Get started

How do AI detectors work?

An AI detector is a classifier trained on human-written and AI-generated text. Given a new passage, it estimates how likely that writing is to have come from a language model.

Updated November 21, 2026

What does an AI detector measure?

A detector doesn't know who wrote a text. Instead, it measures how closely a passage matches statistical patterns it learned from AI-generated examples, and reports the result as a probability.

Language models generate text by picking each word from a probability distribution over what tends to come next. Over a paragraph or essay, those choices leave subtle patterns in phrasing, sentence rhythm, and structure that a trained classifier can spot even when a human reader can't.

What approaches do AI detectors use?

  • Trained classifiers. A neural network (usually a Transformer) is trained on large datasets labeled as human-written or AI-generated. Most current detectors work this way, including Sapling's.
  • Statistical measures. Perplexity measures how predictable a text looks to a language model, and burstiness measures how much that predictability shifts from sentence to sentence. AI text tends to score low on both. They're cheap to compute, though less reliable on their own than a trained classifier.
  • Watermarks. Some AI providers embed a subtle statistical signal in text as it's generated (Google's SynthID is one example). That check only works on output from models that actually apply the watermark.

How does Sapling's AI Detector work?

Sapling's AI Detector uses a Transformer classifier that estimates, for each token (roughly a word), the probability that it is AI-generated. The overall score, sentence scores, and highlighted passages all come from that same model, so they sit on the same scale.

Language detection runs automatically: English text goes to an English-specific model, and 15 other languages go to a multilingual model trained on human and AI writing.

We retrain the detector regularly on outputs from current models, including GPT-6, Claude 5.5, Gemini 3, and DeepSeek-V3.2. The same model powers both the free AI Detector and the AI Detector API.

What does an AI detector score mean?

A score is an estimated probability that the text is AI-generated, not the share of words written by AI. An 80% score means the model thinks there's an 80% chance the passage is AI-generated, not that 80% of its words came from an AI.

To turn a score into a label, a detector applies a threshold. Setting it higher flags less human writing but misses more AI text; setting it lower does the opposite. The guide to reading and evaluating AI detector scores walks through that tradeoff in detail.

Why do AI detectors make mistakes?

  • Too little text to judge. A couple of sentences don't give a classifier much statistical evidence to work with. Sapling's detector becomes much more reliable once a passage is past roughly 50 words.
  • Predictable human prose. Templates, bulleted lists, and formulaic essays share the same high-probability phrasing that language models favor, which raises false positives.
  • Heavy human editing. When someone rewrites or edits AI output, they break up the patterns the model looks for, making the text easier to miss.
  • Human and AI text mixed together. If only one section of a document was generated by AI, the overall score can stay low even while individual sentences score high.
  • Newer language models. A classifier trained on older models still beats random guessing on newer ones, but accuracy drops until it's retrained on recent outputs.

See Are AI detectors accurate? for what this means in practice.

How should an AI detector score be used?

Treat a score as one input to a review, never as proof on its own. No AI detector, including Sapling's, can definitively prove whether a person or a model wrote a passage. Read the flagged sections, check drafts or revision history if you have them, and talk with the author about how the piece was written.