I’m currently editing a nature guidebook for a national park in Australia. Somewhere between the birds of paradise and the stinging trees, there’s a profile of a lizard called the southern angle-headed dragon. Small, quick, lives in the leaf litter, minds its own business. I needed a little inspiration for a closing paragraph, so I asked an AI tool to take a shot at it. It wrapped the piece up by having the dragon fly off into the forest with scaly wings.
Sorry, what?
To be fair to the AI, it did a lovely job. It was evocative stuff…if we’d been talking about a mythical dragon. The moment it saw the word “dragon,” it stopped reading the paragraph and started an entirely fantastical style of writing about a dragon. One with wings and, presumably, fire-breathing abilities.
AI is Over-Confident
AI wasn’t confused. Confusion implies that it noticed that something was off and wasn’t sure what to do about it. Instead, it was completely overconfident. It just wasn’t reading for meaning and it got carried away with a fairytale instead of checking context.
This is the whole problem and the reason I don’t think AI can or should edit or proofread. Even the increasingly good AI that can genuinely help to an extent with structure, pacing, or spotting a clunky sentence, simply can’t use any discernment. It’s just working off codes and programming.
Proofreading and editing aren’t really writing skills. They’re checking, creative skills. You’re holding two things apart in your head at once: 1) what the sentence says, and 2) what the writer actually meant, or what’s actually true. Generating fluent text and verifying it aren’t the same thing. AI is extremely good at the first one, but it has no real mechanism for the second. It doesn’t automatically know what an angle-headed dragon looks like and it can’t use discernment to ascertain that, based on the rest of the piece, it’s likely not mythological. It knows what the word “dragon” tends to be followed by in the enormous pile of text it learned from. In this case, mythology won that contest.
So, while AI may be free, quick, and easy, it’s still essential that a knowledgeable, skilled human being checks the work that it generates.
Other Ways This Plays Out With AI
On this project and others, I’ve observed artificial intelligence messing up in these ways:
- It smooths without checking. Feed it a paragraph with a small deliberate irregularity (I love a short, blunt sentence sitting in the middle of longer, flowy ones), and it’ll often tidy that up into something more uniform. It’s not wrong. It’s just optimising my text for a kind of fluency that isn’t the same as correctness. It can’t always tell the difference between a mistake and a choice.
- It forgets what it decided three pages ago. A species gets a full common name the first time it appears, and a shortened version after that, the way any style guide would ask for. Ask AI to work across a long document and it’ll lose track of what it settled on, sometimes within the same chapter. This makes for a lot of extra work for the final human editor who needs to ensure this consistency across a whole book.
- It fixes what wasn’t broken and misses what was. I’ve had it “correct” a sentence that was fine and sail straight past a factual slip sitting two lines above it. I wonder if it was because the factual slip was grammatically tidy and, therefore, kind of invisible to something that isn’t checking facts in the first place.
- It leaves its own fingerprints while removing yours. Ask it to polish a paragraph and you’ll often get one back with a tell-tale AI trope (such as the overuse of em dashes where there weren’t any, or a “it’s not X, Y, or Z, it’s A” construction that nobody asked for). Your voice is subtly evicted. And your voice needs to be the essence of your work to make it unique and engaging.
But AI Isn’t Useless
It’s been helpful to me as a drafting tool to get me started, a way to simplify a paragraph when I’m being too technical, or a second pair of eyes for structure or minor mistakes (like double spacing). But there’s a difference between a tool that helps you to write and a tool that can be trusted to catch what you missed. Allowing it to proofread and edit independently only works if you can trust the second part completely. And ask any hiker who’s spotted the southern angle-headed dragon of Lamington Park in person and they’ll tell you that AI got this one very wrong and can’t be trusted.
