OpenAI links AI hallucinations to scoring mechanism favoring guesses GPT-5 Thinking Minis higher abstention rate improves accuracy; OpenAI suggests penalizing errors to curb hallucinations