A few weeks ago, I came across a wild post on Reddit\u2019s r\/DHExchange, a subreddit for trading large datasets: \u201cI hoarded a large database of something valuable, just not what\u2019s [sic] you expect\u2026150k stools images.\u201d\u00a0The post, made by a user called Ill_Car_7351, was advertising exactly what it sounds like: A database of poop images, collected from an AI poop analyzing app that he had launched several years ago. Basically, 25,000 people had been taking images of their poop and uploading them to his app. He\u2019d been collecting, analyzing, and annotating these images and now wanted to sell access to them: \u201cI\u2019ve got 150k+ labeled and classified images of \ud83d\udca9 from roughly 25K different people. Jokes aside, I know there\u2019s a lot of value in it (hard to obtain, useful for ML [machine learning] training, cancer studies etc) but not sure on how to move about it. Feels like I\u2019m sitting on a pile of shi..ny coins but can\u2019t find who wants them.\u201d The poster added that \u201cthe images are extremely rare,\u201d and that he was trying to figure out how much money he could sell them for.The comments were from people who were mostly horrified: \u201cWhen I was 5 the teacher taught me how to read. I now regret that happened,\u201d one read. \u201cWhat in the fuck,\u201d another read. \u201cHow to delete someone else\u2019s post,\u201d a third said.\u00a0I messaged the poster and told him I was interested in obtaining the database. Thus began my journey into the Internet of Shit and, by extension, the unpleasant world of the underground sale of highly sensitive, app-collected user data for AI training.\u00a0<\/p>\n
The poop database comes from an app called PoopCheck, an app made by a company called Soft All Things that purports to use AI to analyze images of one\u2019s stool in order to give you a \u201cdaily gut health score.\u201d\u00a0\u201cOur AI analyzes your poop using the Bristol Stool Scale and advanced pattern recognition. Get insights on consistency, color, shape, and what they mean for your digestive health,\u201d the app advertises. The Bristol Stool Scale classifies stools into one of seven types ranging from \u201cseparate hard lumps, like little pebbles\u201d to \u201cwatery with no solid pieces.\u201dThe app also features a \u201ccommunity,\u201d of 151,317 \u201cshared stools\u201d at the time of this writing and a \u201cleaderboard,\u201d where people can share images of their poop for commentary from other users and earn points for participating. I found the posts in the community a bit hard to stomach, with titles \u201clike play dough,\u201d \u201cConcerned,\u201d and \u201cDealing with this on and off for the past 3 weeks.\u201d Pictures are not automatically shared to the community; when you take a photo it asks if you want to share it.\u201cPopular\u201d posts on the app include people speculating as to whether their fellow community members have parasites or colon cancer; in the comments section of a few posts I saw people recommending ivermectin to the original poster.\u00a0Though users have the option to share their poops with other users, the app provides mixed messages about the fact that the data uploaded to the app will be analyzed, annotated, and packaged with other poops into a commercial database to be sold to AI companies.\u00a0On the App Store page for PoopCheck, it says \u201cThe developer does not collect any data from this app.\u201d The link to the privacy policy from within the App Store download page does not mention anything about selling or sharing the data and says \u201cyour health data is encrypted in transit and at rest. Photos are processed securely. We implement industry-standard security measures to protect your data.\u201d\u00a0The PoopCheck website\u2019s About page states \u201cPrivacy First.\u201d And \u201cHealth data is sensitive. That\u2019s why privacy isn\u2019t a feature, it\u2019s our foundation. Your photos are encrypted. You can delete everything at any time. We built PoopCheck the way we\u2019d want our own health apps built.\u201d The FAQ also notes \u201cyour privacy is our priority.\u201dThis is completely different from the \u201cService Agreement\u201d and \u201cTerms and Conditions\u201d people agree to when they actually open the app and make an account. The Service Agreement states that \u201cby uploading stool images or any health-related data to the App, you grant Soft All Things LLC a worldwide, irrevocable, perpetual, unconditional, royalty-free, fully-paid, transferable, sub licensable license to use, reproduce, modify, adapt, distribute, sell, license, and create derivative works from such content for any lawful purpose, including but not limited to research, commercial exploitation, product development, and third party licensing. You acknowledge that your images and data may be used to create, train, improve, and commercialize AI technologies and machine learning models, and that such models and any outputs derived from your data may be licensed or sold to third parties, including medical organizations, research institutions, and commercial partners.\u201dIt adds that \u201cyour data may be irreversibly incorporated into AI models and aggregated datasets. Deletion of your account will remove your personal profile data but does not require the removal of anonymized, aggregated, or derivative data already processed or incorporated into AI models.\u201d Under a section called \u201cSharing of Information,\u201d it adds that the company reserves the right to share or sell the data \u201cfor any business purpose,\u201d including \u201cAI and Data Licensing.\u201d\u00a0<\/p>\n
