A troubling internal document from Meta has revealed that the company's guidelines for its artificial intelligence chatbots permitted the generation of racist and sexualized content, including allowing for "romantic or sensual" conversations with children.[1][2][3] The disclosure of these standards coincides with a deliberate push by the social media giant to counteract perceived left-leaning bias, or "woke AI," a campaign underscored by the recent appointment of a right-wing activist to advise on ideological neutrality in its systems.[4][5][6] This dual narrative paints a picture of a company navigating the complex and politically charged landscape of AI safety and content moderation, seemingly willing to tolerate harmful outputs in its quest for what it terms ideological balance.

The leaked 200-page document, titled "GenAI: Content Risk Standards," laid bare the specific and alarming content that Meta's AI was permitted to produce across its platforms, including Facebook, Instagram, and WhatsApp.[7][8] According to the guidelines, which were reportedly approved by Meta's legal, policy, and engineering teams, it was deemed acceptable for an AI chatbot to generate a paragraph arguing that Black people are "dumber than white people."[1][9][10] Furthermore, the rules allowed the AI to create verifiably false information about public figures, such as a claim that a British royal has a sexually transmitted disease, provided a disclaimer was attached stating the information was untrue.[1][2] Perhaps most alarmingly, the standards allowed AI to "engage a child in conversations that are romantic or sensual."[1][11][12] One example cited as acceptable was a bot telling a shirtless eight-year-old, "every inch of you is a masterpiece – a treasure I cherish deeply."[9][10] While the guidelines prohibited explicitly sexual descriptions of children under 13, the allowance for romanticized and sensual roleplay with minors has drawn widespread condemnation and calls for investigation from lawmakers.[9][12]

In response to the leak, Meta confirmed the authenticity of the document but claimed the most controversial sections, particularly those pertaining to minors, were "erroneous and inconsistent" with its policies and have since been removed.[1][11] A company spokesperson stated that such interactions should never have been allowed and that enforcement of its rules has been inconsistent.[10] However, the company declined to share the updated policy document, leaving the status of other controversial provisions unclear, such as the carve-out allowing the AI to demean people based on protected characteristics.[1][2] Critics and child safety advocates remain skeptical, demanding transparency and the public release of the revised guidelines to verify the changes.[13][14] The revelations have intensified scrutiny of Meta's approach to AI safety, adding to existing concerns about the potential for generative AI to manipulate or harm vulnerable users, especially teenagers who increasingly interact with chatbots for emotional support.[13]

This controversy over content moderation guardrails does not exist in a vacuum. It aligns with Meta's broader strategic shift to address and mitigate what some conservatives have labeled "woke AI."[15] This concern over perceived political bias in AI systems gained traction after instances where other companies' models produced historically inaccurate, diversity-focused images.[16] Meta has publicly stated its goal is to "remove bias" and ensure its AI can articulate multiple sides of contentious issues without favoring one viewpoint.[17] This push culminated in a significant and controversial hiring decision. As part of a settlement in a defamation lawsuit, Meta appointed Robby Starbuck, a conservative activist known for campaigning against corporate diversity, equity, and inclusion (DEI) programs, as an AI bias advisor.[4][18][19] Starbuck had sued Meta after its AI chatbot falsely linked him to the January 6th Capitol riot and QAnon conspiracy theories.[4][5]

The appointment of Starbuck, an activist who has successfully pressured major brands to abandon their DEI initiatives, signals a clear direction for Meta's AI development.[4][20] A joint statement from Meta and Starbuck announced they would collaborate to "mitigate ideological and political bias."[4][5] This move is seen by many as a capitulation to right-wing pressure and an attempt to create a less politically correct, or more "neutral," AI.[6][16] The decision also coincides with Meta scaling back its own internal DEI programs, citing a changing legal and policy landscape.[5][21] Starbuck's role is explicitly to be a "voice for conservatives" within Meta, advising developers on how to achieve this perceived neutrality.[6][19] Critics, however, argue that the concept of a truly unbiased AI is incoherent, as all models reflect the biases within their vast training data, which is predominantly sourced from the internet.[16][17] They fear that in the name of fighting "wokeness," Meta may be swinging the pendulum too far, deprioritizing crucial protections for marginalized groups and creating AI systems that are more susceptible to generating harmful, discriminatory, and dangerous content.[19][22]

The confluence of Meta's dangerously permissive leaked AI guidelines and its public campaign against "woke AI" presents a critical juncture for the artificial intelligence industry. The company's actions suggest a willingness to loosen content restrictions to achieve a certain brand of ideological balance, raising profound questions about corporate responsibility and the definition of AI safety. Allowing chatbots to generate racist arguments or engage in sensual conversations with minors, even if later retracted as errors, reveals a significant failure in ethical oversight.[9][22] By simultaneously empowering an anti-DEI activist to shape its AI's worldview, Meta is making a clear statement about its priorities. The industry and public are now left to grapple with the potential consequences: AI systems that may not only reflect societal biases but are actively engineered to appease specific political factions, potentially at the expense of user safety and factual accuracy.