Are AI detectors accurate?
AI detectors are much better than guessing at spotting AI-generated text, but none are accurate enough to prove on their own that someone used AI. Every detector flags some human writing and misses some AI text.
Updated November 21, 2026
What does accuracy mean for an AI detector?
A detector can get a text wrong in two ways: flagging human writing as AI (a false positive), or letting AI-generated text slip through (a false negative). A single accuracy percentage hides which error is happening.
That number also depends heavily on the test set: how long the passages are, what kind of writing they are, which models generated them, and where the threshold sits. Strong results on one benchmark don't automatically carry over to a different kind of document.
Why can a 99% accuracy claim mislead?
Most of the writing people check is human, so even a small false-positive rate adds up fast. Say a detector flags 1% of human writing and catches 90% of AI writing. In a batch of 200 essays where 10 were AI-generated, it catches 9 of the 10 AI essays and flags about 2 of the 190 human ones. That means roughly 2 of the 11 flagged essays (nearly one in five) were actually written by a person.
The rarer AI-generated text is in your pool, the bigger that false-alarm share gets. That's why false-positive rate matters far more than a headline accuracy number.
When are AI detectors less reliable?
- Short passages. Accuracy drops sharply on a few sentences. Sapling's detector becomes much more reliable past about 50 words, and the API flags short or list-formatted inputs as low confidence.
- Generic or formulaic writing. Standard boilerplate, templates, and five-paragraph-style essays look statistically similar to AI output and trigger more false positives.
- Paraphrased or edited AI text. Human edits or paraphrasing shift phrasing away from typical model output, increasing false negatives.
- Partly AI-written documents. A mostly human document with one AI-generated section can receive a low overall score even though specific sentences flag high.
- Unsupported languages. Sapling uses dedicated models for English and 15 other languages. Text in other languages falls back to the English model and is less reliable.
- Non-native English writers. A 2023 Stanford study found that the detectors it evaluated consistently misclassified writing samples by non-native English writers as AI-generated, while samples by native writers were identified accurately.
What does the research say?
OpenAI reported that its own AI text classifier, released in early 2023, correctly identified 26% of AI-written text while labeling 9% of human-written text as AI-written. OpenAI withdrew the classifier in July 2023, citing its low rate of accuracy.
Both results date from 2023, and detectors and language models have changed substantially since then. That's another reason to test a detector on your own material rather than relying on an older published figure.
How accurate is Sapling's AI Detector?
Accuracy depends on your dataset, text length, and chosen threshold. Sapling's technical whitepaper reports evaluation results on a 1,000-sample held-out set, breaking down false positives and false negatives. You can request the whitepaper from the AI Detector page. As with any benchmark, check the model version and dataset in the report before assuming the same numbers hold on your own documents.
How can I check an AI detector myself?
- Pull together human-written and AI-generated samples you know the origin of, matching the lengths, topics, and languages you deal with.
- Run them through the detector and tally both mistakes: human writing that gets flagged, and AI text that gets missed.
- Record raw counts alongside percentages, along with the date, detector version, and threshold you used.
The guide to reading and evaluating AI detector scores has the full method and the metrics to report.
How should an AI detector score be used?
Use a detector score as a reason to look closer, never as the sole basis for a decision that affects someone. A high score can prompt a second read, a check of drafts or revision history, or a conversation with the author.