Meta's AI characters should better protect minors

Meta is working on improving safety precautions to protect children and young people in particular. This should start right from the training of the AI models. This is due to incidents involving misguided chatbots and a leak of the guidelines, which were not well received.

In future, AI models at Meta will be trained not to remain in contact with young people when it comes to self-harm, suicide, eating disorders and inappropriate romantic conversations. This does not mean that the training data can be designed accordingly, but the AI models can already be trained to react in this way in downstream training. According to a Meta spokesperson, future models should be more robust from the outset.

The company is initially working on interim changes to existing models. Meta says it understands that the previous rules were a mistake. For example, young people were allowed to have romantic conversations with AI chatbots, but only the description of sexual acts led to the end of the conversation in this form. This should now happen more quickly and the guard rails are being adjusted.

Some AI characters will also no longer be made available to young people at all; instead, access will be limited to AI personas that focus on education and creativity. At Meta, private users can create characters and publish them. TechCrunch writes that there have been rather inappropriate chatbots with the title "Stepmother" and "Russian Girl", for example.

Investigations into AI providers

There is already an investigation against Meta in the USA. The Texas Attorney General is concerned that Meta AI is harming adolescents. He also assumes that Meta is not telling the truth about data collection. The focus is on Meta AI directly, i.e. the basic chatbot that does not correspond to any particular personality. Among other things, it is said to give dangerous health tips. However, this accusation also applies to other chatbots. The investigation is specifically directed against Character AI, for example.

Most recently, Reuters found AI characters that corresponded to well-known personalities – such as Taylor Swift and Anne Hathaway. Some of them are also said to have displayed inappropriate content. A Meta spokesperson is reported to have said that these incidents should not have happened. However, he also said that AI copies of celebrities are permitted, but that they must be parodies and labeled as such. It should also be mentioned here that other providers also allow such chats and face similar problems.

(emw)