i suggest for the chess aspect, which is not really the subject of how chat GPT was used. it could be any game.. ratings systems are oblivious to the game rules. it only uses paired outcomes and population of individuals history of those, etc.... to find back the lichess blog where the author tried to find out what it could have learned from PGN now omnipresent on the web, as probably having been part of the unknown large database part of the make up of such chatbot. The blog was about testing chatGPT as annotator, based on the PGN.
The conclusion, was that it was actually probably basing its annotation on the SAN as a purely linguistic type of association. Not really aware of the positions such PGN was actually determining (if decrypted by trained human, or PGN viewers aided humans, which could annotate the PGN languages code moves, knowing the non-explicit FENs hinted by the moves sequence if applied at each turn. The text of the SAN, rarely contains FEN, which could be a hope that the cross-correlation in the FENs, which do explicitly hold the spatial relationships of the information, that the PGNs do not, if always assuming the standard initial position. Can't invent cross-correlations in data, when the data never makes such information visible. because. we have the decrypting key, does not mean that the text contains what we understand.
This topic has been archived and can no longer be replied to.