AI Detectors: What They're Actually Measuring (It's Not What You Think)

Most AI detectors aren't checking whether a human wrote something. They're measuring something much more specific — and more gameable than people realize.

The core concept behind almost every commercial AI detector is something called perplexity. In plain terms, it's a measure of how "surprising" a piece of text is. Language models assign probabilities to every word in a sequence — how likely is this word to appear after the previous ones? Low perplexity means the text was very predictable. High perplexity means it surprised the model.

AI-generated text tends to have low perplexity. That's by design. Models are optimized to produce probable, coherent output. A model almost never writes something it wasn't sort of expecting to write. Human writers, on the other hand, make weird choices. They use unexpected words, change direction mid-sentence, and include things that feel idiosyncratic. That's high perplexity.

The second metric: burstiness

Human writing isn't just unpredictable on average — it's unpredictable in an uneven way. Some sentences are very surprising. Others are completely banal. This variation in sentence-to-sentence perplexity is called burstiness, and it's something AI struggles to replicate naturally.

When you generate text with a model, the perplexity tends to stay in a narrow band. It's predictably unpredictable, which is itself a pattern. Detectors measure this variance and flag text where it's too consistent.

Where detectors fail

Here's the problem: these metrics are proxies for "AI-ness," not direct measurements of it. A highly technical academic paper written by a human might have very low perplexity — because technical writing is constrained and predictable. A student who learned English as a second language might write with patterns that look AI-like because they're careful and formal.

False positives are a real issue. Multiple studies have shown that detectors flag non-native speakers at significantly higher rates than native speakers — which has obvious equity implications when these tools are used to police student work.

What actually works

The honest answer is that no detector is reliable enough to use as sole evidence of anything. The good ones hedge heavily ("likely AI-generated," not "definitely"). The bad ones don't.

What actually catches AI writing is human reading. An instructor who knows a student's voice will notice something's off. A colleague who reads your drafts will feel the shift. The statistical tools are useful as a screening layer, but they're not the final word — and treating them as such causes more problems than it solves.