Module 6 · Lesson 6.2

How to Read AI Detector Scores

2 min read · Updated

"87% AI." It looks like a precise measurement. It isn't. Most detector scores are confidence estimates, and they're easy to misread in ways that cause real harm. This lesson explains what the numbers mean, what they don't, and what to do at each range.

What "87% AI" actually means

There are two common ways to read that number, and only one is usually right:

How people read it What it usually means
87% of the text was written by AI The detector is 87% confident the text, as a whole, matches its idea of AI writing
There's an 87% chance a person cheated The text has features the detector associates with AI. It says nothing about how it was written

Some tools do estimate the share of sentences that look AI-written, and they say so in their results. Check which kind of score you're looking at before you react to it. Either way, a score describes the style of the text, not its history.

What to do at each range

Exact cut-offs differ between tools, but this is a reasonable way to act on a result:

Score What it suggests Sensible next step
0–20% Reads as human to this tool Nothing to do. Many hand-written texts land here, not at zero.
20–60% Uncertain. Common for edited AI drafts and for formal human writing Look at which sentences were flagged. Check them against the 12 tells.
60–100% Strong match to AI patterns If it's an AI draft, run the 10-minute edit. If you wrote it yourself, see lesson 6.3.

Sentence highlights are more useful than the total

Most tools highlight which sentences pushed the score up. That's the part to study. Usually you'll find the highlights cluster in a few places: the introduction, a list, a summary paragraph, or anywhere you used a stock phrase.

The AI detector on this site takes this further. It shows the exact phrases it noticed, a chart of your sentence lengths and your most repeated words, so you can see why it scored the way it did.

Don't over-read small differences

  • Scores move around. The same text can shift by 10 points after a minor model update.
  • Short texts are unreliable. Under about 250 words, treat any score as close to a guess.
  • Tools disagree. A gap of 40 points between two tools on the same text is normal. That's why testing on several helps.
  • Low isn't proof either. A low score doesn't prove a person wrote something. Edited AI text can score low, and that's partly what this course teaches.
If you're a teacher, editor or manager reading this: a detector score alone isn't evidence. Even detector makers say their results should start a conversation, not end one. Ask the writer about their process and look at drafts or version history before drawing a conclusion.
Try this

Go back to the log you started in lesson 6.1. For each result, note which range it falls in and whether the tools agreed. Then look at the highlighted sentences on the highest-scoring result and mark which of the 12 tells each one shows.