AI Detectors: How They Work and Why People Like Karpathy Don't Trust Them
By Venkata Anirudh Devireddy · Endoblog.dev
Every student has had the same moment. You write something yourself, run it through Turnitin out of paranoia, and it comes back flagged as partially AI generated. You didn't use AI. The tool is wrong. And there's no clean way to prove it.
That happens more than people think, and it's worth understanding why.
How these tools actually decide something is AI written
Turnitin's detector doesn't read your essay and understand it. It looks at patterns. Two matter most.
The first is perplexity. This measures how predictable each word is given the words before it. Human writing tends to make less predictable choices. AI models tend to pick the statistically likely next word, so AI text often scores lower on perplexity. Lower perplexity, higher suspicion.
The second is burstiness. Human writing naturally varies in sentence length and rhythm. Some sentences are short. Some run long. AI text used to be more uniform, evenly paced, similar sentence lengths throughout. Detectors were built to notice that evenness and treat it as a signal.
The problem is both signals break down constantly. A student with a plain, direct writing style scores low on perplexity for no reason related to AI. A non-native English speaker who was taught formal, structured sentence patterns in school often writes with less burstiness naturally. Neither is using AI. Both get flagged.
Turnitin says it's accurate. The independent numbers tell a messier story
Turnitin claims 98% accuracy with under 1% false positives. That number comes from Turnitin's own internal testing on curated samples, not independent verification. Stanford's Human-Centered AI institute ran its own test and found detectors flagged 61% of essays written by non-native English speakers as AI generated, compared to a much lower rate for native speakers. That's not a small gap. That's a tool that fails one group of students at a dramatically higher rate than another.
Vanderbilt disabled Turnitin's AI detector entirely in 2023. Other universities followed. Not because the tool never worked, but because the cost of a false positive is too high to justify the tool's error rate. An accusation of academic dishonesty can follow a student for the rest of their academic career. A coin flip's worth of uncertainty isn't good enough for that.
Why Karpathy and people like him don't trust these tools
Andrej Karpathy has said plainly that AI detectors don't work reliably, and he's not alone among AI researchers in saying it. The reasoning is straightforward once you understand what these detectors are actually measuring.
They're not detecting AI. They're detecting a statistical resemblance to AI output, based on stylistic features that also show up naturally in human writing. Simple sentence structure, formal tone, or a plain, unembellished style can all trip the same signal an AI model would produce. There is no ground truth check happening. There's a classifier trained on examples, guessing.
Karpathy has also pointed out that these classifiers get gamed constantly. Run AI text through a paraphraser or a "humanizer" tool, and the detector's accuracy drops fast. So the tool ends up catching people who wrote something plain and honest, while missing people who deliberately dressed up AI text to slip past it. That's close to the worst possible outcome for a detection system built to punish dishonesty.
What this actually means if you're a student
Don't assume a detector flag means you did something wrong, and don't assume a clean scan means you're safe if you did use AI dishonestly. The tool isn't reliable enough to be treated as a verdict either way.
If you get flagged and you wrote the work yourself, ask your professor how the score is being used and whether it's the only evidence being considered. Most instructors know the tools are imperfect. Keep your drafts, your revision history, anything that shows the writing process. That's stronger evidence than any detector score, in either direction.
The bigger point is that academic institutions adopted these tools fast, before the tools were actually reliable enough to justify the trust being placed in them. That gap hasn't closed yet. Until it does, a flagged score should be a conversation starter, not a conviction.
