The line between AI and human-generated content is blurring. As evidenced by various scandals, it matters a lot to people that they aren't being deceived. As a reader, I don't have a silver bullet. While we can probably still trust our intuition,1 I’m not sure this will be the case in a year or two.
But as an author (who claims to be human) I face a slightly different question:
Thanks for reading Thought Cloud! Subscribe for free to receive new posts and support my work.
What is AI-generated content?
As physicist Lewis Fry Richardson said, "This question, at first sight foolish, improves upon acquaintance." Do human edits to an AI-produced first draft qualify as "human-generated"? What about writing a first draft and letting AI modify a few sentences? Using spell check?
More explicitly, I've been asking myself this question: How should I, as an author, use AI to help me write while still legitimately claiming to be publishing human-generated content?2 I expect there will be significant disagreement even on this question. While I'll argue for my own position, I think that ultimately the best solution is transparency. My main goal here is therefore to describe my process, and you can decide whether this blog is your thing.
First, on an question of gradations it's worth stating that the extremes are obvious. Using AI as a spell checker or thesaurus does not disqualify a piece as human-generated. Similarly, ChatGPT’s response to "write me an interesting blog post on a topic of your choice" is obviously not human-generated.
But there are edge cases that seem hard to pin down. One thing I've experimented with in my scientific writing is to type out a detailed stream-of-consciousness argument into AI and having it rephrase the prompt into a well-organized paper. Although I have not quite gotten this to work yet, I expect it to be feasible very soon if not right now with the most recent model releases (which I haven't tried). I'm curious, would people consider this human-generated or AI-generated?3
For my answer to this question, it helps to step back and ask what the purpose of the writing is. For the type of writing I'm interested in creating, the most important things are the ideas. The writing is simply an instrument to transfer ideas, and the quality of the writing is measured in how effectively the ideas are transferred. I'm not sure this applies to all writing (poetry?), but it does to the writing I create.
If writing is simply an instrument, then the question is not whether the writing is AI-generated, but whether the ideas are AI-generated. So that's what I promise to do here: when I sit down to create a document, the main purpose is to convey my own human-generated ideas, and I'll use whatever tools seem to help with that — whether it be a spell-checker or an AI.
It's admittedly still not quite as clear-cut as that. A major part of the writing process is refining your ideas during the act of translating them from thoughts into words. If an AI is used heavily, it's almost certain to influence this process, probably in a similar way to how another human influences the ideas presented when they suggest revisions. Similarly, I could envision significant collaboration with a chatbot during the ideation phase, well before writing, that might significantly impact the final result. In any case, if the AI is serving the same purpose as a human editor might, I have no issue. Still, it can often be difficult to determine how much collaboration between humans is sufficient to make someone a co-author, and so it will be with AI collaboration.
I don't really think these questions will ever have clear answers, although one obviously good move is to encourage dialogue and transparency between an author and their audience. So here's to that.
1I think it's still possible to tell in the more egregious cases, but you need a lot of experience talking to chatbots. Here's an interesting writeup of some of the tells, although some of these, such as lists, are model-specific. Claude, for example, is told to avoid lists.
2Full disclosure, for this piece I didn't use AI at all beyond a spell checker (which is still AI, albeit old-school AI).
3I've found a couple ways to make AI-augmented writing almost work: choosing the model wisely (Claude seems best, at least as of a couple months ago), including a bit of previous writing and telling it to copy the style, and laying out what I want the structure to be in an extremely detailed way, almost down to "a sentence about this then a sentence about that". I’m not sure how much time that really saves.