Inference

The capacity that distinguishes AI systems from other types of systems: the process of deriving outputs through learning, reasoning, or modeling, which transcends basic data processing. Recital 12 identifies two modalities: machine learning approaches (learning from data) and logic-and-knowledge-based approaches (inferring from encoded knowledge or symbolic representation). (Recital 12; Commission Guidelines, 2025-02-06)

The process of using an already-trained AI model to generate a response to a specific input, such as a prompt. Inference is comparatively cheap and happens continuously across every deployed AI product; it is what most end users actually experience.

The act of running a trained model to produce an output for a request. Small per event but recurring at usage volume, making it the dominant carbon source over a deployed feature's life for most buyers.

The act of using a trained model on new input to produce an output; the model applies already-learned patterns without learning anything new from that interaction, in contrast to training.

The stage at which an already-trained model is actually used: a prompt goes in, and the model generates a response in real time, without permanently updating its underlying training from that specific interaction in most current deployed systems.