TrustNoRobot

Category: Distributed Human Identity Test

  • Imagine All The (DeepFake) People

    Imagine the year is 1989, and I put in front of you a computer with a mouse, keyboard, and wired connection to, if not the Internet, CompuServe or one of the other graphical interfaces to remote machines and information.

    Now, if I asked you to project forward twenty years, could you have imagined all that boiled down and expanded 100x in your pocket, connected wirelessly to billions of remote services? Some people have that kind of imagination for sure, but most people would have struggled to see how it would be possible. Or necessary.

    Now I ask you to consider these technologies:

    1) Realtime face mapping to simulate another person, as seen here.

    2) Realtime voice cloning that matches your own voice, as seen on sites like lyrebird.ai.

    3) Enough processing power to enable real-time near-perfect simulation of a human being.

    All three of these things exist today, and we are already seeing the emergence of DeepFakes as a service with sites like deepfakesweb.com. Sure, it’s not anywhere near passable as a human yet, but neither was your Apple IIe as a mobile device.

    And here is the kicker: it’s not going to take 20 years. It’s going to happen over the next two years. By 2021 you will not be able to trust the authenticity of any given talking head video unless it’s certified in some way.

    We don’t need new laws to protect us from this forced-Turing test, existing fraud and copyright laws should suffice. We need new technology. And not tech to identify a fake once it’s out in the world, already damaging our identities, but rather tech to certify the authenticity of a video from its inception.

    Digital watermarking will be an important component of this new technology, but that’s like saying RAM is an important component of your mobile device. We will need to authenticate our devices and use Distributed Human Identity Tests to validate ourselves.

    More about some research on AI-based digital watermarking research:

    It replaces the typical photo development pipeline with a neural network—one form of AI—that introduces carefully crafted artifacts directly into the image at the moment of image acquisition. These artifacts, akin to “digital watermarks,” are extremely sensitive to manipulation.

    In tests, the prototype imaging pipeline increased the chances of detecting manipulation of images and video from approximately 45 percent to over 90 percent without sacrificing image quality.

    It’s a start, but will need to be created into consumer tech rapidly to afford us any protection from the chaos that is about to occur.

    DeepFakes are going to be a lethal new weapon dropped into the ongoing ideological, class, and epistemological battles being waged in our society. The level of disruption they will cause is hard to imagine, but start by thinking about a perfect video of you saying the opposite of your current political beliefs posted on social networks.

    Identifying bots is almost impossible now. When they begin to hijack real identities, or fabricate them whole cloth, the game is over for a trusting society.

    Astounding Science Fiction Oct 1953 cover art

  • In 2016, America Failed the Turing Test

    The problem with the Turing test is we are conducting it against our whole civilization, and in this case the testers, normal Americans using social media, were not exactly looking to help contribute to a philosophical discussion about AI sentience.

    In 2016, America failed the modern version of the Turing test. We proved that we could not distinguish bot networks from real human networks.

    The testers, in the case of Americans dealing with networks of bots posing as humans on social media, were angry, scared, distracted humans who are looking to validate their own beliefs.

    These bot networks still exist and lie in wait for the next election cycle to commence. And no law will fully stop them, though I applaud the effort. Another point, these bot networks are controlled by self-serving assholes.

    Let’s switch this up.

    Instead of the Turing Test, what if the test was harder. Not a little bit harder. A lot. And what if it was designed not to detect if a computer could pass as a human, but instead if a human could validate its status as a known entity.

    The Turing Test asks if a human can be fooled into thinking a computer program is a human. It sets up conditions that do not penalize the computer program for not being biological – by putting the human behind text messaging, the test lets the tester focus on just the content of the messages.

    But see, a human isn’t a text message, and in this case, the medium is truly the (text) message. By crippling the human in this test, and letting the bot hide behind text, we create an illusion of equality. But we are not equal with programs, that is a category error that is deeply confusing to many.

    To help eliminate this confusion, and focus on the real issues around identity and bot defense, I propose the Distributed Human Identification Test.

    1) The DHIT test is conducted via videoconference conversations with multiple remote people over long periods of time, in other words, whenever you video chat with someone the test is in effect.

    2) In the DHIT test, you talk to someone you know in a regular conversation. This is a video chat, and can be a business call or a friendly share.

    3) At the end of each conversation, you will be asked if there was anything to indicate that the entity you were taking to was transformed or artificial in any way.

    4) Over time, and many tests, a validity score can be built up about an individual. If they drop beneath a certain threshold that individual would need to take further steps to confirm their human Identity.

    The point of this test is not to determine if a computer can fool a human. We have bots that can do that now, in the wild, outside of the laboratory. Spoiler alert: They can even get bumbling Presidents elected.

    The point of the Distributed Human Identification Test is to confirm human Identity. If you can pass it, you are a real person, known by other real people. We don’t need Equifax for identity confirmation when we have the people you went to high school with.

    By using video in realtime and normal conversation we make the test very difficult for an AI to ever pass. By conducting the test with people you actually know we make this almost impossible to fool. By making it distributed we use the power of large real social networks to build a constantly evaluated Web of Identity.

    The DHIT will protect us from bots, and turn our attention to the real danger: if realtime deepfake videos of us as individuals start being used to commit crimes and spread misinformation, by shitty humans, we will have no defense.

    An AI bot who can get a home loan over videoconference using our identity can ruin your life in twenty minutes, and bring down a civilization in an hour.

    But your Mom knows you, your friends know you, your boss knows you, the network of real people know you.

    Properly integrated into existing video chat systems, the DHIT could be our greatest defense grid against deepfake fraud. And eliminate the greatest security risk, passwords, once and for all. Because boiling things down to text, be it a Twitter post or a password, is the root of the problem.

    Because text doesn’t matter.

    Ideas don’t matter.

    Only people matter.

    Render by Mike Winkelmann

    Painting by Simon Stålenhag