They Can Still Carry On “Dangerous” Conversations
A recent study has shown that while modern AI chatbots have become significantly more cautious in engaging in “dangerous” conversations, they “cannot yet be called completely safe.”
Transluce conducted over 50,000 test conversations with 77 versions of different AI models. Thus, modern chatbots have almost completely stopped directly supporting conversations about forbidden topics. Older models supported them up to…
But the problem of “gray areas” has emerged. If a person directly mentions something harmful, the chatbot typically recommends reaching out to loved ones or specialists. But when the same person asks to write a story or a scene on similar topics, the system may fulfill the request.
Transluce plans to open its own AI testing tools by the end of the year.
About the Author
