How Accurate Are AI Detectors? Our 2026 Test Results

Uzair Khan

18 min read

AI detectors should not be used in isolation to make a judgment, and none will ever be completely accurate. The more "raw" the AI text is (i.e., unedited), the more accurate the score.

Accuracy will reduce the more human input is added, and can still flag human-written text. So consider the score a prompt to investigate further.

If you've been flagged as a student, please don't panic. If you've been flagged as an educator, please read on for information about what a detector score means and doesn't mean.

We conducted our own set of tests using 6 AI detectors and detail our false positive/negative rates below.

The tests below were performed on 14/11/2026. Our sample is limited to 1 per category and 6 detectors. Please consider this indicative of detector performance only. See the academic sources below for more information on AI detector accuracy.

Please note that we (Phrasly) are the authors of this page, as well as an AI Detector. Our AI detector forms the last column of our tests, and we have reported our false positives/negatives accordingly.

We will cover how AI Detectors work and calculate a score, show our results, debunk 4 common misconceptions, discuss Turnitin, and close with guidance if you have been flagged.


Check a passage with Phrasly's AI detector and read its score alongside the guidance below:


How AI Detectors Score Text, in Plain Terms

Three factors behind AI detector scores: perplexity, burstiness, confidence

A score is simply a probabilistic judgement based on language features, and should be treated as a heuristic to guide further assessment rather than conclusive evidence. It's important to note that no detector can 'detect' ChatGPT per se.

Rather, they predict the likelihood of a given piece of text having been written by an LLM.

In practice, scores can be understood via three main factors:

  •  Perplexity: how predictable each next word is to a language model. Lower perplexity means the wording is easier to guess.

  • Burstiness: how much sentence length and structure vary across a passage. Human writing usually swings between short and long sentences.

  • Classifier confidence score: a probability that the text resembles AI output. It is not a verdict about who wrote it.

Compare two short passages.

A low-variance sample reads: "Time management is important for students. It helps them complete their work on time. It also reduces stress during exams."

A high-variance sample reads: "I missed the 9 a.m. bus twice in October. After that, I started packing my bag the night before, which sounds small but fixed half my mornings."

The second swings in length and carries specific detail. It looks less predictable to a detector.

If you write in a formal style, plainly, or are not a native English speaker, your writing may be flagged as low perplexity/burstiness. This is simply a limitation of AI detectors and does not say anything about your writing.

Detectors weigh several signals at once:

  • Word predictability (perplexity)

  • Variation in sentence length and structure (burstiness)

  • Repeated stock phrases and transitions

  • Uniform paragraph shape

  • Text length, since short passages give the model less to judge

Since AI detectors are calibrated against existing LLMs, their output will vary with the model they have been trained on. This is known as drift, and is expected as new LLMs are released.

For the mechanics, see how AI detectors work.

Our 2026 AI Detection Accuracy Test

2026 AI detector test stats showing false positives on human writing

We tested one example from each of the 6 classes with 6 AI detectors, and broke out the results separately. Reporting only an "overall accuracy" would obscure false positives (human text detected as AI) and false negatives (AI text not detected as AI).

Detectors compared are GPTZero, Originality.ai, Pangram, Copyleaks, ZeroGPT, and Phrasly (disclosed, hence at the end of the table).

We have not included results for Turnitin as it is only accessible via university accounts, but we will mention some publicly available data about its performance below.

Cells show the AI likelihood score assigned to each example by the given detector (n=1 example per class).

Sample type

GPTZero

Originality.ai

Pangram

Copyleaks

ZeroGPT

Phrasly

Human-written (want LOW)

100%

7%

0%

0%

0%

0%

Human-written ESL (want LOW)

100%

92%

100%

100%

100%

0%

Raw AI (want HIGH)

100%

98%

100%

100%

100%

100%

Human-edited AI (want HIGH)

100%

90%

100%

100%

100%

100%

Paraphrased AI (want HIGH)

100%

96%

100%

100%

100%

100%

Mixed human and AI (want HIGH)

51%

42%

41%

100%

33%

100%

'AI detection accuracy 2026 spot-check results table'.

