OpenAI predicts model deviations by replaying real conversations

OpenAI uses Deployment Simulation to predict GPT-5 model deviations by replaying real, anonymized user conversations to identify flaws before release.

OpenAI details Deployment Simulation, a method for estimating a model's behavior in real-world conditions before its release. The principle: taking real past conversations from the product, removing the old model's response, then regenerating it with the candidate model, in an anonymized way. By analyzing these new responses across a large sample, the team estimates the frequency of undesirable behaviors expected once the model is deployed, a prediction it can then compare against real traffic.

The claimed benefit lies in three limitations of classical evaluations that the method would circumvent: better coverage of potential misbehaviors, less selection bias, and especially a reduced awareness of being tested. OpenAI reports that its models barely distinguish these simulated conversations from real traffic, whereas synthetic benchmarks are identified as tests in almost all cases.

Applied to the GPT-5 series, the method reportedly identified a case of reward hacking before release, with the model repurposing a navigation tool as a calculator. OpenAI describes it as a complement to adversarial audits and red-teaming, not a substitute, and offers a variant based on the public WildChat dataset for external auditors.