I really paid attention to when it was launched because it sounded very very interesting in that Flock was saying you can now type in natural language searches. Obviously this is kind of getting on like the AI hype as well right where oh I can ask the LLM to do this I can ask chat GPT to do this.

[00:06:48]Well, I can ask Flock's system to surface footage that may contain people that I'm interested in. So, at the time when Flock announced free form search as in like free form language searches, it was focused on vehicles. So, you could type in stuff like red Ford pickup with a green sticker on it or something like that. something that is descriptive of the vehicle but wouldn't fit into a license plate sort of search box and this was how they were advertising it.

[00:07:22]Um but as this piece shows and as we'll go into in some of the examples like it's not just looking for vehicles you can use those free form searches to look for basically anything. and I just happen to focus on people because obviously that's probably the most important part when you're talking about a uh sort of privacy uh invading technology.

[00:07:43]>> How did you go about finding these? Um like how did you obtain these searches?

[00:07:48]Was this different than the process of figuring out how flock is being used otherwise? Um, so when free form searches were first announced again in 2025, I started doing a bunch of public records requests to various law enforcement agencies who I suspected might soon be using free fall or might already I can't remember exactly how, but often when a surveillance company will come out of a new product, they will include a testimonial from a specific law enforcement agency. Maybe they were testing it out. Like I've done this with Axon before when they introduced um draft one which uses AI to provide summaries of body cam footage.

[00:08:28]So I did a bunch of foyers against law enforcement agencies that said they were demoing the technology. I did the same here with Flock and Free Form and I asked specifically I want your records uh and databases related to free form searches. Um for whatever reason those were denied at the time. So I kind of put it to the side. And I thought, well, I'll come to this later then where maybe it's more widespread and maybe there are softer targets to send um public records requests to and then I'll get like some of these free form searches. Then um I was looking around have I been flocked.com which is a very very interesting website which basically aggregates the results of a bunch of other people's public records requests against flock for the license plate searches that we've done to show local cops who are looking up cameras for ICE and tracked a woman who was seeking uh who sought an abortion and you know all of the data that's contributed to this nationwide conversation on flock have I flock aggregates and maybe not centralizes but puts as much as it can in one place. I was going through there and very handily have I been flocked already has a section for free form searches because they're actually just included in the ordinary network audit as well. When we've sent these requests and when researchers who have acted as sources for us, they've done these requests as well. You ask the police force. I want the network audit of your flock searches. And they'll bring up here's all the license plates that were searched. This is when they were done.

[00:10:05]This is how many cameras were queried either in the town or potentially across the country. And this was the reason why. It turns out actually very conveniently for us, the free form searches are just in there as well. I think it's the column text underscore prompt. So, um, thanks to Have I Been Flocked, I was able to just start going through there, through those records and go, "Oh, okay, there's already a bunch of free form searches out here." And now I can start seeing what sort of people searches police are doing through flock cameras and not just the license plates.

[00:10:37]>> So, what were some of the searches that you found in there? As you mentioned, often when Flock is doing a search, or rather when a police is is searching Flock, they'll just type in a license plate, but uh with this, they can get creative. It's essentially like using Flock as a Google search, more or less.

[00:10:55]So, uh like what what are cops typing in?

[00:10:59]>> Yeah, [snorts] there's kind of like two brackets of searches. either they are quite specific where it will be somebody searching for um for example Atlanta Falcons referencing the NFL team. So presumably looking for like a jersey or a shirt with that. Um maybe it'll be so cop searching for man wearing a black t-shirt and shorts which I guess isn't super specific but it still has some characteristics there. um American flag shirt, Dodger shirt came up as well.