Scores from each AI detector, 14/11/2026. Flag if AI score > 50%. We would expect "human" entries to be scored low (high scores are therefore false positives) and "AI" entries to be scored high (low scores = false negatives). Only one example per class, so scores should not be interpreted as a rate.

Interpreting this table: For the 2 human-generated rows, a high result indicates a false positive. For the 4 AI-generated rows, a score <50% indicates a false negative. You can follow each column to trace a given detector, or across a row to compare detectors. The following 2 rows are of particular interest and shown here with images.

False Positives Landed on Human Writing, Especially ESL

Five of six tools detected the ESL human sample with a detection score of 92-100%. All tools passed the Native-English human sample with the exception of GPTZero, which gave a 100% AI detection. Phrasly passed both human samples.

Turnitin alternative GPTZero false positive on human text.

GPTZero scored the Native-English human-written sample 100% AI, a false positive.

AI detector false positive on ESL human writing.

Originality.ai flagged the ESL human sample at 92% AI. Five of six tools flagged this human text.

Phrasly AI detector clears ESL human text.

Phrasly was the only detector to clear the same ESL human sample as human-written.

More human-sample scans across all tools:

Copyleaks: Human-written ESL: 100% AI

Copyleaks: Human-written ESL: 100% AI  (false positive)

GPTZero: Human-written ESL: 100% AI

GPTZero: Human-written ESL: 100% AI  (false positive)

Pangram: Human-written ESL: 100% AI

Pangram: Human-written ESL: 100% AI  (false positive)

ZeroGPT: Human-written ESL: 100% AI

ZeroGPT: Human-written ESL: 100% AI  (false positive)

Copyleaks: Human-written: 0% AI

Copyleaks: Human-written: 0% AI

Originality.ai: Human-written: 7% AI

Originality.ai: Human-written: 7% AI

Pangram: Human-written: 0% AI

Pangram: Human-written: 0% AI

Phrasly: Human-written: 0% AI

Phrasly: Human-written: 0% AI

ZeroGPT: Human-written: 0% AI

ZeroGPT: Human-written: 0% AI

False Negatives Landed on the Mixed Human-and-AI Sample

All detectors identified the pure-AI, human-edited and human-rewritten content with scores of 90%+. The 50% AI / 50% human piece performed better, scoring <50% on Originality.ai, Pangram, and ZeroGPT, and 51% on GPTZero.

However, it was detected at 100% by Copyleaks and Phrasly. Phrasly highlighted in red what it identified as AI and did not highlight what it identified as human text.

ZeroGPT misses mixed AI text.

ZeroGPT read the mixed human-and-AI sample as 33% AI, below the flag line, a false negative.

Originality.ai misses mixed AI text.

Originality.ai scored the same mixed sample 42% AI, also under the threshold.

Pangram detects raw AI text.

On raw, unedited AI, the tools were reliable: Pangram scored the raw AI sample 100% AI.

More mixed human-and-AI scans:

Copyleaks: Mixed human + AI: 100% AI

Copyleaks: Mixed human + AI: 100% AI

GPTZero: Mixed human + AI: 51% AI

GPTZero: Mixed human + AI: 51% AI

Pangram: Mixed human + AI: 41% AI  (miss / false negative)

Pangram: Mixed human + AI: 41% AI  (miss / false negative)

Phrasly: Mixed human + AI: 100% AI

Phrasly: Mixed human + AI: 100% AI

AI correctly caught: raw, human-edited, and paraphrased, across all tools:

