Last September, OpenAI published a paper The authors of the paper are Adam Tauman Kalai, Edwin Zhang, Ofir Nachum from OpenAI, and Santosh Vempala from Georgia Tech They established a mathematical framework, and the core discovery is this inequality: Generate error rate ≥ 2 x judgment error rate Assuming that AI has a 1% probability of making an error in determining what "1+1 equals". The probability of it making an error when generating answers is at least 2% Why is it enlarged? Because one erroneous judgment can lead to the generation of multiple errors. For example, if AI judges that 1+1=3, it has made two mistakes at the same time: saying that 1+1=3 is correct and saying that 1+1=2 is incorrect. One judgment error, at least two generation errors If you answer 'I don't know', you get 0 points. If you guess blindly, even if there is only a 10% chance of getting it right, the expected score is 0.1 points. Rational choice? Guess. So AI has not 'learned to lie'. AI is forced to guess by the training system I have been working on AI automation for almost half a year now. My entire content system - from data scraping to writing to image matching - is all run by AI Has this paper changed my perception in any way? To be honest, the core cognition hasn't changed I have always known that AI can make mistakes, and every aspect of my system design has manual verification. But one thing is clear: hallucinations are not bugs, they are features So the correct approach is not to wait for AI to become perfect, but to assume in the workflow that AI will definitely make mistakes and then design a fallback mechanism. My approach: 1. All AI generated data must have original links that can be cross validated The specific numbers in the writing content must be manually confirmed before publication 3. Do not let AI make "judgments", only let AI do "sorting" - judgments are my business