[00:11:33]They're sort of more specific um search terms, right? And then you have ones which are just incredibly vague like tall man. [laughter] Please search all of the cameras for tall man. And look, I mean, I'm sure authorities would say that, well, we have this piece of information that this missing person or the subject of an investigation that we're trying to track down quickly was a tall man. And that's all the information we have. Maybe we'll do that and then maybe we'll get more um solid information after that. There's a lot of tall men presumably walking past cameras. So as some of the experts I spoke to, including um Barl Lipton at the uh Electronic Frontier Foundation, these are probably going to sweep up innocent people as well, arguably even more so than the license plate stuff, right? I mean, Jason, you you've been tracking this closely. Like there's been a lot of mistakes recently, right?

[00:12:29]>> Yeah. Yeah. So, I wanted to bring up one. This is actually it's a free form search, but it's a free form search for car. So, like, as you mentioned, this can be used to track people, and we'll get back to that in a second. But, uh, it can also be used to track cars when the police are not sure of a license plate. So, we've seen searches that are just like black SUV in Florida, like things like this, like very vague terms that are going to necessarily catch a bunch of different cars. Um, rather than just like a specific license plate, you can type in like red Acura, uh, and it will it will pull up all of the examples that they have of something like that.

[00:13:12]And so [clears throat] in Florida, uh, this happened in October of 2025. There was a hit-and-run collision in which a black, uh, Dodge Durango hit another car and three people died. And the driver of that car drove off. And so the Florida Highway Patrol put in black Dodge Durango because some witness said, "Oh, it was like a black Dodge Durango."

[00:13:41]and they pulled over this woman named Lindsay Issacs who was driving a black Dodge Durango and they arrested her and they put her in jail for 13 days. Uh but she wasn't it was the wrong car. It was the wrong person. Like very interestingly and and kind of very insanely. Uh they pulled her over in this black Dodge Durango that allegedly, you know, hit another car and caused three people to die and had not a scratch on it. Uh but they still arrested her and it turned out that a different black Dodge Durango had hit the the family and and killed them. And so this was just like this woman spent 13 days in jail because she was falsely arrested. And so I mean this is like an example of a free form search gone wrong. Like they they can do that for cars and then as Joseph has been talking about they can also put in people descriptors in there. And so we've seen both types of searches where they are searching like a vague term for a car uh or they're searching for as Joseph said tool man like protesters like I I've seen searches for signs uh like people holding white signs is like a a search that we have seen um and so these are like the types of things that uh can can go in there. I mean, just because you mentioned like political stuff as well.

[00:15:05]Um, I contacted the Center for Democracy and Technology, you know, this good organization who have their finger on these issues. So, they're very um reliable when you're like trying to work on a story quickly, like please, I need expert comment on this. and I pinged them about it and they were already looking into something uh similar cuz they sent me a link on have I been flocked to some of the results. And one was very explicit in nature where it said um one police department searched 198 networks for quote white jeep with Trump flag end quote. Now I don't know the context like beyond that but it shows that clearly they can look for a vehicle with a certain political symbol on it a political flag on it like no matter what it might be right so that was very very that again that was a vehicle search like the one you're talking about Jason um but then the people searches do show you can like oh you can look for something of a certain branded t-shirt maybe a sports jersey some something like that so I I know these searches can get pretty narrow um when you're not typing in like black SUV and arresting the wrong person.

