I have one porn detection action on tumblr and i have a bad oneBackchannelbusinessculturegearideassciencesafetypodcasts video artificial intelligenceclimategamesnewsmagazineeventswired insiderjobscouponslouise matsakis Tumblr's pornography-detecting ai does one thing, but does a poor job of suppressing it Facebook E-mail email application Content moderation Sector Social networks initial data Images Technology Machine vision What does it say about what a patent order drawing for socks with a troll, a cartoon scorpion in a hard hat, porn ai generators and a cat parkour comic have in common? It is reported that any spokes purchased tumblr-tagged this week as soon as the microblogging platform announced that it better not threaten to allow "erotic content material." But so far, tumblr's method for finding posts that violate the new policy, which goes into effect dec. 17, hasn't worked quite well, to say the least, sultry citizens on twitter who have shared screenshots of innocent tumblr posts that have been flagged in the wrong way. Like nsfw. The ad was met with convulsions in the tumblr scene, which has been a long-running bastion of diy and unpopular porn. But the policy change appears to have an even broader effect than expected. News is flagged that seems to go further than tumblr's definition of content for large, which typically includes photos, clips or gifs showing real human genitalia or female nipples, and all information, including photos, videos, gif-movies and illustrations depicting intercourse. (Visitors can contact a human moderator if they believe their posts have been mislabeled as adult content and nothing can be censored until the new policy goes into effect at the end of this month.) "I" i confess i was naive: when i saw the announcement of some kind of ban on the file of erotic content, "i did not think at all about what this would affect my blogs," says sara burshtein, professor of law college of the university of oklahoma, who noticed most of her posts. Have been noted. “I just write about design patents, not erotica.” Tumblr has acknowledged in its own blog a notice about the newest rules that become bugs, under what circumstances it will start using them. “Filtering this kind of content, say a political protest with nudity or a statue of david, within a framework is not easy,” wrote new tumblr ceo jeff d’onofrio. This is the moment when a social media site mistakenly labels wallpapers with current pg ratings as sexy. In 2018, for example, facebook recklessly banned a woman from broadcasting an advertisement with a statue almost 30 thousand rubles old, due to the fact that the porn bunny contained nudity. But not like facebook mistakes , many of the tumblr errors involve posts where there is nothing that even remotely resembles a naked person. Reportedly, and this is the first time the site has flagged a notice in the platform about wrist rests for a person with a type of connective tissue disease. Computers are now usually good at identifying what is in a photograph. But what does it give? Despite the fact that now the possibilities of machine learning have improved significantly, computers do not understand” images because of the population. They show whether clusters of pixels are similar to the kind of nuance that movies have seen before. Tumblr's automatic content moderation mechanism can detect patterns that the factory doesn't represent you can't explain. “Machine learning is great at suppressing relationships in raw data, but a common technique is for algorithms to catch random biases that can lead to unreliable predictions,” says carl vondrik, professor of computer vision and machine learning at columbia engineering. For example, a poorly trained ai to search for images of food may mistakenly calculate whether there is a plate, but not the food itself. Deployed, trained to detect explicit content using datasets typically containing millions of sex video examples – and non-porn. A classifier is only as good as the data it gets from it, says reza zadeh, an associate professor of computer science at stanford university and ceo of computer vision company matroid.Based on examples of user-tagged content posted on twitter, he will say that in reality, tumblr did not include the required number of copies of things like nsfw cartoons in their personal documentation package. Similar in the course, for example, why the classifier mistook burshtein's patent illustrations for 18+ material. “I think what if they forgot to add enough cartoon details and maybe other types of examples that matter and are recognized by sfw,” he says. “Computers just opened their eyes , and it's naïve to think that humans have great vision. Reza zadeh, matroid Wired tried to post some tumblr posts that were reportedly tagged with the natural classifier matroid nsfw images as adult material, including an image of chocolate. Ghosts, photo of joe biden and chief of burstein's patents, at this stage for led-lit jeans. The classifier correctly identified any of them as sfw, although he believed that there was a 21 percent chance of the phenomenon that chocolate ghosts were nsfw. The test claims that using the services of such images there is nothing "adult" - the essential thing is that they are forced to have sex by different classifiers. Make a mistake. Such mistakes,” zadeh says. "Computers have only recently opened their eyes, and it's naive to think that humans can see perfectly." Tumblr has had problems with nsfw messages being clearly tagged before. Also in 2013, yahoo bought tumblr—its group that never figured out how to make big money—for $1.1 billion in cash. Then, four years later, like russian nesting dolls, verizon bought yahoo for about a few billion dollars. (Both hotmail and tumblr are now a segment of a verizon subsidiary called oath.) Immediately after the second acquisition—perhaps in an effort to shape a unique online resource that is more aesthetically pleasing to advertisers—tumblr introduced "safe mode," a subscription feature that automatically filters out " confidential" material in the toolbar, even in search results. Users quickly discovered that safe mode accidentally filters regular content, even lgbtq posts. In june of last year, tumblr apologized and said that, on average, fixed the problem. Now the blogging platform refuses this functionality, due to the fact that in the near future all tumblr will function in a secure plan . Constantly. It's unclear if the company will end up borrowing the same ai technology it used for the site-wide benign mode. In response to the issue, tumblr did not specify what technology it is able to use to help you comply with the new rules for porn content. A source familiar with the firm said it uses modified proprietary technology. The company said in a message to the workshop that, as well as the bulk of the social platforms. User-generated networks, it plans to use a combination of "machine occupancy classification and human moderation by us trust & safety, the group of people who help moderate tumblr." The company also says it will soon increase the number of moderators it employs. Tumblr's competitors have also benefited from a more than decade-long run. While tumblr has always allowed porn — its former ceo defended allowing explicit content on the site even after it was acquired by yahoo — other sites like facebook have long since banned explicit content. These platforms have spent years accumulating nsfw training data to hone their image recognition tools. Any time a human moderator removes porn from facebook, this example can be used to train its ai to detect such things on its own, as tarleton gillespie, a microsoft researcher and author of guardians of the internet, pointed out on twitter. . Platforms like facebook and instagram have also already faced many more philosophical issues that tumblr has yet to sort out, such as when a nipple should be considered a violation of its policy or not. Tumblr will soon have to decide where it wants to draw the line between art (which it says is legal) and pornographic material, for example. To become an adult content-free platform, tumblr will have to improve its automated tools and likely train its classifiers on larger datasets. But the company will also have to answer a lot of hard questions that only humans can solve. - Using popsocket has changed my fucking life- what is the fastest 100m run? Can a person run?- Amazon wants you to code the ai brain for this little machine- spotify year-end ad highlights the weird and the beautiful- hate traffic jams? Curb your love of online shopping. Get even more information with our weekly backchanneltwitter newsletter Paresh dave Vittoria elliott Will knight Lauren goode Chris stockel-walker