Products SEO & GEO Agent Paid Ads Agent Website Agent Conversion Agent
How it works Pricing FAQ Blog Book a call

AI score meaning: how to read AI detection scores

Updated · 5 min read

Short answer It depends on the tool. Turnitin's percentage is the share of long-form prose it thinks AI wrote. GPTZero and Originality.ai mostly report how sure the model is of its label, so 90% AI means a confident guess, not 90% of the words. Every maker admits false positives, so no score proves who wrote something.

Want help applying this to your business? Talk it through on a free 30-minute call.

Book a call

What does an AI detection score mean?

It's a guess, and each tool words its guess differently. Before you act on "72% AI," find out which kind of number you're looking at:

ToolWhat the number measuresSo "80% AI" means...
TurnitinShare of qualifying text (long-form prose) its model thinks AI likely wroteAbout 80% of the essay-style sentences were flagged; bullets, tables and short text aren't counted
GPTZeroChance its label (human, mixed or AI) is rightThe tool is fairly sure the document is AI, not that 80% of it is
Originality.aiModel confidence that AI was involved, per its help centerHigh confidence that AI touched the text somewhere

Turnitin says its percentage "indicates the amount of qualifying text within the submission" that its model judged "was likely generated by AI," and adds that this "is not necessarily the percentage of the entire submission" (Turnitin Guides).

GPTZero describes its number as a hit rate:

"I.e. 90% means that 90% of the time on similar documents our detector is correct in the prediction it makes." (GPTZero FAQ)

Originality.ai's help center says "An AI score is a probabilistic signal, not the percentage of words written by AI" (Originality.ai Help). Its own blog, updated October 2026, describes the same score as "the percentage of AI involvement in a piece of content" (Originality.ai blog). When even one maker explains its number two ways, read any score as a rough signal.

Is there a safe or "good" AI score?

No number is safe or damning across tools. The scales differ, so 30% in one tool and 30% in another aren't the same claim.

Turnitin itself treats low scores as noise. It found "a higher incidence of false positives" between 1% and 20%, so since July 8, 2024 it shows only an asterisk (*%) there, with "no score or highlights" (Turnitin Guides). A result of 0% or 20–100% shows as a number.

What a score can't tell you:

  • How much of the text a person wrote. Except in Turnitin, the number is confidence, not a word count.
  • Whether the facts are right. A fully human page can be wrong; an AI draft can be accurate.
  • Whether Google will rank it. Google says its "focus on the quality of content, rather than how content is produced" has guided its results for years (Google Search Central Blog). The SEO side is in AI detectors and SEO.

How often do AI detectors flag human writing?

Often enough that every maker warns against relying on one scan. Turnitin says its model "may not always be accurate" and "should not be used as the sole basis for adverse actions" (Turnitin Guides). GPTZero says "There always exist edge cases" where human text is called AI (GPTZero FAQ). Originality.ai says no detector should be "the sole basis for a disciplinary, employment, publishing, or academic decision" (Originality.ai Help).

The public record backs them up:

  • OpenAI pulled its own detector. It caught 26% of AI text and labeled human text as AI 9% of the time. On July 20, 2023, OpenAI withdrew it "due to its low rate of accuracy" (OpenAI).
  • Non-native English gets flagged. Stanford researchers ran seven detectors on 91 essays by non-native English writers; on average 61.22% were called AI-generated (Liang et al., 2023).
  • Accuracy claims don't always hold. The FTC says Workado advertised its detector as "98 percent" accurate, while independent testing found 53% on general content (FTC, April 28, 2025).

Some writing is flagged more than others. Originality.ai lists text that is "short, highly structured, heavily edited, translated, or written in a predictable style," and names product descriptions, instructions and "frequently repeated marketing phrases" as formats that can resemble AI (Originality.ai Help). That describes a lot of small business copy: service lists, FAQs, policy pages.

My content got a high AI score. What should I do?

Check the evidence before you change a word. The makers' own advice adds up to five steps:

  1. Scan the whole piece, not a paragraph. Turnitin skips bullets, tables and short text; Originality.ai says short passages "are more vulnerable to uncertain results" (Originality.ai Help).
  2. Note the tool and model. Two tools, or two models of one tool, give different numbers on the same text. Compare only same text, same model.
  3. List the helpers used. Grammar checkers, translators and rewriters can make human text look AI-like, Originality.ai says, naming Grammarly, DeepL and QuillBot.
  4. Pull the history. Google Docs or Word version history, earlier drafts, notes, and older work by the same writer are better proof than any score.
  5. Ask the writer. A short talk about how they wrote it, and whether they can explain the facts, settles more than rescanning.

Don't keep rewriting good text until the number drops. Originality.ai warns that rewording "solely to lower the AI score" changes the evidence and "can make the writing less natural" (Originality.ai Help).

Should I use an AI score to check a freelancer?

Use it to start a question, never to end one. What you're paying for is accurate, specific content. Check that directly: are the prices, rules and steps right, and does the page say things only your business would know? If your contract bans AI drafting, say so in writing and agree how you'll check, for example by asking for the version history. For what a person should add to an AI draft, see AI content plus human input.

Example

A roofer reviewing a freelancer's blog post

The post scores 85% AI in one tool. The owner opens the Google Doc history: it was typed over three days with edits. Two shingle prices are out of date. He asks for the price fixes and drops the AI question.

A Spanish-speaking salon owner's own website text

She writes in Spanish, translates, then edits. A client runs the text through a free detector and says it's "AI." Translated and plain text is what detectors flag most. Her drafts and version history answer the claim.

An accountant getting "*%" in Turnitin

Her continuing-education paper shows "*%." That means Turnitin detected something below 20% and chose not to show a number, because false positives are more common there. It isn't a finding against her.

FAQ

Does 50% AI mean half my text was written by AI?

Only in Turnitin, and only for its qualifying prose. In GPTZero and Originality.ai, 50% means the model isn't confident either way.

Why do two detectors give different scores for the same text?

They use different models, training data and scales. One may measure share of text, another confidence. Different numbers are expected and neither is "the real one."

What does the asterisk in Turnitin's AI score mean?

AI was detected below 20%. Turnitin hides the number there because its testing found more false positives in that range.

Can a high AI score prove someone used AI?

No. Turnitin, GPTZero and Originality.ai each say their results shouldn't be the only basis for a decision.

Different for you?

Keep learning

NextRewrite ChatGPT output so it's useful, not just undetectedHow to rewrite a ChatGPT draft for your business site: fix facts, add what only you know, cut filler. Why synonym swaps and detector chasing backfire.

Ask us one question

One line is enough. We reply by email.

Sources

  1. Turnitin Guides: AI writing detection in the classic report view (updated September 8, 2025)
  2. GPTZero: Frequently asked questions (read 2026-11-16)
  3. Originality.ai Help: Human-Written Content Flagged as AI: What to Do Next (read 2026-11-16)
  4. Originality.ai: Is That Text AI or Not: What Does AI Score Mean? (October 25, 2026)
  5. OpenAI: New AI classifier for indicating AI-written text (January 31, 2023; updated July 20, 2023)
  6. Liang et al., GPT detectors are biased against non-native English writers (Patterns, 2023)
  7. FTC: FTC Order Requires Workado to Back Up Artificial Intelligence Detection Claims (April 28, 2025)
  8. Google Search Central Blog: Google Search's guidance about AI-generated content (February 8, 2023)