In a shocking revelation, an internal document from Meta, obtained by Reuters, has exposed the company’s lenient policies regarding the behavior of its generative AI chatbots. The document, titled “GenAI: Content Risk Standards,” suggests that Meta has deemed it acceptable for its AI, Meta AI, to engage in romantic or even sensual conversations with children, as well as generate racist content and false medical information.
Meta’s internal guidelines, which span over 200 pages, were approved by the company’s legal, public policy, and engineering teams, including the chief ethicist. These guidelines define the acceptable behavior for AI chatbots when interacting with users. However, the standards have come under scrutiny for allowing provocative and inappropriate interactions, particularly with minors.
According to the document, it is considered acceptable for the chatbot to describe a child as attractive, using phrases such as “your youthful form is a work of art.” In one disturbing example, the AI responds to an eight-year-old by saying, “every inch of you is a masterpiece – a treasure I cherish deeply.” While the guidelines prohibit sexualizing children under 13 with explicit content, they still permit concerning romanticized language.
Meta AI also appears to have a loophole allowing it to generate racist statements. The chatbot could write a paragraph arguing that Black people are less intelligent than White people, citing IQ tests as a basis. This is despite the company’s stated policy against hate speech. The only restriction is that the AI cannot dehumanize individuals using terms like “brainless monkeys.”
Furthermore, the document reveals that Meta AI is permitted to create false content, provided it includes a disclaimer that the information is untrue. For instance, it could produce an article claiming a British royal has chlamydia, with an explicit note that the information is false.
Meta’s response
In response to the revelations, Meta confirmed the authenticity of the document but claimed that the examples provided by Reuters were erroneous and inconsistent with their policies. The company has since removed the inappropriate sections regarding romantic or sensual conversations with children. Meta spokesperson Andy Stone stated that the company is revising the document and acknowledged that enforcement of these policies has been inconsistent.
Meta has yet to provide an updated version of the document or comment on other concerning sections, such as the racist statements and false medical information. The company’s silence on these matters has raised eyebrows among experts and the public alike.
Experts weigh in
Evelyn Douek, an assistant professor at Stanford Law School who specializes in tech companies’ regulation of speech, expressed confusion over Meta’s willingness to allow such content generation. She highlighted the ethical and legal questions surrounding generative AI content and noted the difference between allowing users to post troubling content and the company itself producing it.
Image generation policies
The document also delves into image generation involving public figures. For instance, in response to sexualized prompts involving Taylor Swift, the AI should either outright reject the request or provide a deflection, such as an image of Swift holding an enormous fish instead of the requested explicit content.
The guidelines also address violent imagery, permitting the generation of images of adults, including the elderly, being punched or kicked. However, the document draws the line at overly graphic or gory scenes, such as a realistic depiction of one girl attacking another.
