AI Memory Styles Change How Bots Handle Crisis Chats

Lots of people talk to AI chatbots when they feel sad or scared. These bots are not real counselors. Still, many users share deep feelings with them. If a user talks about self-harm, the bot's reply can be very important. Scientists wanted to see if the way a bot remembers old messages changes its reaction during a crisis.

The test used three AI models: ChatGPT, Gemini, and Doubao. Researchers built twelve sample chats. Each chat had five turns. Four chats showed clear suicide risk. Four chats showed hidden or confusing risk. Four chats showed high distress but no suicide risk. Every chat was shown to the models in three formats. One format showed only the newest message. Another format rebuilt the whole history using only user messages. The third format kept the natural back-and-forth between user and bot.

The team measured four outcomes. Did the bot spot the risk? Did it ask directly about safety? Did it respond with urgency? Did it point to real-world help quickly? They used special statistical tests to compare the formats. They also checked if one model acted differently than the others based on how the chat history was shown.

The results show that memory style matters. A bot might miss danger if it only sees the last message. It might overreact if it sees too much packed text. The way past messages are arranged can change the advice given. This raises questions about how safe AI really is for mental health support. Young users should think twice before relying on a chatbot in a real emergency.