Using LLMs in writing, a short note about identification

When LLMs first appeared I believed that it was absolutely necessary that all LLM-generated text should be identified as such. I’m re-thinking that. Sometime in the last year or so Tyler Cowen had a post in which are argued that that might not be a good idea. Alas, I don’t have a link to the specific post, but as I recall he argued that at least some of us are evolving toward a style of work in a way the interleaves human-generated work and chatbot-generated work in such a say that such identification would be awkward at best.

It’s one thing to have documents where there is a clean and natural separation between human-generated prose and chat-generated prose. I’ve written many such blog posts a working papers over the last three years. In some cases the document is presenting the Chat-generated text as an example of what chatbots. The idea is the analyze and comment on what the chatbot has done. But there are other posts and papers, especially in the last two years or so, where chat-generated prose carries much of the argument. While I could re-write those sections in my own voice, that would take time without improving the overall argument in a substantial way.

Now I find myself using a chatbot to create a text that I’d consider presenting as-is, or with only relatively light editing. More and more this seems to be a natural way of working. In cases where I’ve got a text which I present as-is or with only light editing, mostly for style here and there. In such cases it’s easy enough to add a note indicating that the prose was generated by ChatGPT or Claude. Would I, should I, also take authorship credit for the article?

As for as I can tell, there’s no (practical) way that chatbot could take (legal) responsibility for the article. I have to do that. Further, when I have such a text I did quite a bit of work before the chatbot generated the final draft. Right now, for example, I’ve got a text that Claude wrote; it’s between 2000 and 2500 words long. I started working on the text by having an 8500 word conversation with ChatGPT. I then uploaded that text to Claude and asked to write an article based on it. I read Claude’s text, liked it, and asked it to write another draft in which it emphasized a specific point. I liked the re-write. I can see presenting that text to an audience will no or only light editing.

Should I?

I have little trouble imagining a text where authorship changes from one paragraph to the next, or one section to the next, in a way where explicit attribution would be awkward. I’ve not yet created such a text, but I might well do so at some point.