What does an AI detection score mean?

Updated October 31, 2026 · 4 min read

Short answer

It depends on the tool. In Turnitin, the percentage is the share of the text's qualifying sentences that its model thinks were written by AI. In GPTZero, the percentage is a probability about the whole document, not the share of AI text. Originality.ai's score was a confidence level until its 5.0 models (October 2026) and is now described as the share of AI involvement.

So "40% AI" can mean "about 40% of the sentences look AI-written" or "the tool leans toward human but isn't sure." Check which tool and which version produced a number before drawing conclusions, read the highlighted passages, and treat any score as one piece of evidence.

What does each detector's number measure?

Tool	What the number means	Worth knowing
Turnitin	The share of qualifying prose sentences its model judges likely AI-generated, including AI text changed by paraphrasers or "bypasser" tools.	Not a share of the whole file: lists, tables, code and references are excluded. No score for 1 to 19% (asterisk). Needs 300+ words.
GPTZero	A probability that the document belongs to each class: human, mixed or AI.	GPTZero says "6% AI" does not mean 6% of the document contains AI. Its 4o model has been the default since September 20, 2026.
Originality.ai (before 5.0)	A confidence level: "60% Original and 40% AI" meant the model was 60% confident the text was human.	Its own 2024 explanation: the score "DOES NOT mean that 60% of it was Original and 40% was AI-generated."
Originality.ai (5.0, from Oct 2026)	Described as the percentage of the text that is Likely AI, Likely Original or Mixed.	The same page still calls it a confidence elsewhere, so note which version produced a score.
Copyleaks	An estimate of how much of the document's content generative AI may have created.	Separate from its plagiarism (similarity) score. Recommends scanning at least 350 words.
Pangram	The fractions of the text that are AI-generated, AI-assisted and human, with a label for each segment.	Light copyediting, spelling and grammar fixes count as human. AI-assisted text counts at half weight in its AI fraction.
AI Checker Plus	The shares of the text marked AI-generated, AI-assisted and human-written, with each passage highlighted and its own AI likelihood.	The free check covers the first 300 words; Plus checks up to 5,000 words at once.

The practical difference: a share-of-text score (Turnitin, Copyleaks, Pangram, AI Checker Plus) tells you how much of the writing looks AI-generated, so the highlighted passages matter most. A probability score (GPTZero, and Originality.ai before 5.0) tells you how sure the tool is about the document as a whole.

What does a score under 20% mean?

Usually very little. Turnitin stopped showing numbers and highlights for results between 1% and 19% in July 2024, and shows an asterisk (*%) instead, "to avoid potential incidence of false positives." A low score in any tool can come from a few predictable sentences, a quote, or a phrase many people write the same way.

What does a 40% or 50% score mean?

It's a mixed signal, so the details matter:

In Turnitin and other share-of-text tools, roughly that share of the sentences was flagged. Read the highlighted passages: are they the introduction and conclusion, a section with a different voice, or scattered sentences?
Turnitin's figure can undercount. To keep false positives low, Turnitin accepts missing some AI text: it says that if it identifies 50% of a document as likely AI, it "could contain as much as 65% AI writing."
In a probability tool, 40% to 50% means the model can't tell. GPTZero adds confidence labels to its results and states higher error rates for the less confident ones.
Mixed results often mean editing. Text written by a person and then polished with an AI tool, or the reverse, tends to land in the middle. Pangram and AI Checker Plus label it AI-assisted rather than AI-generated.

What does a 100% AI score mean?

The tool is highly confident, or every qualifying sentence was flagged. It's the strongest signal a detector gives, but it still isn't proof: the University of Waterloo reported that "in more than one instance the product flagged human written text as 100% generated by AI." If the text really is a pasted chatbot answer, everything in it counts as AI, including quotes the chatbot reproduced.

Why do detectors give different scores for the same text?

They measure different things, as the table above shows.
They're trained differently and set different thresholds for what counts as AI.
They define "human" differently. Pangram counts light copyediting and literal translation as human; other tools may score the same edits as partly AI.
They change often. In 2026 alone, GPTZero switched to its 4o model (September), Originality.ai launched 5.0 (October), Pangram released version 4 (July) and Turnitin updated its English model (October). A score from last semester may not match today's.

How should you read a score?

Note the tool, the date and, if shown, the model version.
Check the length: Turnitin won't score under 300 words, Copyleaks recommends 350+, and all tools are less reliable on short texts.
Read the highlighted passages, not just the headline number.
Compare with other evidence: drafts, version history, the writer's earlier work, and a conversation about the content.
Never act on the number alone. Turnitin says its score "should not be used as the sole basis for adverse actions against a student."

Questions people also ask

Is a 30% AI score bad?

Not by itself. In Turnitin, 30% means about 30% of the qualifying sentences were judged likely AI-generated, so look at which passages are highlighted and whether they match drafts or notes. In a probability-style tool like GPTZero, a 30% AI probability means the model leans toward human. Either way, the score is a reason to look closer, not a conclusion.

Why does Turnitin show an asterisk instead of a number?

Since July 2024, Turnitin shows *% with no number and no highlights when it detects 1 to 19% AI writing, because false positives are more likely in that range. The asterisk means a low, unreliable signal.

Can a fully human essay get a 100% AI score?

Yes, it happens. The University of Waterloo said that in more than one instance Turnitin flagged human-written text as 100% generated by AI, one of the reasons it stopped using the feature in September 2025.

Do quotes, references and lists count toward the score?

In Turnitin, no: it only scores qualifying prose, so lists, bullet points, tables, code and references are left out. In AI Checker Plus, an answer you copy from ChatGPT or another chatbot counts as AI-written in full, including any quotes inside it, because the chatbot produced those words.

Sources

By the AI Checker Plus team. We research and draft these guides with help from AI tools, then check every claim against the sources listed above.