[00:16:17]>> Yeah. And and just on that like like I said um of course these cameras are networkkedorked together as we've reported many many times as people know now. But sometimes like the search is both vague and they're searching a lot of different networks meaning a lot of different cities or like across an entire state like this. And so if you're doing a vague search and you're searching, you know, tens of thousands of cameras, it's going to pull up people who are who have nothing to do with the incident that's being investigated. Joseph, does this mean that like every flaw camera is also tracking people or what did you see in these free form searches? Yeah, that that that's probably like the one bit of nuance that people really should take away from it because yes, people should understand that flock cameras are also searching for people and cops are using that based on the searches we've seen, but it's not every single flock camera, right? So, ordinarily when you think of a flop camera and you see it on like a lamp post or whatever or an intersection or or wherever you're seeing it, that's probably a license plate reader just in that I think there are more of those cameras, right? These ones are well one type they can be are these condor cameras which they are explicitly for people tracking uh where somebody will walk past them and the camera will automatically pan and zoom on that person because that way that's what it's looking for. These are not strictly the same as the license plate ones. I think that you can also see this in some of the data where when you look up a search on have I been flocked or if you have the original spreadsheets yourself, it will also say how many cameras were searched and how many networks were searched. When we look at license plate data, when we get that and we've uh I think we've had untold number of those results at this point, I can't even keep track. It is like often like 88,000 cameras or something like that or thousands of networks. And when you see that I feel like you and I Jason and you especially Jason have got a feel for just like oh that's a nationwide search because like it's around 80k 80,000 something like that. So that shows they searched basically every camera on the national flock network for a certain license plate. That doesn't seem to happen with these free form searches. is you'll look at them and like that white jeep with a Trump flag one that was 198 networks still a lot that's probably not the entire country right or even some of them are just like 100 cameras which are probably Condor or other people tracking cameras installed in that single town or as you say Jason maybe in that single state something like that so in the data it doesn't say the type of camera you kind of just have to extrapolate that from looking at this data so much over the past like more than a year at this point. So, I would just stress that yes, um plot cameras are tracking people as we show, but it's not necessarily the same one that's on your street corner.

[00:19:25]It's kind of hard to tell because they look a little bit different uh to the the Condor ones, but we know they track people cuz throwing back to an earlier story, Flock left a bunch exposed, right, Jason?

[00:19:37]>> Yeah, they left a lot exposed. Um, like basically we were able to find the uh I don't know, it was like a testing situation where like a a testing server that Flock was leaving these uh Condor cameras open on. And we did a story back in December on this uh that you should check out. We also did a video with Ben Jordan on this uh who's a YouTuber who's done a lot of work on Flock. And basically like you know you could see how the cameras work where the cameras would follow uh people like as they were walking by. Um I drove up a few hours from where I live and in front of some of these exposed cameras and track myself. I was able to see myself as I was walking around like things like this. And so uh you know Flock has been really marketing the Condor cameras as extremely powerful um like very high resolution. They're called PTZ cameras.

[00:20:30]pan, tilt, zoom, so they can uh like basically zoom in on a situation and on people in any given scene. Um, which is like quite concerning. I think the other thing, and I only learned this the other day, I just missed it when it happened.

[00:20:46]I don't know if you saw this or not, Joseph, but [clears throat] uh condors are like the main video cameras that Flock has been uh promoting. And for a while, uh they were saying that the uh license plate readers were were just basically like taking static images of uh of cars as they pass by.

[00:21:10]But last year, uh, Flock added video capabilities to all of its ALPR cameras as well. Um, and so these cameras are and can film video as well. Uh, they're not as like high resolution as the Condor cameras, and they're not, um, they're basically like not as powerful as the Condor cameras. are not really intended to look for humans in a scene, but they are filming video. And uh Flock went and sort of told all of its customers that they're pushing this as like a free upgrade to all of the um all the license plate reader cameras out there, uh last year. And so, uh I would just check that out, um if you if that's something you didn't know. Um, I'm not sure if the if they can search people on these, like based on sort of the information that you found, um, like it seems like the number of cameras that were involved in any given search suggested that they're probably Condor cameras cuz there's not as many Condor cameras out there. And a lot of these searches were not as wide as um as we've seen with the ALPR cameras where there's sometimes like tens of thousands of cameras in in any given search. But um I think that's something that uh people should know that these can film video.

[00:22:31]And in the um like promo materials, they basically say like you get the license plate uh from our reader, but then you can go back and and check the video to see like which way a car turned like at an intersection, for example. [snorts] >> Yeah, I didn't know that. Yeah, I didn't know that. And yeah, it's [laughter] it's it's just very funny. And this isn't even a flock specific thing, although it very much is uh here, but just like, oh yeah, we're not doing this thing. Surprise, we're doing the thing now actually. And pushing it as a free update is very kind to its customers, but obviously that does set the groundwork potentially for making those into more people search as well if they're high enough resolution, I guess.

