No AI detector is right every time, including ours. This guide is candid about where detection goes wrong, which kinds of writing are most likely to be misread, and how to use a result so that a mistake doesn't turn into an unfair accusation.
Key takeaways
- Every detector makes mistakes in both directions: false positives and misses.
- Formulaic, templated, very short and second-language writing are the most likely to be misread.
- A score is evidence to weigh, never proof on its own.
- Drafts, version history and a conversation with the writer are the best checks on any result.
Why detectors make mistakes
A detector doesn't know who wrote a text. It looks at patterns, such as how predictable the word choices are and how sentences are built, and estimates how closely they match writing it learned from people and from AI models. Human and machine writing overlap. Some people naturally write in a plain, even, predictable way. Some AI output gets edited until it reads like a person. Wherever the two overlap, a detector can get it wrong.
Mistakes go both ways. A false positive is human writing flagged as AI. A miss is AI writing that reads as human. False positives tend to do more harm, because they can lead to an accusation against someone who did nothing wrong.
Writing that's most likely to be misread
Some kinds of writing sit closer to that overlap than others. These are the situations where we'd urge the most caution before reading anything into a result:
- Formulaic writing. Five-paragraph essays, lab reports and other assignments taught to a strict template follow predictable patterns that resemble AI output.
- Non-native writers. People writing in a second language often use simpler, safer vocabulary and textbook sentence structures, which can look machine-like.
- Very short texts. A few sentences don't contain enough signal for a confident result in either direction.
- Heavily templated genres. Cover letters built from a template, product descriptions, compliance language and standard business emails are uniform by design.
- Tool-polished writing. Human text run through grammar or rewriting tools can pick up machine-like patterns. In Winston AI this often shows as yellow, mixed-signal highlights.
Picture a student who learned English as an adult and writes carefully, leaning on phrases from a textbook. Their essay may come back with several red sentences even though every word is theirs. The highlights are a reason to ask how they wrote it, and their drafts will usually settle the question quickly. Without that step, a tool meant to help becomes a source of harm.
How to use results responsibly
- Read the highlights, not just the number. Check which sentences were flagged and whether they match what you noticed yourself.
- Consider the writer and the genre. Is this a second-language writer, a template-driven format or a very short piece? Adjust how much weight you give the result.
- Look for process evidence. Drafts, notes, version history in a shared document and earlier work tell you far more about how something was written.
- Talk to the writer. Ask them to walk you through their choices. People who wrote their own work can usually explain it.
- Match the evidence to the stakes. The more serious the possible consequence, the more you need beyond a score.
- Document your reasoning. Write down what you found and why, not just the score.
What we do on our side
Winston AI scores every sentence so you can see where a result comes from instead of trusting a single number. Yellow highlights separate mixed signals from clear ones, which helps with human writing that a tool has polished. We retrain as new model versions ship. None of that makes the tool infallible, and we'd rather say so plainly than have anyone treat a result as a ruling.
If you've been flagged and you didn't use AI
Gather what shows your process: drafts, outlines, notes, your research trail and document version history. Offer to talk through the work. Our guide on how to prove you did not use AI goes through this step by step.
If you review writing for a living, run a check in the AI detector and then follow the steps above. Teachers can find a full review routine on our page for teachers.