How Is Your AI Percentage Calculated? The Math Behind Scores

How is your AI percentage calculated? Short version: a classifier reports how confident it is that your text matches machine-written patterns, and that confidence gets dressed up as a percentage. It is not a count of how many words an AI wrote, and that one misunderstanding causes more damage than any flaw in the models themselves.

Here is how an AI percentage is calculated in practice: the detector splits your text into tokens, measures how predictable each token is to a language model, compresses those measurements into features like perplexity and burstiness, and feeds everything to a trained classifier. The classifier outputs a confidence value, which gets rescaled into the percentage you see.

The quantity behind the number

A detector never observes who typed the words. It observes statistics. Language models write by repeatedly picking a likely next token, and that habit leaves a signature: text where nearly every word is the word a model would have guessed anyway.

So the score answers one narrow question. How strongly does this text's statistical profile resemble the machine-generated writing in my training data? The percentage is the classifier's confidence in that resemblance, stretched onto a 0 to 100 scale. It is not a census of your sentences.

You can inspect the raw ingredient yourself. GLTR, a 2019 research demo from the MIT-IBM Watson AI Lab and Harvard NLP, colors every token by how predictable it was to a language model. Paste in human writing and the page blooms with unlikely word choices; paste in model output and it runs a smug, uniform green. Commercial tools have moved far past GLTR, but the intuition survives intact. Our guide to how AI detectors work covers the model families in depth; this article stays on the math that turns prose into a number.

How is AI percentage calculated, step by step

Every serious text detector runs some version of this pipeline:

1. Tokenization. Your text is chopped into tokens, the word fragments language models actually read.
2. Probability scoring. A reference model estimates how probable each token was given everything before it. After "the powerhouse of the," the word "cell" is cheap; "harmonica" costs a fortune.
3. Feature aggregation. Token probabilities are compressed into document-level features. Perplexity is the average surprise across the text. Burstiness is how much that surprise varies from sentence to sentence. Many tools stack structural signals on top, like sentence-length rhythm and phrase frequency.
4. Classification. The features, or in end-to-end neural systems the raw text itself, go through a trained classifier that outputs a raw score leaning human or machine.
5. Calibration. That raw output is squashed into a 0 to 1 range and calibrated against validation data, so that a 0.8 tracks something like "80% of texts scoring here were AI" rather than an arbitrary dial position. Multiply by 100 and ship it.
6. Aggregation. Long documents are scored in windows or sentences, then combined into the single number on your screen.

Six steps, and at least four involve judgment calls: which reference model, which features, which calibration data, which combination rule. Hold that thought for the section on why tools disagree.

Why 60% doesn't mean 60% of your sentences

Two very different quantities get displayed under the same "AI percentage" label. Some tools report confidence that the document contains AI text, so 60% means "we lean AI, without much conviction." Others, Turnitin included, estimate the share of the document that qualifies as likely AI-written, so 60% means "roughly six of every ten words sit inside machine-flavored segments, we think." Same digits, different claims, and the interface rarely tells you which one you're looking at.

Either way, the segment-level estimates carry their own error. Turnitin claims under 1% false positives at the document level, a figure that applies to documents scoring at least 20% AI, while acknowledging roughly 4% error at the sentence level. Zoom in and the picture gets noisier. That is exactly why a single document number hides so much.

It is also why AI Detector 360 scans include a sentence heatmap instead of one lonely digit: you see which sentences carry the statistical signal and how strong it is, rather than guessing at what an aggregate hides. And if you're wondering at what point a number should worry you, we've answered that separately in what AI percentage is actually concerning.

Thresholds, where a number becomes a verdict

Behind every "flagged" banner sits a cutoff, and cutoffs are chosen, not discovered. Set the threshold low and you catch more AI at the price of accusing more humans. Set it high and you protect innocent writers while waving more synthetic text through. There is no neutral setting, only trade-offs someone selected on your behalf.

OpenAI's own classifier is the cautionary tale here. At its chosen operating point it caught just 26% of AI-written text while false-flagging 9% of human writing, and OpenAI retired it in July 2023 for "low accuracy." Turnitin leaned the other way, gating its headline false-positive claim to documents scoring at least 20% AI. Neither choice is dishonest. Both are reminders that the percentage and the verdict are separate machines, and the second one is tuned by product decisions, not physics. The documented record of both failure modes lives in how often detectors get it wrong.

Why two tools score the same essay differently

Paste one essay into three detectors and you might get 12%, 48% and 91%. Nobody lied to you. The tools made different engineering choices at every stage of the pipeline:

Design choice	How it moves your number
Reference model	An older reference makes newer AI text look more human, deflating scores
Training mix	A detector trained on essays misreads fiction, code or marketing copy
Feature weighting	Perplexity-heavy scoring punishes plain, predictable human prose
Calibration data	Tools calibrated on native English inflate scores for second-language writers
Decision threshold	The same raw signal clears one bar and trips another
Windowing	Different chunk sizes spread signal differently across a document

The calibration row is not hypothetical. A 2023 Stanford study in Patterns found seven detectors flagged an average of 61.3% of TOEFL essays by non-native English speakers while handling US 8th-grade native essays almost perfectly, largely because careful second-language writing is statistically predictable. That failure, and its cousins, are unpacked in why human writing gets flagged.

Reading your score like an analyst

A short field manual for the moment a percentage lands in front of you:

• Check the length first. Below a couple hundred words, every stage of the pipeline is starving for data, and the output is noise with confident formatting.
• Find the confidence label. A well-built tool tells you when it isn't sure. If yours doesn't, assume it isn't.
• Compare engines, not numbers. An 85 from one product and a 60 from another are not on the same scale. Agreement between independent engines means far more than either digit alone.
• Go to the sentence level. Signal concentrated in one boilerplate paragraph reads very differently from signal spread across the whole document.

That workflow is what AI Detector 360 automates: multiple engines per scan, an explicit confidence level, and a heatmap showing where the signal sits, with the calibration logic published on our methodology page. You can try it on any text up to 5,000 characters with the free AI detector, no sign-up required, and watch how the math in this article behaves on your own writing.

A percentage is the end of a long chain of modeling decisions. Treat it like an instrument reading: useful, repeatable, and only meaningful once you know what the instrument actually measures.