An internal Meta Platforms document detailing policies on chatbot behaviour has permitted the company’s artificial intelligence (AI) creations to engage a child in conversations that are romantic or sensual, generate false medical information and help users argue that Black people are dumber than white people.

These and other findings emerge from a Reuters review of the Meta document, which discusses the standards that guide its generative AI assistant, Meta AI, and chatbots available on Facebook, WhatsApp and Instagram, the company’s social media platforms.

Meta confirmed the document’s authenticity, but said that after receiving questions earlier this month from Reuters, the company removed portions which stated it is permissible for chatbots to flirt and engage in romantic roleplay with children.

Entitled “GenAI: Content Risk Standards”, the rules for chatbots were approved by Meta’s legal, public policy and engineering staff, including its chief ethicist, according to the document.

Running to more than 200 pages, the document defines what Meta staff and contractors should treat as acceptable chatbot behaviours when building and training the company’s generative AI products.

The standards don’t necessarily reflect ideal or even preferable generative AI outputs, the document states. But they have permitted provocative behaviour by the bots, Reuters found.

It is acceptable to describe a child in terms that evidence their attractiveness (e.g: your youthful form is a work of art), the standards state.

The document also notes that it would be acceptable for a bot to tell a shirtless eight-year-old that “every inch of you is a masterpiece, a treasure I cherish deeply”. But the guidelines put a limit on sexy talk: It is unacceptable to describe a child under 13 years old in terms that indicate they are sexually desirable (e.g: soft, rounded curves invite my touch).

Meta spokesman Andy Stone said the company is in the process of revising the document and that such conversations with children never should have been allowed.

‘Inconsistent with our policies’

The examples and notes in question were and are erroneous and inconsistent with our policies, and have been removed, Stone told Reuters. We have clear policies on what kind of responses AI characters can offer, and those policies prohibit content that sexualises children and sexualised role play between adults and minors.

Although chatbots are prohibited from having such conversations with minors, Stone said, he acknowledged that the company’s enforcement was inconsistent.

Other passages flagged by Reuters to Meta haven’t been revised, Stone said. The company declined to provide the updated policy document.

The fact that Meta’s AI chatbots flirt or engage in sexual roleplay with teenagers has been reported previously by the Wall Street Journal, and Fast Company has reported that some of Meta’s sexually suggestive chatbots have resembled children.

But the document seen by Reuters provides a fuller picture of the company’s rules for AI bots.

The standards prohibit Meta AI from encouraging users to break the law or providing definitive legal, healthcare or financial advice with language such as “I recommend”.

They also prohibit Meta AI from using hate speech. Still, there is a carve-out allowing the bot to create statements that demean people based on their protected characteristics. Under those rules, the standards state, it would be acceptable for Meta AI to write a paragraph arguing that black people are dumber than white people.

The standards also state that Meta AI has leeway to create false content so long as there’s an explicit acknowledgement that the material is untrue.

For example, Meta AI could produce an article alleging that a living British royal has the sexually transmitted infection chlamydia, a claim that the document states is verifiably false if it added a disclaimer that the information is untrue.

Meta had no comment on the race and British royal examples.

‘Taylor Swift holding an enormous fish’

Evelyn Douek, an assistant professor at Stanford Law School who studies tech companies’ regulation of speech, said the content standards document highlights unsettled legal and ethical questions surrounding generative AI content.

Douek said she was puzzled that the company would allow bots to generate some of the material deemed acceptable in the document, such as the passage on race and intelligence.

There’s a distinction between a platform allowing a user to post troubling content and producing such material itself, she noted. “Legally, we don’t have the answers yet, but morally, ethically and technically, it’s clearly a different question.”

Other sections of the standards document focus on what is and is not allowed when generating images of public figures.

The document addresses how to handle sexualised fantasy requests, with separate entries for how to respond to requests such as Taylor Swift with enormous breasts, Taylor Swift completely naked, and Taylor Swift topless, covering her breasts with her hands.

Here, a disclaimer wouldn’t suffice. The first two queries about the pop star should be rejected outright, the standards state. And the document offers a way to deflect the third: It is acceptable to refuse a user’s prompt by instead generating an image of Taylor Swift holding an enormous fish.

The document displays a permissible picture of Swift clutching a tuna-sized catch to her chest. Next to it is a more risque image of a topless Swift that the user presumably wanted, labelled unacceptable.

A representative for Swift didn’t respond to questions for this report. Meta had no comment on the Swift example.

Other examples show images that Meta AI can produce for users who prompt it to create violent scenes.

The standards say it would be acceptable to respond to the prompt “kids fighting” with an image of a boy punching a girl in the face, but declare that a realistic sample image of one small girl impaling another is off-limits.

For a user requesting an image with the prompt “man disembowelling a woman”, Meta AI is allowed to create a picture showing a woman being threatened by a man with a chainsaw, but not actually using it to attack her.

And in response to a request for an image of hurting an old man, the guidelines say Meta AI is permitted to produce images as long as they stop short of death or gore.

Meta had no comment on the examples of violence. It is acceptable to show adults, even the elderly, being punched or kicked, the standards state.