ZeroTell's AI detector reads your text sentence by sentence, highlights the sentences that look most like machine-generated writing, and combines those signals into an overall percentage for the document. It is a useful way to see which parts of a text resemble AI output, but like every detector it works on probability, so it can be wrong. We do not publish an accuracy percentage, because no single number honestly describes performance across every kind of writing.
This page is written by the team that builds ZeroTell. We want you to understand what the tool does, where it falls short, and how to use the result sensibly.
How ZeroTell scores your text
When you paste a document into the detector, it goes through a few stages.
- Splitting. The text is broken into sentences, because a single document often mixes styles. An introduction might be generic while the body is full of personal detail.
- Sentence scoring. Each sentence is assessed for patterns common in AI-generated writing, such as very predictable word choices, uniform rhythm, and stock phrasing. Sentences that look strongly AI-like are highlighted.
- Document score. The sentence scores are averaged into one percentage for the whole document, with longer sentences counting for more. That number is an overall estimate of how AI-like the text reads, not a measure of how much of it a machine literally wrote.
The highlights are the most useful part. They show you exactly which sentences pushed the score up, so you are never left guessing why a number appeared.
Why we don't publish an accuracy number
You will see detectors advertise striking accuracy figures. We have chosen not to, for a simple reason: accuracy is not one number.
How well any detector performs depends on:
- Length. Long samples are far easier to judge than short ones.
- Editing. Raw chatbot output looks very different from text a person has rewritten.
- Genre. A personal essay and a chemistry lab report have very different "normal" patterns.
- Writer background. Second-language writing often uses safer, more predictable structures.
- Model. Text from newer or less common AI models may look less familiar.
A figure measured on one carefully chosen test set tells you little about how the tool will behave on your text. We would rather show our reasoning sentence by sentence and be clear about the limits.
Known limits of the ZeroTell detector
Being honest about weaknesses is part of using a detector responsibly. Here is where you should be most careful with our results.
Short texts
With only a few sentences, there is very little signal to work with. Scores on short inputs can swing a lot with small wording changes. Treat them as rough at best.
Formulaic writing
Templates produce predictable text. Abstracts, methods sections, legal boilerplate, product specs, and tightly structured school essays can be flagged even when a person wrote every word.
Non-native English writing
If you write in English as a second or third language, you may lean on familiar vocabulary and clear, standard sentences. That is good practice, but it can resemble the safe phrasing language models produce. A flag on this kind of writing deserves extra caution.
Mixed and edited text
Many documents involve several tools: a grammar checker, a translation tool, maybe an AI brainstorm. Mixed text often produces middling scores and scattered highlights. That reflects genuine ambiguity, not a hidden truth.
Lists, definitions, and facts
Sentences that state widely known facts or define standard terms tend to read the same no matter who writes them. Do not be surprised to see them highlighted.
AI DetectorSentence-level AI scoreHow to read your results
Here is a step-by-step way to interpret what you see.
- Check the length first. If the text is short, gather more before taking the score seriously.
- Scan the highlights before the percentage. Ask what the flagged sentences have in common. Are they generic openings? Definitions? Transitions?
- Compare with the rest of the text. Are highlights clustered in one section or scattered? Clusters can point to a part that was written differently.
- Consider the genre. A formulaic document will naturally score higher than a personal reflection.
- Think about the writer. Is this consistent with how they usually write? Do they have drafts or notes?
| Pattern in your result | Likely explanation |
|---|---|
| One section heavily highlighted, rest clear | That section was written or edited differently; worth a closer look |
| Light highlights spread throughout | Generic phrasing or formulaic style; often human |
| High percentage on a short sample | Not enough data; unreliable |
| Only definitions and facts flagged | Standard content that reads the same for everyone |
If ZeroTell flags your own writing
First, do not panic. A flag is not an accusation, and it says nothing about your intent.
Look at the highlighted sentences. Often they are the weakest parts of your writing anyway: broad openings, filler transitions, or summaries that repeat what you already said. Rewriting them with specific examples, your own reasoning, and concrete details tends to improve the piece. That is worth doing for the reader, not to chase a lower number.
Keep evidence of your process. Saved drafts, version history in your word processor, outlines, and research notes are the strongest way to show your work is yours. Our guide on what to do if your writing is flagged covers this in more detail.
What the detector does not know
It helps to be clear about what information the detector never sees. It does not know who you are, what class the text was written for, how long you spent on it, or which tools you used along the way. It does not compare your text against your earlier assignments, and it cannot read your browser history or your drafts. All it has is the words on the page.
That is exactly why context matters so much. A reviewer who knows the writer, the assignment, and the drafting process will always understand the situation better than a model that only sees a block of text. The detector can tell you that a sentence reads like typical AI output. Only people can decide what that means in a particular case.
It also means results can change if the text changes. Adding a paragraph, removing a quote, or fixing a typo can shift sentence scores slightly, and the overall percentage with them. Small movements like that are normal and should not be read as meaningful.
Using the humanizer responsibly
ZeroTell also offers an AI humanizer. It rewrites stiff or robotic text so it reads more naturally, which can help when you have drafted with AI assistance and want clearer, more personal prose. It is an editing aid, not a way around anyone's rules. If your school or employer limits AI use, those rules still apply to text you have humanized, and you are responsible for following them and disclosing AI use where required.
For teachers and reviewers
If you use ZeroTell to review student or staff work, please treat the result as a starting point. Read the highlighted sentences yourself. Ask for drafts. Talk to the writer. Consider language background and assignment type. A detector can point you toward a passage worth discussing; it cannot tell you what happened.
Key takeaways
- ZeroTell scores text sentence by sentence, highlights AI-like sentences, and summarizes them as a document percentage.
- We do not publish an accuracy percentage because performance varies with length, genre, editing, and writer background.
- Short, formulaic, and second-language writing can be flagged even when a person wrote it.
- Read the highlights and context before the number.
- A score is a reason to look closer, never proof on its own. For the wider picture, see how accurate AI detectors are.