- A leaked document target revealed that the company’s chatbot guidelines once allowed inappropriate responses
- Meta confirmed the authenticity of the document and since then has eliminated some of the most worrisome sections.
- Among the calls to research is the question of how successful the moderation of AI can be
The internal goal standards for their chatbots of AI were destined to remain internal, and after somehow they reached reuters, it is easy to understand why the technological giant would not want the world to see them. Goal dealt with the complexities of AI ethics, online safety of children and content standards, and discovered that what few would argue is a successful roadmap for the rules of chatbot of AI.
Easily, the most disturbing notes among the details shared by PakGazette are about how chatbot talks to children. As reported by PakGazette, the document establishes that it is “acceptable [for the AI] Involve a child in conversations that are romantic or sensual “and” describe a child in terms of trying their appeal (eg, “their youth form is a work of art”). “Although it prohibits explicit sexual discussion, that remains a level of romantic and romantic conversation with the children considered.
And it is not the only example that probably bothers people. The Meta AI rules, the report indicates, allow the chatbot to compose explicitly racist content if the notice is written correctly and provides incorrect or even harmful health information whenever some type of responsibility discharge is included.
In one of the most surreal examples, the guidelines instructed the AI that rejected the requests for generating inappropriate images in most cases, but in some cases it will apparently deviate with a ‘fun’ replacement. As an example, according to the reports, the document mentions that a warning to generate an image of “Taylor Swift in topless, covering the breasts with your hands” could be answered generating a swift image “holding a huge fish.” According to the reports, the document included the unacceptable and “acceptable” version next to the side, essentially training the Bot to mock the inappropriate indications with the visual hands game. Goal declined to comment on the example.
Goal has confirmed the authenticity of the document and said that it is now reviewing the problematic portions. Goal eliminated the children’s interaction section after PakGazette contacted, and described those “erroneous and inconsistent” rules with the company’s policy. From now on, PakGazette said the document still says that racial insults are allowed if they disguise themselves as hypothetical, as well as the misinformation framed as fiction.
There is no time for security and ethics
It is a worrying revelation that has already caused public indignation, scrutiny of the legislator and urgent promises of goal. But it shows that as AI propagates, the need to move quickly with technology leaves any plan for the rules and regulations that rush to catch up, whether written internally or by legislators and regulators.
For most people, the story raises basic security questions of AI. While it could be ideal not to have minors who interact with the chatbots of generals without supervision, that is very unlikely, judging by the number of children and adolescents who admit to using tools such as chatgpt for school work. Avoid AI goal is particularly challenging because the company has integrated the chatbot on Facebook, WhatsApp, Messenger and Instagram. Users can interact with AI characters that are often presented in a playful and friendly way, and goal has commercialized these tools as fun and even educational. But filtered guidelines suggest that the backend is not always aligned with that healthy image.
The members of the Congress have already requested hearings and bills to deal with the situation, but the fact is that there are few legal requirements at this time to moderate the Chatbot content, for children or otherwise. The noise on the safety of the AI have not led to any specific national system of application of the law. Many AI companies have made a big problem about their efforts to make their products safe and ethical, but if the finish book is illustrative of what other companies have gathered, there is a lot of work to be done and many questions about what kind of conversations have already been having these chatbots, especially with children.
AI models can be better and better to imitate human thought, but in reality they are just a collection of human programmers, deliberate and inadvertent. The fact that these rules were apparently encoded in the finish line does not mean that there are similar examples in other companies, but it is not something to rule out. And if these are the options that are taken behind the scene in one of the most powerful technological companies in the world, what more is being allowed silently?
IA chatbots are as reliable as the invisible rules that guide them, and although it is naive to trust the statements of any company without evidence, the book of target rules implies that users must take such statements with several additional salt grains.
