CAPTCHA: Testing our humanity

You’ve answered this question before. If you’ve browsed the web, you’ve answered this question countless times. Often, it’s accompanied by a demand: Prove it. Read this pair of fuzzy words. Tell us whether this is a picture of a storefront. Click on all the boxes that contain a traffic light. These challenges go by the unwieldy acronym CAPTCHA, which stands for “Completely Automated Turing Test To Tell Computers and Humans Apart.” They’re a crucial cybersecurity tool that helps keep the internet from becoming (even more) inundated with spam, fake accounts, and malicious bots that can be used to hack websites. But they’re also a sneaky strategy to develop artificial intelligence. Every time you fill out a CAPTCHA, you’re helping computers get a little better at tasks like reading distorted text or identifying objects in photographs. Over time, the machines have gotten even better at many of those tasks than humans. Humans, read on. Bots, read this: B̵͍̥̾̿ǘ̸̡̢̨̢̝͖̺̝̹̯̦̰͋̂̃͆̾̀̿́̊̎z̷̮͈̻͓̹̤̜̙͍̏̈́͂̈́́̀͑̈́̕̕ͅz̵̡̖̝̞̳̭̙̉̌̔́̕̕͜͝͝ Ì·Í—Ì½ÍƒÍ›Í ÌˆÌˆÍ†ÍšÍ…Ì˜Ì±Ì˜ÌŸÌ¢Ì¨Ì±Ì¥Í–Ì»Íœớ̵̱͙̱̭̼͖̺̺̪̄̑͐f̶͎̌̈́̃̒̍̂̍͐͘͝f̷̧͉̮̫͚̗̫͔͈̜̂̋̅̿̈́. 🐦 [Tweet this!]( 🌐 [View this email on the web]( [Quartz Weekly Obsession] CAPTCHA October 20, 2021 Are you a robot? --------------------------------------------------------------- If you’ve browsed the web, you’ve answered this question countless times. Often, it’s accompanied by a demand: Prove it. Read this pair of fuzzy words. Tell us whether this is a picture of a storefront. Click on all the boxes that contain a traffic light. These challenges go by the unwieldy acronym CAPTCHA, which stands for “Completely Automated Turing Test To Tell Computers and Humans Apart.” They’re a crucial cybersecurity tool that helps keep the internet from becoming (even more) inundated with spam, fake accounts, and malicious bots that can be used to hack websites. But they’re also a sneaky strategy to develop artificial intelligence. Every time you fill out a CAPTCHA, you’re helping computers get a little better at tasks like reading distorted text or identifying objects in photographs. Over time, the machines have gotten even better at many of those tasks than humans. Humans, read on. Bots, read this: B̵͍̥̾̿ǘ̸̡̢̨̢̝͖̺̝̹̯̦̰͋̂̃͆̾̀̿́̊̎z̷̮͈̻͓̹̤̜̙͍̏̈́͂̈́́̀͑̈́̕̕ͅz̵̡̖̝̞̳̭̙̉̌̔́̕̕͜͝͝ Ì·Í—Ì½ÍƒÍ›Í ÌˆÌˆÍ†ÍšÍ…Ì˜Ì±Ì˜ÌŸÌ¢Ì¨Ì±Ì¥Í–Ì»Íœớ̵̱͙̱̭̼͖̺̺̪̄̑͐f̶͎̌̈́̃̒̍̂̍͐͘͝f̷̧͉̮̫͚̗̫͔͈̜̂̋̅̿̈́. 🐦 [Tweet this!]( 🌐 [View this email on the web]( By the digits [500:]( Years of labor humanity collectively spends solving CAPTCHAs every day [99.8%:]( Accuracy rate for modern AI systems solving the hardest text CAPTCHAs [33%:]( Accuracy rate for humans solving those same CAPTCHAs [$1-2:]( Typical wage workers earn for solving 1,000 CAPTCHAs in illicit “click farms,” where humans help bots beat CAPTCHAs [25 million:]( Books in the Google Books archive, which was digitized with help from CAPTCHA test takers Giphy EXPLAIN IT LIKE I’M 5! We're all professors at AI University --------------------------------------------------------------- From the outset, CAPTCHA tests have been built around hard-to-solve problems in AI. In the early 2000s, machines were terrible at reading distorted text, so the earliest CAPTCHAs asked people to identify wavy, fuzzy, crossed-out words. Every time a human successfully solved one of these puzzles, the answer went into a gargantuan dataset that helped train algorithms to read garbled text. The biggest beneficiary of this system is Google, which owns reCAPTCHA, the world’s largest CAPTCHA company. In 2009, Google began scanning the pages of millions of books to create its massive Google Books archive. Whenever its text-reading AI stumbled over a poorly scanned word it couldn’t quite read, Google used that word in a reCAPTCHA test. Human test-takers would then read the word and tell Google the correct answer. In the process, they also generated training data for Google to use to fine-tune its AI. Google would repeat the process to improve other