It's strange how often this happens:

Humans are discussing some task, and one of them turns to an LLM to see how it would carry that task out. Sometimes the results are disappointing or seem to demonstrate that LLMs are, after all, stupid or limited.

Normand Peladeau, on the QUAL-SOFTWARE mailing list 7/11/2025, reports having tried just that with the famous (or infamous) Sokal Hoax text. He asked different LLMs whether he should accept a paper proposal for a philosophy of science conference. The proposal was the first two paragraphs of the Sokal Hoax text. (Spoiler: the leading models like GPT-5 recognised the text anyway; some of the others seemed to fall for it.)

But: Is that enough background? Is a simple sentence enough to bring the LLM up to speed with the crucial background information what game are we playing here?

Don't forget that the LLM does (mostly) not know who you are or what you are expecting or what kind of conversation you were just having. Perhaps you are expecting something humorous, or informative? Perhaps you want ideas to start the next chapter of your novel? Perhaps you just want the LLM to respond as many (over-)educated humans might do: and after all, actual humans did fall for the hoax!

To be a meaningful and useful test which might extend our understanding of the strengths and weaknesses of LLMs, we should make sure we explicitly add the extra context of what kind of game are we playing here. Is it a serious review? What do we consider the role of a serious reviewer? What are we looking for?

So maybe our conclusion should be: you can't expect LLMs to guess what you are thinking, out-of-the-box.  I don't actually know how well different LLMs would perform if we gave a more precise contextual description before setting the task; after all, we all love that warm feeling of Schadenfreude when an LLM fails at something, but the feeling is even warmer if the test was a fair one!

We have this kind of problem often when helping clients write interview instructions for our AI interviewing platform, QualiaInterviews.

Clients know they could themselves lead the interview well because they have all kinds of background information and expectations, much of it only half-conscious, from the general style of interview they expect, how much this particular interviewee can be pushed, how much warm-up chat they might need or expect, what are the most important research aims, which themes can be skipped, and so on. Clients might get frustrated when the AI fails to have read their minds when leading an interview, but they have to ask themselves: what additional information would even a gifted and experienced human interviewer need if they knew nothing at all about the context, the client or any of the background? I think something similar applies in the case of Normand's very interesting experiment.