How AI detectors work — and why they disagree
What AI detectors actually measure, why two tools can give the same text very different scores, and how to read a detection result sensibly.
AI detectors promise to tell you whether a piece of text was written by a person or a language model. Paste the same paragraph into three detectors, though, and you may get three different answers. Here's why — and how to use these tools without being misled.
What detectors look for
Most detectors are themselves machine-learning models trained on examples of human and AI writing. They learn statistical patterns, such as:
- Predictability: language models tend to choose likely next words, so AI text is often more predictable than human text.
- Uniformity: AI sentences are often similar in length and structure; human writing is burstier.
- Vocabulary habits: certain phrases and transitions appear far more often in model output.
- Polish without specifics: fluent text that makes general claims but few concrete, personal details.
Why scores disagree
- Different training data: each detector learned from different examples, and newer AI models write differently from older ones.
- Different thresholds: one tool calls 60% “likely AI”, another needs 90%.
- Short texts are hard: a few sentences don't contain enough signal for a confident answer.
- Edited text is mixed: when a person revises an AI draft, or uses AI to polish their own writing, the result really is a blend.
False positives are real
Detectors can flag human writing as AI — especially formal, formulaic or very polished text, and writing by people using a second language. A high score is a signal to look closer, not proof of who wrote something.
Never make a high-stakes decision — a grade, a hiring choice, an accusation — from a detector score alone.
How to read a result
- Look at the sentences, not just the total. Sentence-level scores show you where the text reads as generated.
- Consider the context: the type of writing, its length, and how it was produced.
- Compare drafts over time rather than relying on a single number.
- Treat the score as a writing prompt: the flagged sentences are usually the ones that would benefit from more specific, personal wording anyway.
How TrulyWrite's detector helps
TrulyWrite colours every sentence green, amber or red and gives an overall estimate, so you can see exactly which lines read as machine-made. From there you can rewrite just those sentences with the Humanizer — and keep everything you wrote yourself exactly as it is.