How to interpret AI detector results: what your score actually means

You paste your essay into an AI detector. It spits back a number: 72%. Your stomach drops. But what does that number actually mean?

Most people treat AI detector scores like a verdict. High score = guilty. Low score = safe. The reality is messier, and misunderstanding what these scores represent causes real problems: students wrongly accused of cheating, writers flagged for work they wrote themselves, and editors making decisions based on numbers they do not fully grasp.

This guide walks through how to read AI detector results from the tools that matter most: Turnitin, GPTZero, Originality.ai, Copyleaks, and Grammarly. You will learn what the numbers actually measure, why different detectors disagree on the same text, and what to do when your score is higher than you expected.

What an AI detection score actually measures

The first thing to understand: an AI detection score is not a percentage of AI-written words. A score of 60% does not mean 60% of your document came from ChatGPT. It means the detector's model is 60% confident that the overall writing pattern looks AI-generated.

This is a confidence score, not a word count. Think of it like a weather forecast. A 60% chance of rain does not mean 60% of your garden will get wet. It means the model thinks rain is more likely than not, based on the patterns it sees.

Originality.ai explains this directly: "An AI score of 60% Original and 40% AI means the model thinks the content is Original (human-written) and is 60% confident in its prediction." The score reflects the detector's certainty, not the proportion of AI text in your document.

Different detectors frame this differently. Turnitin calls it an "AI writing indicator." GPTZero calls it a "probability." Grammarly describes it as "the percentage of scanned text that is likely AI-generated." But they all converge on the same core idea: these are statistical guesses, not measurements.

Score ranges explained across major tools

While tools use different scoring systems, most produce a 0 to 100 output. Here is how the ranges break down in practice, based on how Turnitin, GPTZero, Originality.ai, and Leap interpret them.

0 to 20: Human range. Text reads as clearly human-written. Most tools treat this range as safe. Turnitin no longer surfaces scores below 20%. If your score is in this range, you are almost certainly fine for any institutional context.

21 to 50: Mixed signals. The detector sees patterns consistent with both human and AI writing. This often means: the text was AI-assisted then edited, a human wrote in a formal or technical register that looks AI-like to the model, or the writer is a non-native English speaker. Most institutional policies treat this as a yellow flag. It warrants a closer look, not an accusation.

51 to 80: Likely AI range. The detector reads the text as more AI-like than human. In academic and publishing contexts, this range will typically trigger a formal review. Unedited ChatGPT or Claude output often lands here. If your score falls in this range and you wrote the text yourself, skip ahead to the false positives section.

81 to 100: Detected range. The detector is highly confident the text is AI-generated. Raw, unedited AI output almost always lands here. This range almost always triggers institutional flags. If you are a human who wrote this and got this score, there is likely something about your writing style that the model finds unusually uniform. See the reliability section below.

It is worth noting that Originality.ai uses a different convention. Instead of a single 0 to 100 AI score, it shows both an Original score and an AI score that together add to 100%. A result of "60% Original, 40% AI" means the model has 60% confidence the text is human-written.

Why different detectors give different scores on the same text

You run the same paragraph through three detectors. One says 15%. Another says 48%. The third says 72%. All three cannot be right. Or can they?

This is not a glitch. It is expected behavior, and understanding why will save you a lot of confusion. Detectors disagree because they are built differently at every level.

First, training data. Each detector is trained on different pairs of human and AI text. Originality.ai trains on the latest LLMs including GPT-5 and Claude 4. Turnitin trains primarily on academic writing. GPTZero was built with student essays in mind. A detector trained mostly on academic prose will react differently to a blog post than one trained on web content.

Second, signal mix. Detectors combine different signals to reach their conclusion. Most use perplexity (how predictable each word is given the previous ones), burstiness (variation in sentence length and structure), and model-specific fingerprints (recurring phrases or structural patterns from known LLMs). Different tools weight these signals differently. A detector that leans heavily on burstiness will react more strongly to uniform sentence lengths. One that prioritizes perplexity will flag highly predictable prose.

