How AI Detection Actually Works – A Teacher Explainer
There’s a lot of confusion in teacher circles about what AI detection tools are actually doing under the hood. I put together a breakdown based on what’s publicly known, and what I’ve learned from testing these tools in my CS classes.
AI detectors work primarily by measuring two things: perplexity and burstiness. Perplexity measures how “surprised” a language model is by each word choice – AI-generated text tends to use predictable, high-probability word choices consistently. Burstiness measures variation in sentence length and complexity – human writing is “bursty” (mix of long and short, complex and simple), AI writing is more uniform.
The problem is that neither of these is a reliable standalone signal. Skilled human writers often write predictably. ESL students write uniformly. Academic writing conventions produce low burstiness by design.
The bottom line for teachers: these scores are probabilistic indicators, not forensic evidence. Turnitin and GPTZero are working with the same fundamental limitations. neither of them can tell you with certainty whether a student used AI – they can only tell you that the text has statistical properties similar to AI-generated text.
i’m happy to answer questions about the technical side.
6 Replies
Join the discussion.
Log In to Reply
I've been teaching for 15 years and Turnitin has gone through several iterations of this - plagiarism detection had similar accuracy debates in the early days. The technology does improve. Give it another 2-3 years and the false positive rates will come down significantly as the training data matures. For now, use it as one data point among several.
this is exactly the framing my department needed to hear. admin told us to use Turnitin for detection and also told us we can't formally accuse a student based on the score alone. so... what exactly are we doing with it? a probabilistic indicator that can't be used as evidence isn't a policy tool, it's a vibe.
that's exactly my concern with Liam's point - if the model doesn't know what human-written academic text looks like in a given discipline, it defaults to flagging anything confident and structured. most experienced students write confidently and structurally. the model punishes exactly what good academic writing looks like.
exactly. the policy and the tool are operating on different assumptions. the tool says "here is a probability." policy needs to say "here is a threshold." most schools haven't done that second part, which is why everyone is confused about what to do with a 78% score.
ok so ive been misreading these scores for months lol. i thought higher burstiness was bad. this changes everything about how ive been interpreting my results. thanks for actually explaining this.
useful explainer, sharing with my department. the part about perplexity scores is something most of us have never seen explained in plain language. that alone changes how im reading results.