On Reddit, I messaged Ill_Car_7351 and said \u201cHi – am interested in this database you posted about. Can you share any more info about what you’re looking for \/ details about the app where it was collected? also any chance there’s like, a sample of what the data looks like etc?\u201d They responded quickly and said \u201cHey! The db was gathered by real users, we had 25k users over the last couple years, since we launched the app. It\u2019s called PoopCheck btw if you wanna see it. Let\u2019s maybe talk via email? I\u2019ll be happy to share a sample of the data if that interests you.\u201dI sent an email to someone named \u201cMarco\u201d at Soft All Things, who identified himself as one of the founders of PoopCheck. I said I had reached out on Reddit and was interested in a sample of the data. I used my real email address and real name.\u201cWe can surely send you a sampling of the dataset, would a Google Drive link containing an image folder and JSON data work? We can also figure out other ways if you prefer,\u201d Marco said. \u201cIn terms of the actual dataset you need, what would be the size of it for your needs? And what would you be using it for? Just so we can make sure it\u2019s actually a good fit for your use case.\u201dI told Marco that I wanted 10,000 pieces of data and said I would use it for AI training. I asked him for pricing and what type of data was included.\u00a0Marco responded:\u201cYou’ll find a folder with images and JSON metadata covering the key fields we capture per entry. Let us know if you have any questions about it.To give you a better idea of the dataset and pricing options: we currently have over 150,000 images validated by AI. Around 5,000 of these have also been manually reviewed by a member of our team, who verified the AI output and labeling, making this portion more valuable and priced accordingly. It’s also worth noting that certain types on the Bristol Stool Scale are rarer than others, so availability may vary depending on your specific needs.With that in mind, here there is an estimation of pricing options:\u2022 10,000 unreviewed images (AI-validated) \u2014 $3,000\u2022 5,000 fully human-reviewed & annotated (on top of AI validation) \u2014 $4,000\u2022 5,000 reviewed + 5,000 unreviewed \u2014 $5,000It would be great to have a quick call to take this further as there are a few things about the dataset’s structure and coverage that are easier to walk through live.\u201d\u00a0The sample dataset Marco sent me included 20 images of poop from four specific users (five poops each). Each image was tied to a series of user-reported data points as well as AI analyses of each image. AI-analyzed datapoints included the time the poop was taken, the Bristol Type of each poop, whether it was \u201chealthy\u201d or \u201cunhealthy,\u201d the \u201cshape\u201d and \u201cconsistency,\u201d whether there was blood or mucus in the poop, and the quantity (\u201clarge,\u201d \u201cnormal,\u201d or \u201csmall\u201d), and whether it was \u201cfloating\u201d or not. Each of these data points also had a \u201cconfidence\u201d score for how confident the AI was in its analysis. Each image also had user-reported information, which included the answers to a series of questions including \u201cwhen did you have your last meal,\u201d \u201cany discomfort while pooping? (\u201cHard to pass;\u201d \u201cburning\u201d; \u201csharp pain\u201d etc); \u201cHow long did it take?\u201d \u201cDid it smell stronger than usual?\u201d \u201cCoffee or alcohol in the last 12 hours?\u201d The data also included demographic information, which includes age ranges, sex, height, weight, and sensitivities such as \u201clactose intolerance\u201d or \u201cirritable bowel syndrome.\u201d\u00a0 Each image is tied to a specific user through a field called \u201cexternalIndividualID.\u201d\u00a0Soft All Things is not exactly quiet about the database that it has created. On the Poop Check website, it has a page called \u201cFor Business,\u201d which advertises its database. It sells access to both the \u201cStool Analysis API,\u201d which \u201cturns a stool photo into a structured health report,\u201d as well as the \u201cAnnotated Dataset,\u201d of 140,000+ images to \u201ctrain your own models.\u201d It advertises this as the \u201clargest consumer stool image dataset we know of.\u201dIt maybe should not be terribly surprising that a free app in which you upload images of your poop to a random company would have a business model focused on packaging and selling that data. But this type of data collection\u2014of our literal poop\u2014highlights how almost anything we do on our phones can ultimately end up for sale. The fact that it is advertising this for sale at all indicates that there is an AI goldrush for any and all types of data, even our literal waste.\u00a0Research has shown, over and over again, that de-identified \u201canonymous\u201d data doesn\u2019t necessarily remain anonymous when combined with other datasets. Toward the end of last year, the appliance giant Kohler endured a security shitshow when a researcher showed that its stool-analyzing smart toilet camera was not actually properly encrypting the images that it sent to Kohler. The concern there was that your poop data would be somehow accessed by bad actors. In the case of PoopCheck, anyone can simply buy access.\u00a0\u00a0After I told Marco I was writing an article about PoopCheck and its database, he stopped responding to me and did not answer any of my questions.<\/p>\n","protected":false},"excerpt":{"rendered":"