Is Your Book About to Be Flagged as AI? Here's How to Check, Free
By Mark Hankin
There's a particular worry doing the rounds in author groups lately, and it goes something like this: you wrote your book, every word of it, and now you're frightened a machine is going to decide you didn't.
It's not paranoia. AI detectors have spread everywhere, schools and platforms lean on them, and authors have started quietly running their own manuscripts through the things just to see. Some of them get a nasty surprise. A chapter you bled over for three weeks comes back "85% AI," and suddenly you're arguing with a website about whether you exist.
So let me try to be useful here, because I've spent more time inside this problem than is strictly healthy.
Detectors are guessing, and they're not very good at it
Here's the uncomfortable truth nobody selling a detector wants to say out loud: there is no reliable way to look at a finished piece of text and know whether a human or a machine produced it. None. The Authors Guild, who launched their own "human authored" scheme this year, admitted as much themselves.
What detectors actually do is measure vibes. They look for the things AI writing tends to do, smooth even rhythm, tidy transitions, a sort of frictionless competence, and they score you on how much you resemble that. The problem is that plenty of humans write like that too. If English is your second language, if you're naturally a clean and careful writer, or if you've edited a passage within an inch of its life, a detector will cheerfully accuse you of being a robot. Meanwhile the genuinely AI-written stuff has got fluent enough to walk straight past most of them.
I know this because I built a checker, and the first version was rubbish.
The time my own tool got it embarrassingly wrong
To test it, I asked ChatGPT to write me a short piece of fiction. An alien invasion, ships the size of atoms riding moonlight, the full cosmic-wonder package. It was about as human-written as a parking ticket. Lovely, fluent, completely synthetic.
I fed it into my shiny new tool and it came back with 28 out of 100. "Reads as mostly human." Marvellous. I'd built a smoke alarm that compliments the fire.
The reason was simple, and it's the same reason most detectors fall over. My tool was treating fluency and nice imagery as signs of a human hand. They aren't, not anymore. Modern AI is fluent by default. The actual tell isn't whether the prose is smooth, it's whether it's specific. AI writing reaches for the grand and the general ("a soft wash of silver," "the terrible beauty of it") because it has no particular memory to draw on. Real writing tends to be lumpier, odder, full of the one weird detail only you would have thought of.
Once I taught the tool to weigh genericness and abstraction instead of polish, that same alien story jumped to 88. Which is roughly where it always belonged.
What Amazon actually cares about (it's simpler than the panic suggests)
Worth saying plainly, because the detector anxiety is partly aimed at the wrong target. Amazon KDP does not ban AI. It asks you to disclose it. And the test it uses has nothing to do with what a detector thinks.
The question KDP is really asking is: who wrote the first version of these words, you or the machine? If an AI produced the text and you published it, that's AI-generated and you tick the box, even if you then edited it heavily. If you wrote it and used AI to tidy it up, that's AI-assisted, and you don't need to declare anything. There are three separate boxes, by the way, for text, images and translations, so an AI-made cover counts even on a book you wrote every word of. I went into all of this properly over here: how much AI is too much.
The point is that a low detector score won't save you and a high one won't sink you. What matters is the honest answer to that one question.
So what's the checker for, then?
Fair question, given I've just spent several paragraphs telling you detectors can't be trusted.
Two things. First, it's a sanity check. Paste a draft and you'll get an honest estimate of how AI it reads, which passages stand out, and a plain-English reminder of what KDP expects. If you've been writing in good faith, it tends to come back reassuringly low, and that's a nice thing to see before you hit publish. If it comes back high on something you thought was all yours, that's worth a look too.
Second, and this is the bit I actually care about, it's a demonstration of why detection is the wrong tool for the job. A score is a guess about the past. It can never prove anything, because it's looking at the finished page and trying to reverse-engineer the story of how it got there.
The only thing that can prove how a book was written is a record made while you write it. That's the whole idea behind gowrite. As you work, it keeps track of what the AI suggested and what you actually kept, so when someone asks "did you write this?" you have an answer that's evidence rather than a shrug. No detector on earth can offer you that, because they weren't in the room.
How to use it
It's free, it doesn't need an account, and it doesn't keep the text you paste. Drop in a few hundred words, see where you stand, and read the note about disclosure while you're there. If you'd rather have a score you can actually stand behind, start writing the next bit in gowrite and let the record build itself.
Your name's going on the cover. It's worth knowing the book behind it is one you can defend without flinching, ideally before a dropdown menu asks you to swear to it.