How to Read an AI Detection Report

An AI detection report is a probability estimate, not a verdict. The headline percentage is the detector's confidence that the text is machine-generated — so "78% AI" means the model is fairly confident the text is AI-written, not that 78% of the words are AI or that you have 78% proof of misconduct. To read a report correctly, look past the headline number to the sentence-level highlights, the confidence band, and the sample size, and always combine the report with your own judgment.

This guide breaks down every element of a typical report and shows you how to interpret it without over- or under-reacting.

The Three Things People Get Wrong

Before the walkthrough, clear up the three most common misreadings:

1. "It's a percentage of the text." It is not. A 60% score does not mean 60% of sentences are AI. It is one probability for the whole sample (or for a passage).
2. "It's a measure of certainty." A score is a model's estimate. The model can be wrong, and it is more often wrong on short, formal, or non-native English text.
3. "A number can decide an outcome." No responsible workflow lets a detector score alone determine a grade, a firing, or a rejection. The report is evidence; a human weighs it.

If those three are clear, the rest of the report becomes much easier to read.

Anatomy of an AI Detection Report

A typical report from the xeviora AI Detector has four parts. Read them in this order — not top to bottom.

1. The headline probability

This is the big number: an AI-generation likelihood from 0 to 100%. It is useful as a triage signal — it tells you whether to look closer — but it is the least precise part of the report. Note it, then move on.

2. The confidence band

Better detectors do not just give a number; they give a band or label: "likely human," "uncertain," "likely AI." This matters because the same 55% score in a narrow, high-confidence band means something different from 55% in a wide, low-confidence band. The "uncertain" zone is real and honest — respect it instead of forcing a black-and-white call.

3. The sentence-level highlights

This is the most valuable section. The report color-codes or flags individual sentences by how strongly they pushed the score up. Read these flagged sentences directly. Do they show the tells described in how to check if text is AI-generated — uniform rhythm, hedging filler, no specifics? If the flagged sentences look genuinely templated, the score has substance. If they look like ordinary careful writing, be skeptical of the headline.

4. The metadata line

Sample length, language detected, and sometimes the metrics behind the score (perplexity, burstiness). This line tells you how much to trust everything above it.

How to Interpret the Score by Band

Use this as a calibration table. The exact thresholds vary by tool, but the behavior is consistent.

The middle row is where mistakes happen. A 55% score is the detector telling you it does not know. The correct response is more investigation, never a snap judgment.

Why Sample Size Changes Everything

Detection statistics need text to work with. A 40-word answer gives the model almost nothing to measure, so its score is volatile and unreliable. The same model on a 1,200-word essay is far steadier.

Practical rule: be very cautious with any report on under ~150 words, and treat reports on under ~50 words as little more than noise. If you must assess short text, gather more samples from the same author rather than over-reading one tiny score. This sample-size sensitivity is one reason detectors disagree, explored further in how accurate are AI detectors.

Reading a Report When the Stakes Are High

For academic or employment decisions, change how you read the report entirely.

• Never quote the percentage as a fact. Say "the detector flagged several passages," not "the report says it's 80% AI."
• Lead with the highlights. Bring the specific flagged sentences to a conversation with the author. Ask them to walk through how they wrote those passages.
• Cross-check. Run the text through at least one other detector. Agreement across tools is meaningful; disagreement is a clear signal to slow down.
• Account for known false positives. Non-native English writers, formulaic genres (lab reports, legal summaries), and naturally concise writers are flagged more often. See why AI detectors flag human writing.

Educators can see a realistic version of this process in can teachers detect ChatGPT, and our Solutions for students and educators covers fair-use policy alongside detection.

Reading the Report on Your Own Writing

If you ran the report on something you wrote and it came back high, work through it calmly:

1. Look at the highlighted sentences, not the number. Which passages drove the score?
2. Diagnose the pattern. Highlights often cluster in formal, list-heavy, or definition-style sections — the genuinely templated parts of human writing.
3. Decide whether to revise. If you want honest writing to read less mechanically, an AI Humanizer varies sentence length and phrasing on those specific passages so they stop matching machine patterns. Our guide on why AI writing sounds robotic explains the underlying patterns you are trying to break.

A flagged report on your own work is usually a style signal, not an accusation — your formal voice happens to overlap with how models write.

What a Good Report Should Always Tell You

When choosing or trusting a detector, look for reports that include:

• A probability and a confidence band — not just a bare number.
• Sentence-level highlights so you can verify the score yourself.
• The sample length and language so you know the reliability.
• Honest "uncertain" labeling instead of forcing every result into AI/human.

A report that gives you only a single percentage with no breakdown is asking you to trust a black box. A good report shows its work.

The Bottom Line

Read an AI detection report from the inside out: highlights first, confidence band second, headline number last, with sample size as the reliability check on all of it. The percentage starts the investigation — your reading of the flagged sentences, the context, and a second opinion finishes it. Ready to generate a report? Run your text through the AI Detector, then use how to check if text is AI-generated to act on what you find.