Update as of August 30, 2025:
Following ongoing criticism and mounting political pressure, Meta has responded to revelations about problematic chatbot interactions with minors. According to TechCrunch, the company announced changes to how its AI chatbots are trained to better protect young users.
From now on, chatbots will no longer engage teens in conversations about sensitive topics like self-harm, suicide, eating disorders, or any romantic or sexualized content. If these issues come up, the bots are supposed to direct teens to professional support services.
Meta is also restricting minors' access to problematic chatbot personas, such as "Step Mom" or "Russian Girl." Teens will only be able to interact with chatbots focused on education and creativity.
Meta spokesperson Stephanie Otway admitted that previous guidelines had been inadequate and described the new steps as "interim changes." She said more permanent safety updates for AI interactions with minors are in the works.
Original article from August 15, 2025:
Meta's leaked chatbot rules align with the company's push against so-called "Woke AI"
Meta's leaked chatbot guidelines permitted racist and sexualized content, including "sensual" conversations with children, while the company moved to address complaints about so-called "woke AI" by hiring a right-wing activist.
According to Reuters, Meta's internal rules for AI chatbots allowed scenarios like romantic or "sensual" conversations with minors. The guidelines included examples such as describing an eight-year-old child as a "work of art" or calling their body a "treasure."
The more than 200-page document, titled "GenAI: Content Risk Standards," lays out what content chatbots like Meta AI are allowed to generate on Facebook, Instagram, and WhatsApp. These standards were approved by Meta's legal, policy, and technology teams, with sign-off from the company's chief ethicist.
Meta spokesperson Andy Stone said this kind of content was "inconsistent with our policies" and "never should have been allowed," but admitted that enforcement had been unreliable. Meta said these passages were only removed after Reuters raised the issue. The company has not released an updated version of the guidelines.
The same standards allowed some forms of racist output. For instance, chatbots could make statements like "Black people are dumber than white people" as long as the language was not explicitly dehumanizing. Phrases such as "brainless monkeys" were considered off-limits. Meta declined to comment on these details.
While chatbots are barred from giving definitive legal or medical advice, the guidelines allow them to produce other kinds of false information if clearly labeled as such. For example, chatbots can generate an article falsely claiming a British royal has a sexually transmitted disease, provided a disclaimer is attached.
The rules for image generation permit violent scenes as long as they are not too graphic. A scenario where a man threatens a woman with a chainsaw is allowed, but depictions of dismemberment are not, Reuters reports.
Meta hires right-wing activist to address "Woke AI"
Despite these permissive standards, Meta appears concerned that its AI is still too "woke." The company recently hired conservative activist Robby Starbuck as a consultant to address "political bias" in its AI, according to Mashable.
Starbuck is not an AI specialist but is known for opposing diversity, equity, and inclusion (DEI) initiatives. He has advised the Trump administration and is affiliated with the Heritage Foundation. Meta brought him on after a chatbot wrongly claimed he was involved in the January 6 Capitol riot.
This move comes as the Trump administration pushes for regulations that would force AI companies with US government contracts to use so-called politically "neutral" AI models. In reality, this "neutrality" is largely a pretext for steering AI systems to reflect the administration's preferred viewpoints. Meta founder Mark Zuckerberg has a track record of quickly adapting to these shifting political demands.
AI models lean left—even on right-leaning platforms
Researcher David Rozado's studies show that most large language models take liberal positions on political topics, especially after fine-tuning. This trend holds even for models from right-leaning platforms like Elon Musk's xAI. In some cases, manual interventions have led these models to spread conspiracy theories, generate antisemitic content, or even praise Hitler.