products. In 2012, the company wanted to use algorithms to sift through images taken from Google Street View to identify street signs and address numbers on buildings to improve the accuracy of Google Maps. So, reCAPTCHA began asking humans to identify numbers and words in grainy images of addresses and signs. Later, reCAPTCHA began presenting humans with an array of pictures and asking them to click on any photo that contained a certain object, like a boat or bicycle. Those tests helped Google improve the accuracy of its image search results. Today, many reCAPTCHA tests ask humans to identify key elements in pictures of traffic—for example, buses, crosswalks, traffic lights, and so on. But Waymo, the self-driving car company owned by Google’s parent company Alphabet, says none of that data is used to train its driving algorithms. Google told Quartz in July that it stopped using CAPTCHA data to train AI entirely, but declined to elaborate on when or why it made that decision. Quotable “Every time our CAPTCHAs are solved, that human effort helps digitize text, annotate images, and build machine learning datasets. This in turn helps preserve books, improve maps, and solve hard AI problems.” —The reCAPTCHA homepage, [as of April 4](, celebrating the use of CAPTCHAs for the advancement of AI. “Today, reCAPTCHA data is only used for security purposes.” —A Google spokesperson on July 19, indicating that Google no longer uses CAPTCHAs to train AI algorithms at all. The [reCAPTCHA homepage]( now makes no mention of AI. LOOK INTO THE 🔮 A world without CAPTCHA --------------------------------------------------------------- CAPTCHAs as we know them will soon be obsolete: The head engineer on Google’s reCAPTCHA team [told The Verge]( that by the end of this decade, standard CAPTCHA challenges—reading funky text, identifying images—will disappear from the web. The demise of the CAPTCHA is mainly the result of rapid improvements in the field of AI. CAPTCHA’s research mission has succeeded so thoroughly that machines are now as good or better than humans at every task we’ve turned into a CAPTCHA test. We’re running out of challenges that humans are universally good at, but machines can’t handle. Besides, people find CAPTCHAs annoying, so researchers have been looking for ways to make testing our humanity less cumbersome. Since 2014, Google has been steadily replacing CAPTCHAs with a system of “continuous authentication,” essentially, a surveillance system that monitors how we browse the web to determine whether we’re acting like humans or bots. Humans, for example, move their cursors differently than robots: We move with meandering imprecision, while bots tend to take the most direct path between two points. While this new method of humanity verification doesn’t require us to waste time on mindless tests, it has [prompted privacy concerns]( from people who don’t like the idea of Google watching their every mouse move. Brief history [2003:]( Carnegie Mellon researchers coin the term “CAPTCHA” in an academic paper that lays out their vision for using CAPTCHAs to block bots and advance AI research. [2007:]( The Carnegie Mellon researchers launch reCAPTCHA, a company that quickly becomes the web’s top source of humanity-proving tests. They later partner with the New York Times to help the newspaper digitize its archives going back to 1851. [2009:]( Google acquires reCAPTCHA and uses it to help digitize the Google Books archive. [2014:]( Google launches No-CAPTCHA, which differentiates between humans and bots without requiring a test. Users simply click a box that says “I’m not a robot.” [2015:]( A Massachusetts woman files a class action lawsuit against Google arguing the company unfairly exploits web users for free labor. A judge later [dismisses the case](. [2017:]( Google launches an “invisible” CAPTCHA that silently tracks web users’ behavior to determine whether they’re human or not. [2020:]( CloudFlare, a company that provides security services for many websites, stops using reCAPTCHA, citing privacy concerns. 