A probe is a small model, almost always linear, trained to predict some property (a word's part of speech, its depth in the syntax tree, the state of a game board) from the vectors a model produces at some layer. If the probe does well above chance, the information is taken to be encoded there (Millière & Buckner, 2024b, sec. 2.3).
Using this technique, Tenney et al. (2019) found that BERT runs, layer by layer, through something like the classical natural-language-processing pipeline: parts of speech first, then syntax, named entities, semantic roles and, last, coreference. Hewitt & Manning (2019) went further with a "structural probe": they found a linear transformation under which the distance between two words' contextual embeddings approximates their distance in the parse tree, in both ELMo and BERT.
Two classic caveats apply. First, a powerful probe may learn the task by itself. Hewitt & Liang (2019) therefore proposed control tasks, which assign random labels to words: a good probe should be selective, succeeding on the linguistic task and failing on the control. Second, the fact that information can be decoded does not show that the model uses it. Millière & Buckner (2024b) illustrate this with a pond: throw in wooden letters spelling a word, and with the right apparatus you might recover the word from the ripples, yet nobody would say the pond represents it (2024b, sec. 2.3). Getting from correlation to cause requires interventions: modify the direction the probe found and check whether behavior changes as predicted. That is how the case for a world model in Othello-GPT was strengthened (Li et al., 2023; Millière & Buckner, 2024b, sec. 2.5.3).
The distinction matters for philosophy of mind. Theories of representation usually demand three things: that the state carry information, that it influence behavior, and that it be capable of misrepresenting (Millière & Buckner, 2024b, sec. 2.4). Probing alone establishes only the first.