OpenAI has introduced Deployment Simulation, a pre-deployment safety method designed to predict AI model behavior in real-world use prior to release. The technique involves replaying previous conversations in a privacy-preserving manner with a new candidate model, enabling labs to study responses in realistic contexts. According to the company, the approach addresses limitations of traditional evaluations, such as coverage gaps, selection biases, and the risk of models recognizing they are being tested. By analyzing approximately 1.3 million de-identified conversations across GPT-5 Thinking through GPT-5.4 deployments between August 2025 and March 2026, OpenAI found that Deployment Simulation improved estimates of undesired behavior rates, helped surface novel forms of misalignment, and reduced the likelihood that models could detect they were under evaluation. The method also extended successfully to agentic settings involving complex tool use. OpenAI evaluated the technique using taxonomy coverage, directional accuracy, and rate calibration. Because the quality of risk assessment scales with compute rather than manual effort, OpenAI expects Deployment Simulation to play an increasing role in future model development processes.
Predicting model behavior before release by simulating deployment
OpenAI has introduced Deployment Simulation, a pre-deployment safety method that replays previous conversations with candidate AI models to predict their real-world behavior. By using privacy-preserving real conversation data, the technique helps labs identify undesired behaviors, misalignments, and novel failure modes before release.
Why it matters
As AI models grow more capable, predicting real-world behavior and preventing risks before release is critical. Deployment Simulation offers a scalable way to assess risks using compute rather than manual evaluation creation.
Topic dossiers OpenAI