Training adjusts a model’s weights: the numerical parameters used to process input and generate output. Broad training can teach language patterns and many useful relationships. Later training can shape instruction-following, reasoning, refusals and how uncertainty is expressed.

That learning is not a complete, perfectly indexed library. A model can get an older fact wrong, lack a niche detail or reproduce a misconception. A stated knowledge cutoff is a limit on training coverage, not a guarantee that every earlier event is known.

Current or private information can also reach an answer through the task’s input or available tools. Keep learned capability, supplied evidence and retrieved information distinct.