Copyleaks: Raw AI: 100% AI

Copyleaks: Raw AI: 100% AI

GPTZero: Raw AI: 100% AI

GPTZero: Raw AI: 100% AI

Originality.ai: Raw AI: 98% AI

Originality.ai: Raw AI: 98% AI

Phrasly: Raw AI: 100% AI

Phrasly: Raw AI: 100% AI

ZeroGPT: Raw AI: 100% AI

ZeroGPT: Raw AI: 100% AI

Copyleaks: Human-edited AI: 100% AI  (false positive)

Copyleaks: Human-edited AI: 100% AI  (false positive)

GPTZero: Human-edited AI: 100% AI  (false positive)

GPTZero: Human-edited AI: 100% AI  (false positive)

Originality.ai: Human-edited AI: 90% AI  (false positive)

Originality.ai: Human-edited AI: 90% AI  (false positive)

Pangram: Human-edited AI: 100% AI  (false positive)

Pangram: Human-edited AI: 100% AI  (false positive)

Phrasly: Human-edited AI: 100% AI  (false positive)

Phrasly: Human-edited AI: 100% AI  (false positive)

ZeroGPT: Human-edited AI: 100% AI  (false positive)

ZeroGPT: Human-edited AI: 100% AI  (false positive)

Copyleaks: Paraphrased AI: 100% AI

Copyleaks: Paraphrased AI: 100% AI

GPTZero: Paraphrased AI: 100% AI

GPTZero: Paraphrased AI: 100% AI

Originality.ai: Paraphrased AI: 96% AI

Originality.ai: Paraphrased AI: 96% AI

Pangram: Paraphrased AI: 100% AI

Pangram: Paraphrased AI: 100% AI

Phrasly: Paraphrased AI: 100% AI

Phrasly: Paraphrased AI: 100% AI

ZeroGPT: Paraphrased AI: 100% AI

ZeroGPT: Paraphrased AI: 100% AI

Methodology, for anyone who wants to repeat it:

  • Samples: One per category (n=1), each about 180 to 240 words: a human-written sample, a human-written ESL sample, raw AI, human-edited AI, paraphrased AI, and a mixed human-and-AI document.

  • How the AI samples were made: The raw sample was generated with a current large language model; the paraphrased sample was that text run through a paraphrasing tool; the human-edited sample was the raw text revised by hand; the mixed sample paired a human paragraph with an AI paragraph.

  • Threshold: A score of 50% AI or higher counted as "flagged." The same threshold was applied to every tool and sample.

  • Tools and date: GPTZero, Originality.ai, Pangram, Copyleaks, ZeroGPT, and Phrasly, all tested on 14 November 2026 on their current public versions.

  • Limits: With one sample per category, this is illustrative, not statistical. It shows behavior on specific texts, not a reliable accuracy rate; the research sections below carry the weight for general claims.

This matches existing findings. False positives were seen predominantly on human and to a lesser degree ESL texts. False negatives were predominantly on edited/mixed and not unedited AI.

Phrasly was the only tool to produce no errors across our samples, but the caveat is that we have only n=1 per condition.

Of these detectors, which produced the lowest number of false positives? Phrasly produced none, whereas GPTZero produced false positives on both human texts. This is based on our limited test only (n=2), and we point the reader to other independent evidence on the rest of this page.

Some detectors have a limited free tier. Check individual websites for details. Phrasly’s AI detector is free to use, and please see Phrasly pricing page (phrasly.ai/pricing).

Myth 1: 99% Accurate means 99 of 100 Essays are Judged Correctly

What's the actual accuracy of AI detectors in practice? Most claim somewhere between 98% and 99.9%, typically based on testing against their own corpus of AI-generated text. Of course, this is relatively straightforward to identify as AI.

In Scribbr's 2023 detector test, Copyleaks scored 66% against the 99% accuracy Copyleaks advertises (Scribbr, 2023 test).

Scribbr's data is from 2023, so the tools may have changed since.

For the current numbers, see these AI detection statistics for 2026.

A number of vendors will release studies of their own, so it is worth reviewing these. Originality.ai has a useful collection of studies. Winston AI reports that its Model 4.0 has a 0.03% false positive rate (again, this is self-reported by the vendor).

