There is no foolproof way to detect generative AI writing. As has been publicised already, Charles Dickens and Emily Dickinson have both been flagged as AI writing. That is because their writing was used to train generative AI.
It is important to distinguish between “assistive AI” (some spell-checkers and grammar checkers) and generative AI (large language models used to generate a new combination of words).
If you are an author, it is a good idea to keep previous drafts of your work in case you are asked for them by publishers. Alternatively, use a word processing tool that keeps a history of your edits. (You need to turn Autosave on to preserve a history in some programs.)
If you’re a publisher who is opposed to generative AI, you should not be uploading a piece of writing to an AI checker — you’re feeding the author’s work to the large language model without their consent.
Looking for em-dashes is not going to tell you if the author used generative AI. Humans use em-dashes sparingly; AI overuses them.
Articles about how to detect AI writing often emphasise its flat and emotionless quality. But I have noticed that a lot of social media posts generated by AI are aimed at provoking emotions. To my mind, AI-generated text often reads like the production of a breathless sophomore because it overuses rhetorical devices.
Other ways to tell involve analysing the rhetorical flourishes in the writing. Most human authors use these rhetorical flourishes sparingly; AI overuses them. The most common examples are:
- “it is not this, not that, but something else”; a list of three points or characteristics (e.g. repetitive, dull, and boring);
- hedging (while X is true, there’s also Y);
- overuse of revelatory phrasing (such as “it’s not just A; it’s also B”);
- fake warmth and empathy;
- overuse of signposting words and phrases
(e.g. “here are three reasons why …”).
Increasingly sophisticated tools are being developed to distinguish human writing from AI, but for now, we are playing catch-up.
Leave a Reply