[00:23:16]Um, all right, we'll leave that there.

[00:23:18]Well, when we come back after the break, we're going to talk about Emanuel's new story about how AI companies are buying a ton of books, uh, it looks like from a particular company. We'll be right back after this.

[00:23:32][music] There aren't so many things you use every single day. maybe your phone, your wallet, your keys, and then if you wear glasses, well, you wear those everyday.

[00:23:51]So, I've always thought that it makes sense to actually like the pair you're wearing. That's one of the reasons I've been happy with Warby Parker. Warby Parker has made buying glasses feel a lot less intimidating. Their virtual tryon feature lets you use your phone to see what different frames actually look like on your face before you order, which is way better than trying to imagine it from a product photo.

[00:24:12]These are the ones that I ended up with.

[00:24:14]They are the Elio, I believe. Yes, they're they're the Elio in these olive green color. Uh it's actually like a really pretty color. Uh I quite like them. They're a little translucent. Uh and they've become my everyday pair.

[00:24:28]I've worn these pretty much every day out in the LA sun. They're lightweight, comfortable, and they look just as good at the end of the day as they did when I first put them on. Another thing I really like is that Worby Parker keeps everything really simple. Prescription glasses start at $95. And they also offer sunglasses, contacts, online eye exams, plus more than 300 retail stores if you'd rather shop in person. And if you have vision insurance, they'll even help verify your benefits and apply them automatically. I've been wearing these LEOs every day this summer. Uh I don't even think about them anymore. They go with almost every outfit that I wear.

[00:25:02]They're great for the beach. They're great for walking my dog. So, right now, buy one pair of glasses and get 20% off any additional pairs at warbyarker.com/404.

[00:25:13]That's 20% off any additional pairs when you purchase one pair at wbbyarker.com/44.

[00:25:22]We have at this point written an embarrassing number of stories about data brokers, the companies that quietly compile your name, your home address, your phone number, your relatives, your daily habits, and then sell that file to basically anyone with a credit card.

[00:25:34]advertisers for sure, but also spammers, scammers, and occasionally the person you'd really prefer didn't have your address. You never opted in. There's hundreds of these brokers, and opting out of them one by one is effectively a part-time job. That's what Incogn does for you. It acts as your personal privacy advocate, identifying the brokers holding your information and handling the legal leg work of demanding that they delete it. Incogn follow-up process on a recurring schedule. Because this data has a notorious way of reappearing. You set it up once and watch the verified removals stack up on the dashboard. The practical result is fewer spam calls, less junk mail, and a much smaller footprint for anyone trying to find or impersonate you. For the most comprehensive protection, we recommend their unlimited plan, which includes custom removals. As journalists, we know that sensitive data doesn't just sit on random broker databases. It winds up on public forums, directories, blogs, and news portals. If you find your personal details exposed on a specific site, you simply send incognity experts will step in to manually handle the corporate take down request for you.

[00:26:45]I've done these before and uh it's really a huge pain to do it manually.

[00:26:49]So, the fact that incogn does this is amazing. Take control of your data footprint risk-free with a 30-day money back guarantee. Go to incogn.com/44 media and use code 404media for 60% off an annual plan. That's code4 media at incogn.com/404 media.

[00:27:09]Here's a scenario that's not a hypothetical anymore. Your mom gets a call. It sounds like you says you're in trouble. Need money? Don't tell anyone.

[00:27:17]That voice is a generative AI model built from clips that were already out there. Digital fraud is the fastest growing crime in the country. You and I can probably spot most of it, but even we're not untouchable. Now, imagine the people in your life who aren't steeped in AI and technology every day. They're the target. You may have heard us mention Savvi last week. It's a new app that just launched. Savvi protects your whole family from AI powered scams across calls, texts, and voicemail.

[00:27:45]Going beyond what's built into your phone, it filters scam texts before they land, screens unknown callers through voicemail, and can join any live call silently, and flag it in real time if something feels off. 404media listeners can try Savvi free for 14 days. Go to savviescurity.com/404media.

