OpenAI announced new safety features for ChatGPT that improve its ability to detect signs of self-harm, suicide, and violence by analyzing context across conversations. The update introduces temporary 'safety summaries' that capture relevant context from earlier messages, allowing the model to identify escalating risks rather than treating each message in isolation. This comes as OpenAI faces lawsuits and investigations over claims that ChatGPT mishandled dangerous conversations, including a federal lawsuit linking the chatbot to a mass shooting and a state lawsuit from a family alleging it encouraged drug use. The company collaborated with mental health experts to refine model policies and training, focusing on acute scenarios like suicide and harm to others. OpenAI noted that context matters in sensitive conversations and that the summaries are short-term, not used for permanent memory or personalization. Future expansions may include other high-risk areas such as biology or cyber safety.

Key facts

• ChatGPT uses temporary safety summaries to capture context across conversations.
• Focus on detecting signs of suicide, self-harm, and violence.
• OpenAI faces lawsuits over alleged mishandling of dangerous conversations.
• Company collaborated with mental health experts to refine model policies.
• Safety methods may expand to biology or cyber safety in the future.

KeyAudit data perspective