Third, threshold calibration. Tools set their own thresholds for what counts as "likely AI." A cautious detector might flag anything above 30%. A more conservative one might wait until 70%. Turnitin publicly states it prioritizes precision over recall, meaning it would rather miss some AI text than falsely accuse a student. Other tools might make the opposite trade-off.

The practical takeaway: cross-checking with two or three detectors gives you a more reliable picture than trusting any single score. If one tool says 85% and two others say 15 to 25%, the outlier is probably wrong. If all three agree above 60%, the signal is strong.

What affects result reliability

Not all texts are equally easy for detectors to classify. Several factors can push a score up or down regardless of who actually wrote the words.

Text length. Short texts give detectors very little signal to work with. A 50-word paragraph is far more likely to produce an unreliable score than a 500-word essay. Most tools perform best on texts above 200 to 300 words. If you are scanning something short, treat the result as directionally interesting at best.

Writing register. Formal academic writing, technical documentation, and legal prose all share features with AI-generated text: consistent tone, dense hedging, predictable structure. A 2023 Stanford HAI study found that seven major detectors flagged 61% of non-native English student essays as AI-written. The writing was human. The style just happened to match what detectors associate with machines.

Non-native English. This deserves special attention. ESL writers often produce text with less variation in vocabulary and sentence structure. Detectors read this uniformity as AI-like. The Stanford study's 61% false positive rate on non-native essays is not a footnote. It is a structural limitation of how these tools work. If you are an ESL writer, cross-check aggressively and keep your draft history.

Heavy editing. A document that has been through multiple rounds of editing, especially with tools like Grammarly, can start to look suspiciously uniform. Each editing pass smooths out the rough edges that human writing naturally has. The result is cleaner prose that, ironically, looks more machine-like to a detector.

Hybrid workflows. Most writing in 2026 falls somewhere between purely human and purely AI. You might outline with AI, write the draft yourself, then use AI for a second pass. Or you might write the first draft and ask AI to tighten the prose. These mixed workflows produce mixed signals. A score in the 40 to 60 range on hybrid text is normal and does not mean the detector is broken. It means the writing is genuinely ambiguous.

False positives: when human writing gets flagged

A false positive happens when a detector marks human-written text as AI-generated. It is the most damaging type of error because it accuses someone of something they did not do. And it happens more often than most people realize.

Originality.ai reports a false positive rate of 0.5% to 1.5% depending on the model. Turnitin acknowledges that false positives are "a possibility." But those numbers are from controlled benchmarks. In the wild, on real student essays and real-world writing, the rate is almost certainly higher, especially for non-native speakers and technical writers.

A widely shared Reddit post from the PromptEngineering community in early 2026 flagged that "AI detectors have a 15% false positive rate. That means they flag real human writing as AI constantly." While the exact number varies by tool and context, the core point stands: false positives are not rare edge cases. They are a known and documented limitation.

Several writing patterns coincidentally produce high scores on original human work: highly technical prose, heavily edited drafts, formal academic register with dense hedging, and short texts under 200 words where the detector has too little signal. If you are a human who writes in any of these styles, a high score is more likely and less meaningful.

If you have been falsely flagged, document your writing process. Keep draft histories. Save versioned files. Use Google Docs or a platform with edit history enabled. The best defense against a false positive is not arguing about detection methodology. It is showing the receipts. For more on this, read our guide on AI detection false positives.

What to do with your score: a practical checklist

You have your score. Now what? Here is a practical workflow that works regardless of which tool you used.

Step one: do not panic. A high score is a signal, not a sentence. Do not delete anything. Do not immediately rewrite the whole document. Take a breath and move to step two.

Step two: cross-check with a second tool. Run the same text through a different detector. If you used Turnitin, try GPTZero. If you used Originality.ai, try Copyleaks. Two independent results tell you more than one. If they disagree widely, the outlier is probably unreliable. If they agree, the signal is worth investigating.

Step three: check the highlighted sections. Most modern detectors highlight which sentences triggered the highest AI signal. Read those sentences in context. Do they sound generic? Overly smooth? Disconnected from specifics? If the flagged passages are genuinely weak, rewrite them with more concrete detail and personal voice.