[00:28:02]That's savvcurity.com/404media.

[00:28:06]Plans start at 525 a month for your whole family. Available on iOS and Android. [music] All right, and we are back. Uh, Emanuel, I hope I included the correct headline.

[00:28:24]If not, this this is the one we're going with in the podcast, and I'll just update in the show notes. AI companies are buying tons of old books because they're free of AI slop. So, um, a book seller you're speaking to, Emanuel, noticed something happening in April, um, of this year. What did they see exactly?

[00:28:46]>> Yeah, I guess before I get into that, have you realized that this is a business that operates on this level?

[00:28:53]Like obviously, you know, you buy books from Amazon, you buy books from bookstores, but um I didn't realize that there are people who make a living trading books at scale.

[00:29:05]So, they'll go to a library that is shutting down or a school that is shutting down, uh maybe a bookstore that's going out of business, and they just buy up the entire stock, then put [clears throat] it up for sale on one of these marketplaces and sit in it for maybe years until somebody shows up and and wants to buy it. But um that is this guy's business and he's been doing it for a few decades and uh it's not uh the worst business.

[00:29:33]It's not uh a high volume business either. He told me that um on a good week maybe he'll sell 20 books and that has been the case for years and years.

[00:29:46]And then suddenly starting in April, he started to get orders for hundreds of books a day. So he will um uh check on his uh store on various marketplaces and see that he got an order for 300 books, 700 books, 200 books. Um and that is very unusual. It is uh unusual because that is a historic high and then also the trend has been going in the other direction. He said because libraries and schools have been getting less funding, not more funding. Um, and when he went online to see if other book sellers were experiencing the same thing, they have these message boards where they talk about their business and yeah, everybody has has said the same thing. They've seen historic spikes in sales.

[00:30:36]>> Gotcha. So, what does he suspect around these sales? Like obviously as the headline suggests we're going to get to AI companies but like why does he suspect this is like an AI company potentially or maybe multiple AI companies behind this spike? What's the suspicion there? So first it is a volume right as I said it just nobody historically has bought this many books.

[00:31:04]Then also in the past when people buy a bunch of books at the same time, they usually follow a theme. So imagine you're a library and you're trying to build out a section on whatever the the Spanish Inquisition, right? Then you'll buy a bunch of books about that subject.

[00:31:25]And here there's no rhyme or reason to the books that are being bought. Uh, the only thing they have in common is that they have ISBN's, which is the serial number and barcode on the back of probably every every book that you've ever read. Um, and those are the only books that are being bought in these huge batches. And this seller as well as other sellers um he specializes in a certain foreign language uh and also particular subject uh which includes a lot of rare books and a lot of these don't have ISBNs and none of those books are being bought. So imagine if you're an AI company and you're trying to methodically scan maybe every book in the world, then the easiest way to do that is to just go through a database of ISBN's and go through the list and scan every book. H >> how's he feel about that? How's he feel about the idea that he might be selling um a bunch of books not to a library that's gonna serve, you know, maybe kids or adults who are interested in something or even just, you know, a school or whatever. How does he feel about potentially get selling books to an AI company?

[00:32:47]>> It's complicated. Uh he uh it's obviously very good for his business. He says that usually uh the storage space where he keeps his inventory is overflowing and now there are empty shelves. So it's very good for business.

[00:33:05]He's making money. Other people online are talking about and said that they're making money and not complaining. But uh as someone who's in this business because they love books, uh there is a likely possibility that the books that he is selling are getting destroyed because the cheapest, fastest way to scan books is not to take a picture of a page, turn the page, take another picture. It is to cut the spine of the book and then feed the pages to a machine. uh that can scan them very quickly. So you're creating a digital copy of the book that is maybe being used for AI training data, but the physical book that is maybe rare uh is now more rare because the copy that he sold is no longer in circ circulation.

[00:33:57]It's cut and pulped and gone forever.

