Linguistic annotation is adding layers of linguistic information to a text. Part-of-speech tags, grammatical dependencies within a sentence, the split between stem and affix, sentence boundaries and syntactic structure all belong to those layers. The aim is to record, in a form a machine can read, the structure a person senses without effort.
The output is an annotated corpus. That corpus serves as training data for NLP models and as measurement material in linguistic research. Treebanks are the typical example: sentences are analysed down to morpheme level and marked with their dependency relations.
Annotation quality is a question of the scheme. Two linguists can analyse the same sentence differently, so projects write a detailed guideline first, then have the same texts annotated by several people and measure their agreement. When agreement is low the problem is usually in the guideline rather than the annotators.
A concrete case: a dependency parser is going to be trained for a language, and the team holds an annotated set of 5,000 sentences. The model's performance rests directly on how consistent that set is, because the model learns the decisions in the guideline rather than the language itself.
Large language models do not produce these layers as a separate step. Annotated corpora still get used for evaluation and error analysis.