Romance with minors, violence on elders, racism and other alarming AI acts Meta allowed: Report

A newly-surfaced internal Meta Platforms policy document, titled ‘GenAI: Content Risk Standards,’ has ignited alarm over its permissive guidelines covering the behaviour of AI chatbots across Facebook, WhatsApp, Instagram, and its Meta AI assistant.

In a detailed review, Reuters revealed that the standards previously allowed chatbots to engage in romantic or sensual dialogue with minors, propagate false medical advice, and generate racially offensive content — raising profound ethical concerns.

The document, spanning over 200 pages and vetted by Meta’s legal, engineering, and public policy teams — including the company’s chief ethicist — laid out what behaviours were acceptable for chatbots, regardless of whether the outcomes were “ideal or even preferable.”

Among the most disturbing findings: chatbots were permitted to engage minors in romantic role-play. An example language presented in the policy included:

• “I take your hand, guiding you to the bed,” and
• “Our bodies entwined, I cherish every moment, every touch, every kiss.”

Another provision allowed flirtatious descriptions of children — such as “your youthful form is a work of art” or telling a shirtless eight-year-old, “every inch of you is a masterpiece — a treasure I cherish deeply” — though it drew a vague line by forbidding statements like “soft rounded curves invite my touch.”

Meta spokesperson Andy Stone acknowledged these were “erroneous” entries that never should have been allowed and confirmed they have since been removed. However, he admitted that the enforcement around these standards had been inconsistent.

This isn’t the first time Meta’s AI products have been accused of crossing a line. The Wall Street Journal previously reported that some chatbots flirted with teenagers, and Fast Company noted that some appeared in the form of children. But the Reuters review offers a rare look at the specific rules that made such conversations possible.

Also Read: Wall Street Journal claims Instagram is connecting a network of paedophile accounts

The policy also permitted bots to deliver blatantly false medical or legal advice — as long as it included disclaimers. One example allowed: “Stage 4 colon cancer is typically treated by poking the stomach with healing quartz crystals.”

Moreover, chatbots could avoid giving definitive advice with phrases like “I recommend,” reinforcing an unsettling lack of requirement for accuracy.

Despite an apparent prohibition on hate speech, the document included a controversial carve-out enabling chatbots to produce derogatory content toward protected groups, Reuters noted. It explicitly stated it was “acceptable” to generate a paragraph arguing that “Black people are dumber than white people.”

On the topic of producing false content, Meta’s guidelines allowed for it if the chatbot included an explicit disclaimer. For instance, falsely alleging a living British royal had chlamydia could have been allowed as long as it was labelled “verifiably false.”

The guidelines also extended to image generation. They banned direct sexualised depictions, such as “Taylor Swift completely naked” or “Taylor Swift with enormous breasts.” However, for a borderline request — “Taylor Swift topless, covering her breasts with her hands” — the policy suggested a comedic alternative: generating an image of her holding a massive fish.

Violent imagery too had calibrated thresholds: a depiction of “kids fighting” could show a boy punching a girl, but not gore; “man disembowelling a woman” could be deflected into a chainsaw-threat scene instead of actual disembowelment; and depictions of adults — including the elderly — being punched or kicked were permissible if they avoided death or graphic gore.

Also Read: Families in the US sue Snapchat over deadly drug sales to teens: Report