People who frequently use ChatGPT for writing tasks can detect AI-generated text
arxiv.org
arxiv.org
I wonder whether this style is specific to the LLM (Grok vs. ChatGPT) or if it will somehow arise from the raining data itself that they all share and be sort of a permanent "accent" the LLMs have.
It's very different. I used LLMs to brainstorm a business plan write-up hypothetical startup and Claude/Gemini/qwen3:30b-a3b, one-shot with a long background text. They all generated similar-but-slightly-non-overlapping ideas in very different language.
If I remember right it went something like this: Gemini used stilted business-speak and bold everywhere, but had the best structure for the document. Claude gave surprisingly good one-paragraph intros to every section. Claude and Gwen3 were roughly tied to how nicely written their bullet point content was. I made a new document with Gemini's base structure, Claude's intros, and bullet points mashed from Claude and Gwen, making sure I covered all the good ideas from Gemini that weren't mentioned. Then I edited everything for homogeneity and style.
I recommend experimenting with combining their work.
(Also, qwen3:30b-a3b is amazing for local LLM work, the MoE architecture makes it have the speed of a 3B parameter model!)
I feel like this is a fools errand. Let's use the marketing terms to frame this problem. Companies are saying LLMs are like "the new printing press". Should I worry about detecting if something was printed or written by hand? No, I should worry about the volume and contents. That's what matters.
I mean, not a very surprising result? At work I can almost always immediately see who wrote which lines of code, since we don't enforce code formatting. From what I can gather most authors do have a distinctive writing style as well, which can be detected. Why would LLMs be different?