[00:34:00]>> Yeah. I I guess this is probably the best place to mention what happened with Anthropic before and like the lawsuit there and what came out then about them scanning books like what happened with Anthropic.

[00:34:12]So we know that AI companies are very hungry for books as AI training data.

[00:34:19]And we know this because there has been a few stories about them using essentially huge collections of pirated books. Um Facebook was caught doing this I believe and Anthropic was caught doing this. And the reason we know about the physical book operation is a bunch of authors sue sued anthropic uh because they use pirated books and training data. And through that lawsuit uh and discovery, we found out that they had something called Project Panama, which was an a mission to scan millions of books and to u yeah destroy them in the process and this has been verified by um Anthropic itself that that they have been doing this.

[00:35:21]>> Yeah. So, I mean, this is hinted at in in in the headline I read out, but like why why would AI companies want to scan a bunch of books on top of scraping basically the entire uh web uh as well?

[00:35:38]Like why are they hunting out these books and and scanning them like in in this bulk way? Like why do that?

[00:35:44]>> So, there are two reasons. One is it's just another source of data.

[00:35:49]Obviously, scraping the entire internet doesn't sound that easy and I'm sure it's not, but it's basically been done, right? Like the egg companies have taken everything that's online and use it for training data and now they're hungry for more training data because that can make the models better. And they're out of like digital material. So you have to go where you haven't been which is physical material textbooks, printed books, fiction books, anything that you can get your hands on. Um so so that's the first reason. And then the second reason which is made explicit by one of the companies that is now servicing this effort for AI companies to get printed books is that uh we have talked before about uh this recursive problem with training AI models on AI generated data which is called model collapse. If you take AI output and use that to train another AI model and you keep doing that cycle after cycle, the AI model starts to get worse and more prone to errors and and just generally not as good as if it was trained on human created content. And if you're scanning books from 2022 or or earlier, you're guaranteed to be training on human created content because before 2022, the large language models weren't readily available. People weren't using them to flood Amazon with AI generated books. uh and it's just sort of a guaranteed high quality uh and most importantly uh a source that is free of AI generated slop that these very AI companies are basically uh producing.

[00:37:40]>> Yeah, the the key date being November 30th, 2022, the date that chat GPT was released. So, you're getting a book before that. It's probably not somebody trying to grift the Amazon marketplace by mass-producing some AI slot book.

[00:37:55]Yeah, that that totally makes sense. So, this brings us, I think, to the most interesting part of the piece, which is that like you found a company that is explicitly facilitating the mass sale of books to AI companies.

[00:38:11]And like, it's kind of crazy how they just say this um all out loud. So what is ISBNDB and and like what are they offering?

[00:38:21]What do they say on their website?

[00:38:23]>> ISBNDB has been around for many years.

[00:38:26]Um it is sort of like IMDb which is something maybe more people are familiar with. It's just a database of all the books that have ISBNs.

[00:38:40]This has been previously very useful for libraries and schools and retailers who are looking to buy certain books or find out what's in their inventory and organize their inventory and things like this.

[00:38:54]But since the generative AI boom, they've discovered a new and deeped new client type which is a companies and AI labs that are trying to to get these books. So now they won't only provide you with this database, they will also go out and source these books for you. So they'll find them, buy them uh with with the AI company's money and uh collect them all in one spot and then deliver them uh directly to wherever the AI company does the scanning. Uh there's a few companies that do this. ISBNDB is one of them.

[00:39:37]Uh there are other companies that have done this previously for uh schools or libraries or retailers, but now they are directly offering their services to AI companies that are making these huge huge purchases and for them the the money is is no object because as we know they they have billions and billions of dollars in investment. And probably most importantly um for these AI companies especially after the the anthropic stuff um ISBNDB says it can offer these companies secrecy right like what does it say about like the need to keep this on the down low.

