The Machine That Cried Robot

Watching Kids Write While a Machine Guesses

Photo by Katja Ano on Unsplash

I watched a kid write an essay once. Watched the whole thing, start to finish. She chewed her pen. She deleted a paragraph and wrote it again, worse, then better. She came to my office and asked if “in conclusion” was cheating, and I told her no, just tired, and she laughed and went back and fixed it herself. I conferenced with her twice. I have the drafts. I have the timestamps. I have my own memory, which still works, mostly.

Then I ran it through an AI detector, and it looked at her heart on a page and said: 98% machine.

Ninety eight percent. As if a girl who bled ink for a week is somehow two percent human and the rest is server farm.

I want to laugh about this, and some days I do, because the alternative is despair, and despair doesn’t grade papers any better than a bad algorithm does.

Here’s the trouble, plainly: the tools built to catch the fake are guessing, dressed up to look certain. A percentage that precise looks measured. It isn’t. Nothing was measured. Something was estimated and then handed a number so it would sound like science instead of a hunch.

I just finished teaching two online courses, Social Studies 20 and 30, and I saw the other half of that same failure. Not students wrongly accused, but the real thing, undisguised. AI wasn’t tweaking sentences in those assignments. It was writing them outright, start to finish, and it was exhausting to read and respond to week after week. No voice. Flat, impersonal, almost mechanical, the kind of writing that answers the question without ever seeming to have thought about it. I didn’t need a detector for any of it. I needed eyes and a red pen. So I built a policy: where I suspected AI had written the work and the student couldn’t show me otherwise, I marked it at less than half value. I gave every one of them the chance to bring proof, a draft, a screenshot, anything showing the thinking was theirs. Not one student did. That’s the tell a percentage will never catch, and it’s also the limit of a percentage’s usefulness. The real cases didn’t need measuring. They needed someone paying attention.

And the tool tends to flag the students who write the cleanest. The ones who learned grammar. The ones who use a semicolon on purpose. The ones who write in full, structured sentences because someone taught them that structure isn’t the enemy of voice, it’s the frame voice hangs on. Write well enough and you trip the wire. Write messier, more human in its unevenness, and you slide through untouched. So the tool ends up quietly rewarding the rougher draft and punishing the polished one, which is exactly backward from what we’re supposed to be teaching.

I did not sign up to coach students toward writing worse on purpose, just to stay under suspicion.

And now Substack, the platform I hand my own words to every week, the place I came to for exactly the opposite of metrics and panic, has partnered with Pangram, a detection company. On a writing platform. I keep turning that over and I can’t find the part where it makes sense for writers.

Let’s be honest, everyone is using AI nowadays. I can’t run a search on Google that doesn’t come back with an AI insight first. Log into your Telus account and the chat that greets you is AI. It’s stitched into the plumbing of ordinary life now, whether anyone asked for it or not. I use AI too, and let me qualify that. I’ve said so, I’ll keep saying so, because I’m not interested in performing a purity I don’t have. I have ADHD and a job that eats hours like a wood chipper, and sometimes I want the scaffolding built fast so I can climb up and do the actual writing, the part that’s mine, the part with my fingerprints on it.

So every letter I send to parents, every message to the community, every article I post, gets run through AI first, not to write it, but to check the tone, to catch the sentence that reads calm in my head and combative on the page. Last month it caught a line in a letter to parents about a schedule change, a sentence I’d read as neutral that landed, on a second read, like an accusation aimed at the people I was writing to. I cut it before it went out. That’s not a machine writing my letter. That’s a second pair of eyes that never gets tired and never takes it personally. A carpenter who uses a nail gun is still a carpenter. Nobody hands the house back and says this doesn’t count, I heard a machine helped.

Here’s the part I can’t get around, writing this. An essay about detection paranoia is, itself, a test of that same reflex. If admitting all of this makes you trust the essay in front of you a little less, that’s the exact instinct I’m describing, working on you in real time. I’d rather name it than let it work on you quietly.

But I can write. I have always been able to write, long before any of this existed, and I love it, the actual sitting down and wrestling a sentence until it behaves, and that love is what’s under threat here. Not from the technology. From surveillance dressed up as quality control. From a number that claims to know the difference between a voice and a void, and doesn’t.

What I’m actually afraid of isn’t robots writing. It’s humans no longer trusting other humans to have written. A teacher looking at a beautiful sentence and feeling suspicion instead of joy. That’s the real damage. Not plagiarism. Distrust, manufactured and sold back to us as vigilance.

I want my students to become good writers. Great ones, even, the kind who make you stop on the sidewalk and read a line twice. I can’t build that on a foundation of paranoia. You don’t grow trust by handing a percentage generator the keys to the classroom.

So here is where I’ve landed, and it isn’t complicated: watch the writing happen. Sit beside the kid. Read the draft that was worse yesterday and better today. Ask a question you don’t already know the answer to, and see if they can answer it in their own words, out loud, without a screen between you. That’s the only detector that’s ever worked. It runs on attention, not algorithms, and it’s accurate every time, as long as you actually show up.

Online, I can’t sit beside anyone. Social Studies 20 and 30 gave me no hallway, no desk to lean over, no face to watch while a kid works out a hard sentence. So I built the next-best thing: the burden shifted to the student, proof as the price of credit. It’s not as good as presence. Nothing is. But it beats a percentage, because it still asks a human being to account for their own work, instead of asking a machine to guess at it.

The machines can keep guessing. I’ll keep watching kids write, badly and then better, one real sentence at a time.

Vince Hill is a teacher and principal in rural Alberta. He writes weekly on life, politics, and education.