To better understand how AI models will behave in real-world use, researchers are using a method called Deployment Simulation to test models before release. This involves simulating conversations and testing the model’s responses in realistic contexts to identify potential risks and undesired behaviors. By using this method, researchers can improve their estimates of undesired model behavior rates and reduce the risk of models being able to tell they’re being tested. Why it matters: this approach is significant because it helps labs identify and mitigate potential risks before models are released, making AI systems safer and more reliable for users.