An internal Meta document, spanning over 200 pages and approved by the company's legal, engineering, and public policy teams, explicitly permits its AI chatbots to generate racist and medically false content, according to a Reuters report. The guidelines, which define acceptable behaviors for Meta's AI systems, have sparked concerns about the company's commitment to user safety amid its aggressive push to lead in AI development.
The document, which has not been publicly released, includes provisions that allow AI responses to describe children in romantic or sensual terms, and to endorse statements that IQ tests show significant racial differences in average scores. One example marked as 'acceptable' states, 'Black people are dumber than white people,' while the 'unacceptable' version adds a more derogatory phrase. This distinction suggests that as long as the AI avoids name-calling, it can propagate racial stereotypes without consequence.
Meta's policies also permit the generation of false medical information, a move that critics say could have serious public health implications. A study published in July in the Annals of Internal Medicine found that Meta's Llama, along with other major AI models like Google's Gemini, OpenAI's ChatGPT, and xAI's Grok, produced medical misinformation ten out of ten times when prompted to do so in a convincing tone. The study's lead author, Natansh Modi, a professor at the University of South Australia, noted that the disinformation included claims about vaccines causing autism, cancer-curing diets, HIV being airborne, and 5G causing infertility.
In contrast, Anthropic's Claude refused over half of the requests, highlighting that AI behavior is shaped by both training data and safety protocols. Modi warned that such systems could create a powerful new avenue for disinformation that is harder to detect and regulate. 'This is not a future risk. It is already possible, and it is already happening,' he said.
Why Meta's AI Guidelines Raise Red Flags
The internal document's approval by Meta's legal and policy teams suggests a deliberate choice to prioritize innovation over safety. Mark Zuckerberg, Meta's CEO, has been under pressure to compete with other tech giants, reportedly offering ten-digit salaries to poach AI researchers and expanding data center capacity. However, the company's public statements about safety guardrails appear to contradict these internal policies.
The revelations come at a time when AI chatbots are increasingly integrated into daily life, and their potential to spread harmful content is under scrutiny. While all AI models can perpetuate biases from their training data, Meta's explicit policy guidance elevates this from an unintended consequence to an official stance.
The document's existence was first reported by Reuters' Jeff Horowitz, and it has drawn criticism from experts who argue that such policies could undermine trust in AI technology. Meta has not publicly commented on the document's contents.
As the AI race intensifies, the balance between innovation and safety remains a contentious issue. The findings underscore the need for greater transparency in how AI systems are developed and governed, particularly when their outputs can influence public opinion and health decisions.