Step four: consider the context. A 45% score on a 100-word email is noise. A 45% score on a 2,000-word academic paper carries more weight. A 70% score on technical documentation is less surprising than a 70% score on a creative essay. Context shapes how seriously you should take the number.

Step five: if the score stays high, rewrite with intention. Do not just shuffle words hoping the number drops. Vary your sentence length. Add personal examples. Remove hedging phrases that pad without adding meaning. Break up long, uniform paragraphs. One or two short, punchy sentences in a sea of long ones can dramatically shift how the detector reads your rhythm. Our guide on how to edit AI writing to sound human has more specific techniques.

Step six: know your policy. Different institutions have different thresholds. Most universities that have formalized AI detection policies treat anything below 20% as background noise. Some programs have zero-tolerance policies that treat any non-zero score as cause for inquiry. Know what your specific context expects before you react to a number.

One final thought: no detector should be treated as the final word in a high-stakes decision. Use the score as a review cue, not a verdict. The strongest interpretation combines the AI score, your own reading of the flagged passages, and the broader context of who wrote the text and why.

If a detector flags your work, learn why AI detectors give false positives before changing your writing.

The hybrid zone: when 40 to 60 is the honest answer

Most real writing in 2026 is hybrid: drafted by a person, tightened by a tool, edited again by hand. Those mixed workflows produce mixed signals, and a 40 to 60 score is often the correct reading of that process.

That is not a detector failure. It is the detector correctly reporting that the text shows both human and machine patterns. The mistake is treating a middle score as a verdict when it is really a description.

When you see a middle score, read the flagged sections instead of the number. If the highlighted lines are generic, rewrite them with concrete detail. If they are personal and specific, the detector is reacting to rhythm, not authorship.

For a full picture of how tools handle hybrid text, our accuracy comparison across detectors shows the same sample scored by seven tools.

Keep a record of your scores over time. If the same document jumps from 20 to 60 after a tool update, you know the change was the model, not your writing.

And if you are choosing which detector to trust, our 2026 ranking of AI content detectors covers which tools survived independent testing.

The practical rule: treat any single score as a hint, treat a middle score on hybrid work as normal, and treat flagged sections as your editing list.

September 2026 update: how to read a score after a model change

Detector makers ship model updates quietly. The interface looks the same, the score still comes back as a percentage, and nothing on the page tells you the model underneath changed. That matters because a 55% today and a 55% in June are not the same measurement.

What breaks first is your reference points. If you calibrated your instincts in the spring, a summer update can move every score you see by 10 to 20 points without a single word of your writing changing. Treat any single number as a relative signal from one tool on one day, never as a standing verdict about authorship.

Most tools now blend a style model with statistical signals such as perplexity and burstiness. When one half of that blend is retrained, scores shift in a pattern rather than at random. Long, even sentences move first. Short, choppy ones move last. Watching which of your own sentences move tells you more about the update than the headline number does.

Build a calibration set you can re-run:

Keep five texts whose origin you already know, from a clean human draft to a raw AI draft to one heavy edit.

Re-run the set once a month and note the date beside each score, so a jump has a timestamp attached.

Compare the ranking inside one run rather than the raw numbers across runs. The order is more stable than the value.

Check the tool's release notes before you panic. Most score shifts trace back to a version change, not to your writing.

Two checks before you act on a score. First, run a second tool on the same text and see whether both agree on the direction of the change. Second, read the text yourself for the patterns that push scores up, which our roundup of AI detection tools for writers covers in detail.

If a score decides something real, a grade, a contract, a client deliverable, ask what the policy says about human review. In classrooms that usually means a teacher read, and our guide to free detectors for teachers explains where teacher tools are strict and where they are lenient.

If you are appealing a score, send the reasoning rather than the number. Explain which sections were drafted with assistance, what you changed in them, and how the final text was produced. Reviewers respond to a specific account of process far more often than to a screenshot of a percentage from a competing tool.

None of this makes the score useless. It makes it one input among several, which is exactly what the hybrid zone already told you.

How to push back on a score you think is wrong

A flag is a claim, and claims can be checked. The mistake most writers make is rewriting immediately, which destroys the only evidence they had. Ask three questions before you touch a sentence: which tool produced the score, what version or date was it run, and which passages were marked.

