The Arms Race Is Real and Students Are Winning
spent the entire last weekend trying to understand exactly how my students are getting around detection. i’m exhausted.
Here’s what I found out: students are not just using ChatGPT and submitting. The savvy ones are using it to draft, then using AI humanization tools to rewrite it, then manually editing the result to add specific details and voice. By step three, Turnitin, GPTZero, and Originality.ai all come back negative or low confidence.
This whole detection workflow is based on the assumption that AI text has detectable signatures. Those signatures exist for raw AI output. They mostly disappear after one or two passes through a humanizer plus light editing.
The students who are doing this aren’t the lazy students. The lazy students submit raw ChatGPT. They get caught. The motivated students who also want to cheat are invisible.
So our detection net catches the low-effort cheaters and misses the high-effort ones. it’s an inverted filter.
6 Replies
Join the discussion.
Log In to Reply
inverted filter is exactly right. and nobody in admin wants to hear that. we bought the turnitin subscription, we're detecting cheating, case closed. except we're only detecting the students who aren't trying very hard.
Jacob's right that the incentive structure is the problem. students aren't using AI because they're lazy - they're using it because the assignment design signals that content generation is the goal rather than thinking. until we redesign for process rather than product, detection is fighting symptoms.
been saying this for two years. detection is always one step behind. Turnitin spent decades chasing essay mills and never eliminated them. Same story now with AI. The answer was never detection - it was assignment design and process documentation. If students have to demonstrate their learning in a way that can't be faked by generating text, the arms race becomes irrelevant.
the arms race framing is accurate but what it leaves out is the institutional response lag. students adapt in weeks. policy changes take years. detection tools get one update cycle per semester if we're lucky. the gap is structural and isn't closing - if anything it's accelerating.
i've stopped running detection on anything. redesigned every assignment to make it undetectable-proof. its more work up front but way less drama after.
chatgpt is free for students but my school can't afford a working projector. and we're supposed to be winning an arms race against teenagers with unlimited time and direct motivation. great.