Detectors look for patterns, not proof

Most AI detectors are classifiers: models trained to tell apart text written by people and text produced by language models. They don't know who wrote your paper. Instead, they look at statistical signals in the text itself.

Two ideas come up often:

• Predictability. Language models tend to choose likely, "safe" words. Text where almost every word is the most expected next word can look machine-generated.
• Variation. Human writing usually mixes short and long sentences, changes rhythm and takes small detours. Very uniform text can raise a detector's score.

Neither signal is proof. A careful human writer producing a formal, tidy paragraph can look "predictable," and a lightly edited AI draft can look varied.

Why the score is a probability

When a detector shows a percentage, it's expressing how confident the model is, based on the patterns it learned. That's why:

• the same text can get different scores from different tools;
• short passages are much harder to judge than long ones;
• false positives happen, and research has found that some detectors flag writing by non-native English speakers more often.

Even OpenAI withdrew its own AI text classifier in 2023, citing low accuracy. Detection is genuinely hard.

How to use a score sensibly

A detector is most helpful as a revision tool, not a judge:

1. Look at which passages are flagged, not just the overall number. Flagged sections are often the most generic ones.
2. Add specifics. Your own examples, course concepts and opinions make writing both stronger and more clearly yours.
3. Read it out loud. Sentences that sound stiff when spoken are worth rewriting, no matter where they came from.
4. Know your course's rules. Instructors set different policies on AI use. Following them matters more than any score.

Why we care about this

We build Draftout — AI writing toolkit for students (https://draftout.com). Our AI Detector highlights the passages most likely to read as AI-generated, so students can see where to focus their revisions before they submit. In future posts we'll write more about how writing tools work and what we learn from building one.