Tools disagree with each other on the same text, and they disagree with their own past selves after a model update. So a single score with no version attached is not evidence of anything. A score with a tool name, a date, and the highlighted spans is at least a starting point for a conversation.

Then bring what only you have: draft history. Version history, a saved outline, the notes file, the earlier draft with different opening lines. Writing that grew through revisions looks different from writing pasted in one block, and most reviewers accept that when it is shown to them.

One more piece of context is worth having ready. A 2023 Stanford study sent 91 TOEFL essays written by non-native English speakers through seven detectors and 61.22 percent of the human-written essays came back as AI generated, while the same detectors were close to perfect on essays from US eighth graders. That gap is about style, not authorship, and it is the strongest argument a flagged writer has.

The study is at Stanford HAI on detector bias, and the pattern it describes shows up again in the false positive data we track.

Ask for the specifics in writing: which tool, which version, which spans, and what the policy says about detector-only evidence. Many course and workplace policies now require more than a score, precisely because the tools are not reliable enough to stand alone.

Ask for the passages rather than the percentage. A number gives you a feeling, a highlighted sentence gives you an edit. Then read the flagged sentences together and look for the trait they share, which is usually length, list shape, or that even, balanced rhythm detectors score highest.

If the flag arrives with no tool name, no date, and no marked passages, you can say the check isn't reviewable yet. That's not a technicality, it's the practical version of the same request: name the tool, show the spans, and let the writing answer the rest.

If you are the one running the check rather than the one being flagged, start with the tools built for writers instead of a general purpose detector, and read the score as a signal to look closer, never a verdict.

Writing in a second language makes a false flag more likely, and why detectors flag non-native English writing walks through the specific rhythm patterns that trigger it.

What percentage of AI is acceptable?

Under 20 percent is treated as human at every major detector we track. Turnitin treats anything below 20 percent as too weak to stand on its own. Between 20 and 50 percent, read the flagged sections instead of the headline number. So a 12 or 16 percent score sits inside the safe band, and quoting the threshold back is a stronger answer than defending the number itself.

Those cutoffs are policy choices, not measurements. An institution that publishes a rule usually picks a number like 20 percent as the line where a flag becomes a conversation, and another picks 50. That's why the same 16 percent result can be a non-event at one school and a meeting at another. Check the policy first, then look at the score again.

The bands come from the tools. Originality.ai reports an Original score and an AI score side by side, so its 40 percent reading describes the model's confidence rather than the share of your document a machine wrote. If you have been flagged at a low score, the argument is in the research: the false positive rates each detector produces on human writing show how often real people get caught by mistake.

One more number worth keeping ready. A 2023 study at Stanford HAI ran 91 human-written TOEFL essays through seven detectors, and 61.22 percent came back as AI generated, while the same tools were close to perfect on essays from US eighth graders. A threshold that produces that kind of error on real human writing is a policy line, not a measurement of authorship.

When a score lands in the middle, comparing tools on the same text is the quickest way to see how much the number depends on the model: the same sample scored by seven detectors rarely lands on one verdict.

The three questions behind most of the traffic to this page, answered in one line each. Is 16 percent AI detection bad? No, it sits in the safe band at every tool we track. Is 12 percent bad? No, and a lot of fully human writing lands there on a formal register. What percentage is acceptable? Under 20 percent at most institutions, 20 to 50 percent means read the flagged lines, and above 50 percent means the tool is confident enough that you should be ready to explain your drafts.

Here is how that sounds in practice. An appeal that says the score was 16 percent and the institution's threshold starts at 20 gives a reviewer a policy to check. An appeal that says I promise I wrote it gives them nothing to check. Same score, same writer, different outcome.

Publishers and clients rarely publish a threshold at all. When someone asks what score you can guarantee, the honest answer is none, because the tool, the date, and the length all move the number. What you can offer is the drafting history and a fresh run from the tool they use, on the same day they ask.

Which detector should you trust for your document type?

