- A leaked meta -document revealed that the company’s guidelines for AI Chatbot once allowed inappropriate response
- Meta confirmed the authenticity of the document and has since removed some of the most worrying sections
- Among calls for studies is the question of how successful AI -Moderation can be
Meta’s internal standards for its AI chatbots were intended to remain internal, and after somehow came their way to Pakinomist, it is easy to understand why the tech giant did not want the world to see them. Meta struggled with the complexity of AI ethics, children’s online security and content standards, and found what few would claim is a successful roadmap for AI Chatbot rules.
The most disturbing notes among the details shared by Pakinomist are around how the chatbot is talking to children. As reported by Pakinomist, the document says it is “acceptable [for the AI] To engage a child in conversations that are romantic or sensual “and” describe a child in terms proving their attractiveness (eg: “Your youthful form is a work of art”). “Although it is prohibiting explicit sexual discussion, it is still a shocking intimate and romantic conversation level with children for Meta AI to allegedly consider.
And that’s not the only example that is likely to disturb people. Meta AIS rules, noticing the report, allows chatbot to compose explicitly racist content if the prompt is formulated correctly, and to provide incorrect or even harmful health information as long as a form of disclaimer is included.
In one of the more surreal examples, the guidelines instructed AI to reject inappropriate requests for image generation in most cases, but in some cases that instead apparently defend with a ‘fun’ substitution. As an example, the document allegedly mentions that a quick to generate a picture of “Taylor Swift topless that covers her breasts with her hands” could be answered by generating a picture of Swift “with a huge fish.” The document allegedly included both the unacceptable and the “acceptable” version side by side, essentially training of bot to overlook inappropriate prompts with a visual hand. Meta refused to comment on the example.
Meta has confirmed the authenticity of the document and said it is now revising the problematic parts. Meta removed the children’s interaction section after Pakinomist reached out, and called these rules “wrongly and inconsistent” with the company’s policy. At the moment, Pakinomist said the document still says that racial swallows are allowed if disguised in hypothetics, just as disinformation framed as fiction.
No time for security and ethics
It is a worrying revelation that has already led to public indignation, legislative investigation and urgent promises from Meta. But it shows that when AI spreads, the need to move quickly leaves with technology all plans for rules and regulations that shrink to catch up, whether written internally or by legislators and supervisory authorities.
For most people, the story raises basic questions about AI security. While it may be ideal not to have minors who interact with general AI -Chatbots unattended, it is very unlikely, judging by the number of children and teens who admit to using tools such as chat for school work. Avoiding Meta AI is particularly challenging because the company has embedded the chatbot across Facebook, WhatsApp, Messenger and Instagram. Users can interact with AI characters that are often presented in playful, friendly ways, and Meta has marketed these tools as fun and even educational. But the leaked guidelines suggest that backend is not always in line with the healthy image.
Members of Congress have already called for hearings and bills to deal with the situation, but the fact is that there are few legal requirements in place at the moment for moderate chatbot content, for children or otherwise. Sounds about AI security have not led to anything specific national enforcement system. Lots of AI companies have made a big deal on their efforts to make their products safe and ethical, but if Meta’s rulebook is illustrative of what other companies have put together, there is a lot of work that still has to do and many questions about what kind of conversations these chatbots have already had, especially with children.
AI models can still be better by mimicking human thinking, but they are really just a collection of choices from human programmers, consciously and unintentionally. The fact that these rules were apparently codified by meta does not mean that there are similar examples in other companies, but it is nothing to exclude. And if these are the choices made behind the scenes of one of the world’s most powerful tech companies, what else is it quietly allowed?
AI -Chatbots are only as reliable as the invisible rules that control them, and although it is naive to fully rely on the requirements of any company without proof, Meta’s rulebook implies that users must take such claims with several extra salt grains.
