What to Check Before You Trust an AI Detector Score

AI detectors are common in lots of jobs. Editors use them to check freelance work, hiring teams use them on cover letters, and platforms use them to check posts and reviews. They all claim to settle things with a single percentage.

Of course, they don't really do that. AI detector scores are estimates, and the same piece of text can get very different scores depending on which tool you use. Before you act on a score, whether you're evaluating someone else's text or defending your own, it's good to know what exactly that score means and how it can be inaccurate.

How AI detectors come up with a score

AI detectors look for repetitive patterns in text. One way to compute a score is to measure how predictable a piece of text is, a measure known as perplexity. AI tends to pick the most likely next word, so its writing is usually pretty predictable. Some tools also look at burstiness, meaning how much sentences vary in shape and length. Others use a classifier, a machine learning model trained to distinguish human and AI text.

Something important to remember is that no method can see how something was actually written. They compare the text they're given with patterns they've learned and give a score based on how similar it is. That's why you get a probability rather than a yes-or-no answer, and it's also why a detector can be wrong in either direction.

Why human writing gets flagged

False positives happen more than people expect, and a few causes come up again and again.

• Predictable writing. Some writing is supposed to follow a formula. Technical documentation, legal writing, product descriptions, and instructions are all formulaic and clear, which can look machine-made to a detector. In fact, some detectors have rated the U.S. Constitution as likely AI-written.
• Non-native English. In 2023, Stanford researchers found that detectors flagged a large share of writing by non-native English speakers as AI-written, while rarely flagging native speakers. Simple vocabulary and sentence structure can easily be mistaken for model output.
• Editing tools. Grammar checkers and rewrite features can even out the little quirks and peculiarities in how people naturally write. What's left is clean and uniform, which is exactly what a lot of detectors think is AI-written.
• Short passages. A few sentences aren't a large enough sample to reliably measure writing patterns, so scores on short passages tend to jump around a lot.

Even OpenAI pulled its own AI text classifier in 2023 because of its low rate of accuracy.

Check these things before trusting a score

1. Is there enough text? Be careful when scores are given for short sections of text. If the tool says it needs a minimum text length and the text doesn't meet it, the score tells you very little.
2. What does the number mean? Some tools estimate the probability that a passage was generated by AI, while others estimate how much of the text looks AI-generated. A score of 60% means different things in each case, so always check how the tool explains its result.
3. Do other detectors agree? It helps to check the text with several tools, like GPTZero, Copyleaks, and an AI detector that shows results sentence by sentence. If one tool says the score is 90% AI and another says it's mostly human, you can't base a decision on the score alone.
4. Have the tools been thoroughly tested? Look for tools that publish their test results, including how often they mistakenly flag human writing. Public benchmarks, such as HumanizerBench, provide examples of how AI detectors can be evaluated using consistent test samples and scoring methods. These comparisons can help illustrate differences in detector performance, although results may vary depending on the type of text being analyzed.
5. Could the writing style explain it? Before assuming someone used AI, consider whether the text is formulaic by nature, written by a non-native speaker, or polished through heavy editing.
6. What else do you know? A detector score should start a conversation. Drafts, a document's revision history, notes, sources, and a quick chat with the writer will often tell you more about how something was written than any single number.

If your own writing gets flagged

Being flagged is infuriating, but it's common and usually clears up. A few things help.

• Keep your process visible. If you write in a tool that saves your revision history, like Google Docs, you'll have a timeline that shows your process.
• Find out which tool was used and what it said. Then run your text through a couple of other detectors. If they don't agree, you can use the contradictory results to push back on that one score.
• Describe your writing style. If you work in a formulaic field, lean on grammar tools, or write in a second language, say so. Those are all known causes of false positives.
• Offer to talk it through. Walking someone through your sources and choices is often the quickest way to clear things up.