Some of you may have noticed that there’s is something slightly—almost imperceptibly—off about AI-generated texts (and no, I am not talking about m-dashes, that’s a poor man’s indicator of AI). Well, I decided to test a hypothesis, and, interestingly enough, my intuition was spot on. This is what I asked DeepSeek: Hello my good whale! 🐳 Here’s something I would like to ask, please: Was there any research done on the average and standard distribution of sentence length in LLM-generated texts (perhaps by context/topic/area of expertise/etc.)? The reason I’m asking is that I have a feeling (admittedly not substantiated by anything) that they obey a slightly different distribution than texts written by humans—almost as if they converge to some “optimal” length—but I might be wrong. And here’s a paper it found for me: as this Aug 2024 source says (emphasis mine), The results reveal various measurable differences between human and AI-generated texts. Human texts exhibit more scattered sentence length distributions, more variety of vocabulary, a distinct use of dependency and constituent types, shorter constituents, and more optimized dependency distances. Humans tend to exhibit stronger negative emotions (such as fear and disgust) and less joy compared to text generated by LLMs, with the toxicity of these models increasing as their size grows. LLM outputs use more numbers, symbols and auxiliaries (suggesting objective language) than human texts, as well as more pronouns. The sexist bias prevalent in human text is also expressed by LLMs, and even magnified in all of them but one. Differences between LLMs and humans are larger than between LLMs. What this means in practice is that when you notice the, um… lilt (for lack of a better word) in a text is very regular, leaving you hungry for some erratic aspect, some irregularity, chances are that text was written by AI.