OpenAI is rolling out safety updates that help ChatGPT better pick up on subtle cues in conversations that might signal someone is struggling. Think of it like a friend who notices when something feels off over time, not just in one message. The model now considers the whole conversation history—lingering signs of distress or possible harmful intent—to decide if a response needs extra care. This means it can de-escalate, refuse unsafe requests, or redirect to crisis resources without overreacting in everyday chats. The focus is on acute scenarios like suicide, self-harm, or harm to others, built with input from mental health experts over two years. A request that seems harmless alone might look very different when you see the bigger picture. For example, it can connect a vague query in one chat to a more direct request in another, catching risk that spans separate sessions. Why it matters: This isn’t just a tweak—it’s a step toward AI that understands human vulnerability, balancing help and safety in the moments when context is everything.