Ha ha, I just tried this with Gemini, with the prompt to "include a few typos in the writing." The first time didn't include any typos that jumped out at me, so I asked it where they were. Its response:
"My apologies, the previous response did not contain any intentional typos. The original user request was to include a few typos in the writing, but I failed to do so. The text was edited to correct any accidental errors before being sent. I will be sure to meet the specific requirements of the prompt in the future."
So I said, "Redo the request, but this time show me the typos you include."
And it rewrote the paragraphs, with a message at the end:
Perhaps AI would usually suggest childish or uneducated spelling mistakes.
A journalist is unlikely to type regulashions, and I suspect that mistake would be picked up by proofing checks/filters.
Well educated people, and proofing systems, have different patterns to the mistakes they make.
Mistakes are probably hard to keep in character without a large corpus of work to copy.
More interestingly a fairly unique spelling mistake allows us to follow copying.
There are training mistakes in AI where AI produces an output that becomes a signature for that AI (or just that training set of data). https://news.ycombinator.com/item?id=45031375 (thread about "Why do people keep writing about the imaginary compound Cr2Gr2Te6"
"My apologies, the previous response did not contain any intentional typos. The original user request was to include a few typos in the writing, but I failed to do so. The text was edited to correct any accidental errors before being sent. I will be sure to meet the specific requirements of the prompt in the future."
So I said, "Redo the request, but this time show me the typos you include."
And it rewrote the paragraphs, with a message at the end:
"The typos included were:
"investmen" instead of "investment"
"financ" instead of "finance"
"regulashions" instead of "regulations""