2021: Google tells Quartz it no longer uses CAPTCHA data to train its algorithms, and scrubs all mention of AI from [the reCAPTCHA homepage](. Giphy Pop quiz Which of these is NOT a real proposal for a CAPTCHA test? Classify human faces by expression, gender, and ethnicityIdentify which of three jokes is funnyComplete a nursery rhyme popular in your geographic areaIdentify the hidden picture in a magic eye image Correct. Bots and humans alike would struggle with this one, since humor varies widely between and within cultures. Incorrect. There’s at least one group of academics out there who believe this is the best way to distinguish between humans and bots. If your inbox doesn’t support this quiz, find the solution at bottom of email. Visual representation [A stick figure sits at a desk in front of a computer screen that reads, ] Fun fact! In 2017, Amazon patented the “Turing Test via failure”—a CAPTCHA so hard that only computers could get the answer right. People [prove their humanity by getting it wrong](. Giphy Podcast Want to know even more about CAPTCHA? Then listen to this week’s episode of the Quartz Obsession podcast, which features reporter [Nicolás Rivero]( going down the rabbit hole. [Click here to listen](, or subscribe on [Spotify](, [Apple](, [Google](, or wherever you get your podcasts. Sponsored by American Express [Listen now]( Try this! Improving AI's tin ear --------------------------------------------------------------- To make the web accessible for users with visual impairments, most CAPTCHA tests offer an audio alternative: Those who can’t see a fuzzy word can prove their humanity by listening to a scratchy recording of a series of numbers. If you’ve never done an audio CAPTCHA, you can [try your ear here](. If you already use audio CAPTCHAs, [listen to these proposals for new tests]( that aim to trick speech recognition bots by layering multiple human voices on top of each other or interspersing real words with gibberish. Watch this! The mother of all Turing tests --------------------------------------------------------------- If you think CAPTCHAs are tough, imagine being a human at the annual Loebner Prize competition. Each year, a panel of judges participates in a series of five-minute text conversations with the world’s best chatbots and a handful of human volunteers. The judges are trying to figure out who is a human and who is a machine. The bot that fools them most consistently wins the Loebner Prize—and the human who most consistently convinces the judges they’re not a robot is crowned “the most human human.” Quartz chatted with 2009’s most human human about psychology, artificial intelligence, and the nature of humanity. Giphy Poll How do you feel about Google using your CAPTCHA responses to train its AI? [Click here to vote]( I’m glad my responses went toward something usefulGoogle better pay me for my contributionsWho cares? 💬let's talk! In last week’s poll about [randomness](, 58% of you said that a good night’s sleep is the hard-to-price thing you value most, while 30% of you think the comforting sound of a good friend’s laughter is worth more, and 12% of you value randomness over either of those. 🤔 [What did you think of today’s email?](mailto:obsession%2Bfeedback@qz.com?cc=&subject=Thoughts%20about%20CAPTCHA&body=) 💡 [What should we obsess over next?](mailto:obsession%2Bideas@qz.com?cc=&subject=Obsess%20over%20this%20next.&body=) 🎲 [Show me a random Obsession]( Today’s email was written by [Nicolás Rivero](, edited by [Annaliese Griffin](, and produced by [Jordan Weinstock](. [facebook]([twitter]([external-link]( The correct answer to the quiz is Identify which of three jokes is funny. Enjoying the Quartz Weekly Obsession? [Send this link]( to a friend! Want to advertise in the Quartz Weekly Obsession? Send us an email at ads@qz.com. Not enjoying it? No worries. [Click here]( to unsubscribe. Quartz | 675 Avenue of the Americas, 4th Fl | New York, NY 10011 | United States