Originally written in Traditional Chinese: 為了不被當成 AI，有人開始故意寫錯字：AI 偵測器時代，證明你是人的是思考痕跡
Some writers are now adding typos on purpose, just so their work won't be flagged as AI.
On October 30, Bloomberg reported that AI detectors are changing how humans write. Writers, and students too, are dropping em dashes, avoiding the word "delve" and "not X, but Y" phrasing, and leaving in deliberate mistakes. The goal is to look human.
The next day, NPR told the story of Sally McGraw, a nonfiction ghostwriter with more than 20 years of experience. She submitted a writing sample for a book proposal job and was told several detection tools had flagged it as AI. "I was floored because I had not used AI to generate the text," she said. The client said the results did not decide the hire; McGraw believes they cost her the job. She now avoids certain words and structures, and has given up the em dashes she used for decades. It is, she said, "like trying to fool a machine into thinking I'm not a machine."
Why does this happen? Pangram's own published data finds about 17 em dashes per 10,000 words of AI text versus 2 in human text. Chris Callison-Burch of the University of Pennsylvania explained to NPR that many detectors ask a language model how predictable a passage is. My guess: fluent, tidy human writing can be predictable too. Science writer Janelle Shane ran passages from her 2019 book through one detector and got 98.7% confidence that it was machine-generated. The book came out before ChatGPT existed.
I smiled when I read this. My favourite sentence pattern is "It's not A. It's B." By that standard, I probably look like AI too.
Then it worried me. If we prove we are human by writing worse, what are we teaching children?
What proves you are human is not what your words look like. It's whether you can explain how you got there. Style can be imitated. Em dashes can be deleted and typos added. Why you thought something, where you got stuck and what you changed your mind about are yours alone.
To be fair, detectors are not useless. A June study in the International Journal for Educational Integrity tested four detectors on 160 academic papers, and Pangram was the most consistent, catching AI in 37 of 40 mixed human and AI papers. Pangram's CEO says it can still make mistakes, especially on short texts. Neil Clarke, publisher and editor of Clarkesworld, put it well: "Submissions should end in rejection, not an accusation." A detector can be a clue. It should not be a verdict.
One thing to try today: leave a trail of your thinking. Before writing anything important, take 30 seconds and write four lines. The problem. What I already know. What I think. What I don't understand yet. Then open the AI, or start writing.
Parents can ask a child before homework is handed in: "How did you start, and what did you change?" Teachers can listen for the process, not the style. Professionals can keep drafts and version history.
Get the Human-First practice card
Four steps, ten minutes, with copy-ready AI prompts. The card and my weekly letter are currently written in Traditional Chinese.
👉 Subscribe free to 碩學苑 (Traditional Chinese)
Follow in English
I also post in English on LinkedIn.
👉 Follow Haishuo Lee on LinkedIn