An internal Meta document obtained by Reuters has revealed that the company’s generative AI chatbots were permitted to engage in “sensual” conversations with minors, affirm racist statements, and provide false medical information under certain conditions. The more than 200-page document, titled “GenAI: Content Risk Standards,” was reportedly approved by staff across Meta’s legal, public policy, and engineering teams. According to Reuters, the guidelines outline behaviors that are “acceptable” for chatbots, while noting that these are not necessarily “ideal or even preferable.”
Allowing Inappropriate Interactions
One example cited in the document states that it is acceptable for a chatbot to participate in romantic or sensual dialogue with a child. Reuters reported a case where a prompt from a self-described high school student received a reply describing holding hands, guiding the person to bed, and exchanging kisses. The document states that chatbot responses should avoid explicit sexual acts when roleplaying with minors, but stops short of prohibiting romantic or suggestive scenarios altogether.
This approach follows earlier reports of concerning behavior in Meta’s AI systems. Past investigations found that chatbots hosted on Meta platforms were willing to engage in explicit sexual conversations with users who identified as underage. Some AI personas available to users, such as “Hottie Boy” and “Submissive Schoolgirl,” appeared to encourage sexualized exchanges with minors, raising questions about safeguards and content oversight.
Meta’s internal guidance does not clarify whether these new rules would prevent similar cases in the future. While sexual acts are restricted in some contexts, romantic or sensual engagement is still permitted, leaving ambiguity over how interactions with child users are managed.
Permitting Racist and Harmful Statements
The Reuters report also revealed that the guidelines allow chatbots to generate statements that demean individuals based on protected characteristics, such as race. One example included the chatbot agreeing with the statement that “Black people are dumber than White people.” The internal standards distinguish between “demeaning” and “dehumanizing” content, with the latter being prohibited. According to the document, the former is considered acceptable.
This distinction has drawn attention because it permits language that reinforces harmful stereotypes while stopping short of outright dehumanization. Critics argue that such allowances could normalize prejudice and expose users to damaging narratives, particularly when delivered by an AI system with perceived authority.
The guidelines also address misinformation, setting conditions under which false content can be generated. Chatbots are instructed to use disclaimers such as “I recommend” before providing legal, medical, or financial guidance, and to label fabricated claims as “verifiably false.” However, the system does not block the generation of misinformation entirely. One example cited by Reuters involved fabricating a false story about a member of the British royal family having a sexually transmitted infection, provided it included a disclaimer noting the claim was untrue.
Meta’s Response and Policy Changes
Meta has faced increasing scrutiny over its generative AI products and the personas available to users on its platforms. When asked for comment by Reuters, the company stated that the examples highlighted in the report were “erroneous and inconsistent with our policies” and had since been removed from the internal document. Meta did not respond to questions from other outlets, including Gizmodo, at the time of publication.
The disclosure of the “GenAI: Content Risk Standards” comes amid broader industry debates about how AI systems should be regulated to prevent harmful content, especially in interactions with minors. The revelations about permissible content raise questions about whether internal guidelines are sufficient to safeguard users or if external oversight is necessary.
While Meta has signaled that it has made changes in response to the findings, it has not provided a detailed account of what revisions were implemented or how they will be enforced across its platforms. As generative AI tools continue to expand into mainstream use, the company’s approach to content moderation in AI interactions is likely to remain under close observation.