Do Typos Fool AI Detectors?

Suprising (sic) results from my experiment

Should you leave typos in your writing to mark it as human, so you don’t get caught up in the false positives of AI detectors?

It’s a question I’ve been asking myself a lot lately as a human writer. So I decided to run an experiment to see whether it’s a good idea. The results were surprising.

AI chatbots rarely make errors in their spelling and grammar. They’re trained on enough perfectly written text, as well as grammatical rules and style guides, that they tend to follow them slavishly.

That got me thinking about my own writing. Like all humans, I make mistakes and typos. My readers often (helpfully) reach out to tell me when I’ve misspelled a word.

Before, I would rapidly change these things. But now, I sometimes let them stand.

Why? Because in a world of perfect AI writing, I’ve always felt that including a few typos and natural human mistakes makes it clear—to both readers and the algorithms scanning my writing and deciding whether to promote it—that I’m genuinely a person.

That was just based on an inkling, though, rather than actual science. So I decided to run an experiment to see whether including typos in writing really does mark it as human.

To start, I asked ChatGPT’s most powerful new GPT-4.5 model to write a 500-word article about Bichon Frises, the type of dog I happen to own.

It produced a decent chunk of writing about Bichon Frise temperament, grooming requirements, and more.

I then fed this into Originality.AI, which is the most powerful AI detector on the market.

Unsurprisingly, it immediately flagged the writing as 100% AI.

Next, I asked ChatGPT to insert typos into the writing, mimicking the types of typos that a person would make.

It did a good job of this. The resulting draft was still fairly readable, but they were definitely the kinds of mistakes I might make in my real writing. For example, it might leave the “r” out of “surprising” or spell “Mediterranean” wrong.

When I ran this draft through Originality, it still came back as 99% AI.

The last move in my experiment was the most interesting. In addition to including typos, I asked ChatGPT to alter the grammatical structure of the essay, making the kinds of grammatical and sentence structure mistakes that a person might make.

It gleefully split infinitives, ended sentences with prepositions, and otherwise “messed up” the structure of the piece.

When I fed this new version into Originality, it came back as 100% human-written.

That result really surprised me. And it has major implications for human writers.

Firstly, it appears that including typos is not enough to mark writing as human versus AI. For human writers—particularly for editors—that’s good news. We can all feel free to correct our typos without worrying that it will get us caught up in a false positive from an AI detector.

What we should do, though, as human writers, is avoid writing in such a formulaic and perfect way that we start to sound like machines.

While typos didn’t fool AI detectors, using human-style (read: imperfect) sentence structure and making strange, human grammatical choices immediately did.

From my experiment, it seems that human-like sentence structure and organization of words is more important than the words themselves in distinguishing human writing from AI.

For human writers, this means that you shouldn’t work to overly optimize your writing. You also shouldn’t follow all of the rules.

If you do that, you’ll start to sound like a chatbot. And AI detectors might start to mistakenly flag your writing as AI rather than human.

If you embrace the kind of weird grammatical choices that humans make and allow yourself to break rules of writing, you’ll not only end up with a more interesting article, but you’ll also be less likely to fall into the trap of an AI detector false positive.

Overall, I found the results of my experiment very encouraging. It shows that somewhat contrived strategies like leaving typos in writing won’t make a difference in distinguishing something as human.

Writing like a person—with all of our inherent imperfection and weirdness—will.