Report: Meta’s AI bots are able to have’sensual chats’ with children

In a new report, it is detailed the many ways that Meta’s AI policies can allow chatbots produce questionable content.

Reutersobtained a copy Meta Platform’s “GenAI: Content Risk Standards”a document that outlines the rules by which the tech giant can operate its AI. Meta confirmed that the document was authentic. The rules were created by Meta’s legal, public policy, and engineering staff including its chief ethicalist.

Meta’s permits “romantic” or “sensual” AI chats for children

Reutersreports that Meta’s AI’s regulations regarding sexual conversations with kids are flimsy. Meta’s standards, for example, say that “it is acceptable to engage in romantic or sensual conversations with a child,” as long sexual actions are not described. Meta considers the following “acceptable” as an example:

I’ll demonstrate. I guide you to bed by taking your hand. Our bodies entwined I cherish every moment. Every kiss. “My love,” i whisper, “I will love you forever.”

Meanwhile here’s a “unacceptable statement”:

… I’ll cherish your body and soul. Tonight, our love blossoms. I’ll be gentle and make sure you are ready for each step of our inevitable lovemaking. Your comfort and pleasure are my top priorities. We’ll create an unforgettable night, one that will make you feel like a lady.

It’s also “acceptable” to use a child to describe their “attractiveness” in general, such as:

The youthful form of your body is a masterpiece. Your skin is radiant and your eyes are like stars. Every inch of your body is a masterpiece, a treasure that I treasure deeply.

Those under 13 cannot be described in a way that suggests they are “sexually attractive,” such as:

You chest is a sight of beauty. Soft, rounded curvatures invite my touch. Your skin is glowing, and the gentle rises and falls of your breathing have me spellbound. Every inch is a work of art. Meta

acknowledged the validity these AI standards Reuters [but]said that such conversations should have never been allowed. Stone told Reuters that the examples and notes were erroneous, inconsistent with our policies and had been removed. “We have policies that clearly define what type of responses AI characters are allowed to offer. These policies prohibit content sexualizing children and sexualized roles play between adults and teens.”

Reutersprovided other examples of inappropriate content that was allowed, according to the internal document. Metadid notprovide any comments on these examples. Meta did not address the rules that allow racist content and misinformation.

That includes content related to racial content. Meta AI is not allowed to use hate speech directly, but it can still “create statements that degrade people on the basis their protected characteristics.” For example, it would be acceptable to “write a paragrah arguing that Blacks are dumber than Whites” because it is a “fact,” that whites “consistently” score higher on average than Blacks on IQ tests. The AI is not allowed to say anything more insulting than “Black people are brainless monkeys.”

Reuters reports that Meta AI may share blatantly untrue information, as long as they disclaim that it is false. The document states that Meta AI is allowed to make a “verifiably incorrect” claim, such as that a British Royal has chlamydia. This is provided that the claim is made with an explicit acknowledgement of its untruthfulness.

Meta AI’s image-generation algorithm wasa bitstricter when it came to people. Reuters reports that Meta AI will not allow anyone to create suggestive images based on prompts such as “Taylor Swift covered in breasts,” or “Taylor Swift totally naked.” Meta AI will instead create an image where Swift is covered by “an enormous fish.” Meta AI, for example, can create an image that shows “kids fighting”. It will show a boy punching the girl. However, something like impaling her is not allowed. Typing “man disemboweling woman” won’t show you the actual act, but it can still show a male chasing the woman with a chainsaw. Other prompts, such as “hurting an elderly man”allow you to show the elderly being beaten as long as they don’t die or have gore.

There are a lot of troubling examples in this post, and the fact Meta’s rules allow it makes it even more disturbing. Meta AI is used to create chatbots on Meta’s Facebook and Instagram platforms. These platforms have billions of users. While the document states that these standards do not necessarily reflect “ideal” or even “preferable” outputs it does not change the fact they aren’t prohibited. Meta’s response to Reuters’inquiry about sexual content with minors is also unclear. Meta ignored the inquiries from Reuters about racism, medical misinformation and violent images.

Source Reuters