Match the tool to the writing. Academic submissions run through Turnitin, which trains on academic prose and withholds any score between 1 and 19 percent. Web and marketing content sits closer to what Originality.ai and Copyleaks are built on. Short text under 200 words is unreliable everywhere, so treat any score there as noise.

For a second opinion, cross-check with a tool trained on different data. A detector built on student essays reads a blog post differently than one trained on web content, which is exactly why two results beat one.

If you want independent numbers rather than marketing pages, our accuracy comparison across seven detectors retested the field, and our ranking of AI detection tools covers which ones survived. For the reader's side of the same problem, see how to detect AI generated text.

One rule covers every document type. The score tells you where to look, never what is true. Read the flagged sentences, compare them against your own drafts, and treat the number as a pointer rather than a verdict.

Length matters more than the tool. A 60-word paragraph gives any model too little signal to work with, and most tools only behave above 200 to 300 words. If your document is short, no detector choice will make the score meaningful.

The same logic applies to how you read a flagged section. A detector marks sentences that look statistically smooth, which is not the same as sentences a machine wrote. Heavy editing with a grammar tool produces exactly that smoothness, so a writer who has polished a draft three times can score higher than one who typed the first version straight through.

Keep the flagged sentences and your drafts together. If a reviewer asks, the pair answers the question better than any explanation of how detectors work, because it shows the writing changing under your own edits rather than appearing in one block.

Related reading: Why AI humanizer results change every time you run them · which AI detector do college professors trust

How reliable is a detector score in 2026?

A band is a reading convention, not a property of your text. The clearest evidence arrived in 2026, when a University of Florida team cloned roughly 6,000 conference papers written before ChatGPT existed and ran both versions through the five most popular commercial detectors. False positive rates ranged from 0.05% to 68.6%. False negative rates ranged from 0.3% to 99.6%. Two detectors held up until the researchers asked the model to write with more complex vocabulary, and then their reliability fell apart. The work went to the IEEE Symposium on Security and Privacy, and it makes one thing plain: a number tells you what that tool does with your text today, not what your text is.

Put that range next to the band table above and the practical rule changes. A score of 12% from a detector with a 68% false positive rate carries less information than a score of 40% from one with a rate under 1%. You can't compare across tools, and you can't compare across versions of the same tool. Turnitin's own published figures make the same trade from the other side: a document-level false positive rate under 1% for work it scores at 20% or higher, a sentence-level rate near 4%, and an accepted miss rate of roughly 15% of AI text. Missing 15% isn't an oversight. It's the price of keeping false accusations low, and it should shape how you read a low score from them.

So use the bands as triage, not proof. Under 20 means read the piece for rhythm and specifics. Twenty to 50 means read the flagged lines and check whether the pattern is real or an artefact of uniform paragraph length. Above 50 means you'll be asked to explain your drafts, so keep them. What the bands are not good for is settling an argument with an institution. If someone holds up a score as a verdict, the vendor's own published error rate is the better page to put beside it. Our accuracy comparison across detectors shows how far apart two tools can land on one sample.

One more reason the bands move: the same detector is a different product from one release to the next. A 2026 benchmark on Turkish academic writing tested six supervised detectors and one deployed commercial system under locked thresholds. The best held-out model reached 0.9154 AUROC with a human false positive rate of 4.0% and AI recall of 69.5%. The deployed system reached 0.86 recall and still falsely flagged 35.0% of human texts. Those aren't bad researchers. That's what a detector looks like when it's tuned hard toward catching machines, and it's the same tuning that produces the scary score on your document. The measurement study is worth reading if you're setting a policy rather than arguing about one document.

The two-tool agreement test still holds, and it's cheap. Run a second detector on the same text. If both tools move in the same direction after an edit, you've learned something about the text. If they disagree by 40 points, you've learned something about the tools. Either way, save the file you started from. A draft history beats any single score when you have to show how the work was made, and it's the only evidence in the conversation that costs you nothing but a version history you probably already have. Open the file list, check the timestamps, and let the version trail do the arguing. That's usually enough to end it, and it beats a screenshot of a score. Keep the first file, the flagged file, and the final one, and name them so a stranger can follow the order.

Frequently asked questions