Shadow mode is a deployment and validation technique in which a new AI model or system runs on real production inputs, but its outputs are only logged and never used for actual decisions or actions. The logged outputs are compared with those of the existing system or with human judgment to assess accuracy and safety. It is also called shadow deployment or shadow testing.
Unlike A/B testing or a canary release, in which some users actually receive the new system's results, shadow mode has no effect on users or operations. It is used as a validation stage before full rollout in areas where errors are costly, such as machine learning model operations, autonomous driving software, and agents that automate business tasks.