[00:40:19]>> Yeah. So they they talk quite a bit about the optics the the optics problem of doing this which is nobody wants uh a headline about their company that says open AI or whatever uh has bought a million books using them for training data and then destroyed them. That just doesn't look good. Uh, I don't know what what what image comes to mind for you, but it's like it's a subject of science fiction and uh Fahrenheit uh 451 or like book burning and and stuff like that. It's just like destroying printed books is like not a good uh prediction to see.

[00:41:03]>> Yeah. Not a good vibe. Uh so part of their pitch is we'll keep you completely anonymous. This is kind of standard I've discovered for the industry, right? So it's like if you're on one of these marketplaces where you're selling books, if somebody makes a purchase, you don't know who it is, right? You're selling it via the marketplace, it goes to them as a middleman and then they ship it to the client and the client's identity uh uh stays hidden. So they make clear that that is there to protect them. Then they also say that they make people sign NDAs. Uh, I assume that this is normal as well, but basically it's like don't worry, nobody's going to find out uh that you're doing this with us.

[00:41:48]>> That is odd though because like uh on eBay for example, as a seller, like you know who you're selling to, you get a shipping label. You get you there's like a username associated with who you buy who buys your product. And like maybe that username isn't identifiable. Like it can just be a random username, but you get a shipping label and you know exactly where you're sending it and there's like not a middleman. So it is interesting that just like I wouldn't consider buying and selling books to be like very sensitive like under normal times, you know. Yeah. Yeah. I I believe this is me speculating, but I believe that the reason that this became standard is not originally uh for privacy's sake. I think it was because purchases are made in the dozens, hundreds, now probably for the AI companies in the thousands of books.

[00:42:42]And part of the service that these platforms offer is you don't send it directly to the client, you send it to them. They go through the books, make sure that they meet whatever standard you have, then they repackage them and send them to you, right? So, there's like a quality assurance aspect uh to this because you're buying such large numbers of books, but now it has and then therefore you like you send it to an address that's not the final address.

[00:43:13]Um, but now but now it has the added benefit of keeping clients anonymous.

[00:43:17]>> Yeah. And I think just lastly that like this is some of the first English language reporting on this, right? But there's like rumblings of other stuff. I think I saw some Spanish press and then I think you throw to a piece out of the the Netherlands. Like just what what are other countries seeing? Is it like much of the same thing or like they suspect the same thing or what's going on?

[00:43:39]It's the same thing. And I can't prove this, but I have a hunch that um the reason that we've seen rumblings about this in uh Spanish language media and now uh the story is from the Netherlands and uh that story says that they talk to a few rare book traders in Europe and I think that's because the AI companies are going for the rare books first, right? Like imagine if there is a race to scan all the books, all the printed books in the world. It would be easier to get like a printed whatever Harry Potter. Like obviously there's a digital copy of that as well that's already in the training data, but um like a popular book would be easier to get uh than a rare book. So you'll buy the rare foreign language books first, put put those into the database, um destroy them, and then you might have like a leg up on the competition.

[00:44:36]Yeah. Uh, if you are the one scanning and destroying it, then the other company can't scan it because you've [ __ ] burned that book, obviously.

[00:44:44]Yeah. Um, all right. We will leave that there. If you're listening to the free version of the podcast, I'll now play us out. But if you are a paying 404 media subscriber, we're going to talk about the big uh hack and everything we learned about that. We were the first to reveal that. the hacker contacted us and then we published all of these details um about Sunno. You can subscribe and gain access [music] to that content at 404media.co.

[00:45:12]As a reminder, [music] 404 media is journalist founded and supported by subscribers. If you do wish [music] to subscribe to 404 media and directly support our work, please go to 404media.co.

[00:45:24]You'll get unlimited access to our articles [music] and an adree version of this podcast. You also get to listen to the subscribers only section where we talk about a bonus story each [music] week. This podcast is produced by Alyssa Midcaf. Another way to support us is by leaving a fivestar rating [music] and review for the podcast. That stuff really does help us out. I have a few more of those to read out which I'll do soon. This has been 404 Media. [music] We'll see you again next week.