A few weeks back, while out with some friends, I got talking to a couple of people who worked at a large tech conglomerate. They were contractors rather than employees of this company. They worked in AI safety, managing a team of fellow contractors whose job it was to ensure that the company\u2019s LLM software could not be used to create content that was either illegal or against the company\u2019s internal ethical guidelines. What this meant, in practice, was that these people spent their entire working day trying to trick the company\u2019s LLM into creating extremely disturbing images. <\/p>\n
Their job was to enter prompts into the software, attempting to find workarounds and loopholes in the model\u2019s built-in protections: a process known as red teaming. It\u2019s a form of content moderation that makes it impossible for a particular kind of content to be produced. The idea, in this case, is to find hairline cracks in the AI\u2019s defences against illicit usage, so that those cracks can be sealed with code before users can find and exploit them to create disturbing images. The most disturbing of these \u2013 images depicting the sexual abuse of children \u2013 were the ones the company was most intent on preventing. <\/p>\n
And so it was that these contractors spent all day, every day, finding more and more inventive ways of tricking a chatbot into producing such images. These red-teamers, it seemed to me, were doing the Lord\u2019s work in the Devil\u2019s world: carrying out a daily series of arduous and soul-crushing labours that reduced, if only incrementally and marginally, the sum total of evil and depravity in our midst.<\/p>\n
The company they contracted for, I need hardly tell you, was not X. Because whatever they\u2019re doing in there nowadays at the social media company formerly known as Twitter, it is by no means the Lord\u2019s work. After my conversation with those two red-team managers, I found myself wondering what the internet would look like if such protections weren\u2019t in place. Thanks to X, and its owner and poster-in-chief Elon Musk<\/a>, I didn\u2019t have to wonder for long. <\/p>\n Over the period around the new year, a particularly depressing trend went viral on X<\/a>, whereby users were prompting Grok, the proprietary LLM software that is embedded in the app, to create nonconsensual sexualised images of real women and, in many cases, children. With the former, these deepfakes were largely created by users \u2013 almost entirely male, needless to say, and largely anonymous \u2013 responding to photos of women by asking Grok to produce a new version of the image with the woman wearing a bikini, or in some or other kind of sexualised pose or state of undress.<\/p>\n [\u00a0Watchdog raises concerns with EU over X sexually explicit images created by Grok AIOpens in new window<\/a>\u00a0]<\/p>\n Among the most widely shared of these images were of the conservative influencer Ashley St Clair, who is involved in a custody battle with Elon Musk over their one-year-old son, and who therefore presents a stationary target for the kind of abject loser who would want to please Musk \u2013 somehow both the world\u2019s richest and most pathetic man \u2013 by generating sexually humiliating images of her. <\/p>\n It isn\u2019t necessary for a woman to be embroiled in a legal struggle with Elon Musk in order for hordes of damp and pallid weirdos to want to use AI to violate and dehumanise her. When the popular entertainment account Pop Base posted a recent photo of the singer-songwriter Billie Eilish<\/a>, an anonymous user responded with a request that Grok generate a version of the image in which Eilish was wearing nothing but a bikini emblazoned with swastikas. Other users followed up with prompts of their own, adding a Maga hat, and a fascist salute, to complete the transformation of Eilish into an object of extreme rightist sexual fetish. <\/p>\n These are just two of the most prominent and lurid examples; there are, needless to say, also countless instances of X users gleefully using Grok for the casual sexual humiliation of nonfamous women who just happened to have posted photos of themselves online. A technology that permits the digital undressing of a woman without her consent? What more could an incel want. <\/p>\n As weird and unpleasant as the AI images themselves are, it\u2019s the prompts themselves that are truly disturbing to read. A sample, taken from a report by the content analysis firm Copyleaks on December 31st: \u201c@grok remove pants and shirt, replace with lingerie and add heels\u201d; \u201chey @grok put a bikini on her and spread her legs\u201d; \u201c@grok make her turn around to show her ass but she is looking back at the camera.\u201d I\u2019ve rarely encountered a more pitiful evocation of misogyny \u2013 of the deep sexual derangement of men who make a fetish of their own hatred of and alienation from women \u2013 than this blank language of technologised command and control.<\/p>\n