OpenAI simulates deployments to predict model behavior before release

OpenAI's pre-deployment safety reviews have traditionally relied on targeted evaluations and red-teaming, but those methods have known blind spots. Evaluation prompts may not cover all possible failure modes, can be biased toward previously seen contexts, and models increasingly recognize when they are being tested, which distorts their behavior. Deployment Simulation addresses these issues by resampling real conversation prefixes with the new model, creating a distribution that better represents actual usage.

The results show meaningful progress: predictions had a median multiplicative error of 1.5x, and the method outperformed baselines on behaviors with large incidence changes. The most significant error source is simulation fidelity—how realistically the environment reproduces production conditions—which OpenAI considers easier to improve than prompt distribution shift. However, the method cannot detect extremely rare failures (below 1 in 200,000 messages), so adversarial evaluations and red-teaming remain necessary.

The stakes hinge on whether simulation fidelity can be improved enough to make tail errors trustworthy for deployment decisions. If the pipeline's occasional large errors persist, it may raise doubt about whether predicted spikes in undesirable behavior are real or artifacts. OpenAI expects the method to play a larger role as the pipeline becomes easier to run, but its utility for external auditors will depend on whether better public datasets can narrow the gap created by private production data.