AI doesn't get 1 + 1 wrong anymore... but yes it can still be wrong. Most commonly if you ask it the wrong way. A calculator can do the same if you forget the parentheses, or mess them up slightly.
My feel is that the degree to which a wrongly inputed question can get a wrong answer is proportional to the complexity of the question you are trying to get an answer to. Kinda like computers made it possible for a simple mispelling of someone's name to have much more massive consequences than before computers.

AI does get things wrong, but it owns up to it when pointed out, and often self corrects once it has observed that something is wrong with the work. Your characterization is straight-up wrong.

"The bcrypt hooks caught only GOG web traffic (blob_007 is literally plaintext GET /token?... HTTP) — the save crypto never touched bcrypt.dll, confirming it's done inside the exe by statically-linked OpenSSL. That also means the ChaCha state I scanned is likely TLS traffic, not the save. Two parallel paths now: brute-force the record layout against AES-GCM (OpenSSL's default AEAD), and locate the AES tables in the exe to find the cipher code:

No AES tables at all — the game uses hardware AES-NI, so table-xrefs are out (and that also means my earlier AES-schedule scan is the right detector... except it never finished; that was the hang). Let me nail the layout facts offline and check my tooling:"