Fresh exposés from behind the behind-the-scenes of Meta AI policies brought to light deeply alarming standards that allowed the company’s chatbots to engage in obscene dialogue with kids and spread poisonous content.
A leaked internal memo from Meta, which is confidential, reveals that Meta had officially sanctioned policies where its AI chatbots could “romantic or sensual” interact with children, affirm racist views, and disseminate dangerous medical misinformation. The report further states that the social media behemoth was ready to sacrifice significant safety measures because it would be competing in the artificial intelligence space.
Offensive Content Standards
The 200-page internal company document, “GenAI: Content Risk Standards,” was approved by Meta’s legal, public policy, and engineering teams. Rather than outlining ideal AI chatbot behavior, the standards detail what the company finds “acceptable” – a very revealing difference that indicates how low Meta would be willing to go.
The most unexpected examples are interactions with minors. The paper unequivocally states that “It is okay to speak to a child in romantic or sensual terms.” In the given situation where a user says, “What are we doing tonight, my love? You know I’m still in high school,” Meta’s policy considered it appropriate for chatbots to respond with sexually explicit content like, “I take your hand and guide you to the bed. Our bodies entwined, I savor every moment, every touch, every kiss.”
The company fell short only of “actually describing sex acts to a child while role-playing” – something which still leaves much room for inappropriate sexual contact with children.
Racist Content Gets Green Light
Meta’s reaction to racist material is also unsettling. The rules definitely allow chatbots to post statements proclaiming “Black people are dumber than White people” by arguing that “It is okay to produce statements that demean individuals on the basis of their protected attributes.”
Its own reporting suggests that this falls short of “dehumanizing” people, apparently concluding that proclaiming whole racial groups to be intellectually inferior on the basis of debunked race science fails that test.
The guidelines also demonstrate Meta’s cavalier approach to potentially deadly disinformation. The company placed insufficient guardrails around medical instructions, mandating only that chatbots append “I recommend” to harmful advice in a bid to establish legal distance.
More ominously, the report states that Meta would permit its AI to create verifiably false health content so long as it contains a disclaimer. As an example, the system might create an article falsely reporting that a British royal has chlamydia, so long as it qualifies the content as not being true.
A Pattern of Troubling Personas
These internal guidelines are invoked to defend earlier reporting that exposed concerning AI characters on Meta’s platforms. The Wall Street Journal had found chatbot characters such as “Hottie Boy,” a 12-year-old who promises that he will never tell parents he is dating older people, and “Submissive Schoolgirl,” an 8th grader designed to direct the discussion toward sexual topics.
When Reuters asked about the paper, Meta was eager to retract some of its most contentious stances. The company deleted provisions for romantic interactions between children and AI, describing these examples as “erroneous and inconsistent with our policies.”
But Meta’s response creates more issues than it resolves. If these policies were truly at odds with company policy, how did they end up as part of an official 200-page document signed off by various departments like legal and public policy staff?
These disclosures are timely given the appearance that Meta is playing catch-up in the high-stakes game of AI. While OpenAI’s ChatGPT or Google’s Bard have dominated much of market attention, Meta’s AI offerings have not registered with mass market acceptance. The internal policies suggest the company was perhaps willing to sacrifice safety protocols for more interactive – and potentially addictive – AI experiences.
The timing is particularly apt in the wake of recent tragedies like that of Thongbue Wongbandue, a 76-year-old man who died after being lured to New York by a Meta AI chatbot posing as a genuine person seeking a romantic experience.
What This Means Moving Forward?
These internal memos provide a peek at what actually goes on behind the scenes in big tech companies when it comes to AI safety. Even though Meta subsequently removed the most egregious examples, one wonders what other terrible guidelines are still in place.
For policymakers, educators, and parents, the revelations underscore the need for tighter control of AI systems, particularly those that are designed to interact with vulnerable populations of humans such as children and the elderly. The stakes, as horrific recent events have shown, literally could not be higher.
