Abstract

For a number of reasons, which I will discuss, it is difficult to gauge the state of the art of current AI in carrying out commonsense reasoning. That said, it seems that ChatGPT-5 Thinking can almost always correctly carry out commonsense reasoning over a large and important (though by no means exhaustive) class of problems. Specifically, ChatGPT-5 Thinking rarely produces an answer that is clearly foolish in answering a problem that (a) is easy and commonsensical; (b) is sensible and has a clear-cut answer; (c) is purely textual in the sense that input and output are text and answering the problem does not require referring to some non-text information source; (d) does not involve very long chains of thought; and (e) do not involve interpersonal relations between the bot and the user. I'll present some examples of the kinds of failures that still arise when these conditions are not satisfied or when models less sophisticated than ChatGPT-5 Reasoning are used. I'll discuss the challenges involved in overcoming these limitations and the contributions that insights from cognitive science might make.