It is important to note the details of how each study was conducted, i.e., by whom, on what text, and what length.

Five questions sort a strong accuracy claim from a weak one:

  • Who ran the test, the vendor or an independent group?

  • Were human samples included, and were any written by non-native English writers?

  • Were edited or paraphrased AI samples included?

  • Are false positives and false negatives reported separately?

  • When was the test run, and against which AI models?

Base rates matter more than a headline percentage. A 1% false positive rate applied to 10,000 honest students still flags about 100 of them. That is simple arithmetic, not a study finding.

Institutions have acted on these limits.

OpenAI shut down its own AI Text Classifier on July 20, 2023, citing its low rate of accuracy (Search Engine Land).

Several universities have since chosen to disable AI detection, including Vanderbilt (Vanderbilt).

Myth 2: Human Writing Never Gets Flagged

Can a false positive occur with an AI detector? It is possible - where a piece of human-written text is incorrectly identified as AI.

For example, if a student writes a concise lab report conclusion using standard language, it may be detected as AI-generated, as it is low-variability (see example above).

The clearest evidence is about non-native English writers.

Liang et al. (2023), published in Patterns, found that detectors flagged more than half of TOEFL essays written by non-native English speakers as AI, while rarely misclassifying essays by native speakers (Liang et al., 2023; Stanford HAI summary).

It's commonly cited that the study showed a 97% false-positive rate. In fact, 97.8% of TOEFL test papers were detected as AI by at least one detector, with the false-positive rate per detector averaging 61.22%.

So the rate of bias is indeed high, but the precise number should be quoted with this caveat.

This can be attributed to what was described earlier: lower perplexity & burstiness due to less diverse word usage and more uniform syntax.

Other writing also tends to trigger false positives:

  • Plain or highly formal styles

  • Technical or scientific writing with fixed terminology

  • Templated assignments, such as lab reports or structured responses

  • Short texts with too few sentences to judge reliably

Turnitin states a document-level false positive rate under 1% for documents with more than 20% AI writing (Turnitin).

That is a vendor figure measured at specific score thresholds, and the base-rate math from the previous section still applies.

In our spot check, we got similar results. Our ESL human sample was detected as 92-100% AI-generated by all but one detector (Phrasly). While this is, of course, only 1 sample, it's consistent with the findings above.

It can understandably be upsetting to have your genuine writing marked as AI. We won't claim your appeal is guaranteed to be successful - just that having your documentation may assist.

Myth 3: Detectors Catch any AI-Written Text

No! AI detectors have a fairly high false negative rate. They are more likely to fail if the AI text has been altered - e.g., rewritten by a human or using a spinner, blended with human content, if the piece of text is very short, or generated with a model newer than what the detector uses.

An independent benchmark shows how far accuracy can fall.

RAID (Robust AI Detection) evaluates how detectors hold up against edited and adversarial text, and reports large drops under paraphrasing and simple edits (RAID paper, arXiv 2405.07940; leaderboard).

Record the date you check the leaderboard, since rankings change.

Several factors push detection accuracy down:

  • Human editing of AI drafts

  • Paraphrasing tools

  • Hybrid human and AI writing

  • Text length

  • Model drift as new AI models are released

  • Language of the text, since most detectors are trained mainly on English

For more on how rewording changes a score, see whether detectors spot paraphrased AI text.

Our spot test was consistent with these results. Every detector identified the unedited, edited, and rephrased AI outputs at >90%, with only the hybrid (50% human/50% AI) output returning scores <50% from Originality.ai (42%), Pangram (41%) and ZeroGPT (33%). Therefore, it is the presence of both human and AI text that evaded detection, rather than any degree of human editing.

This means that while AI detectors are helpful for flagging potential AI use, they should not be relied on exclusively to establish whether or not a student used AI. Other evidence, e.g., prior drafts/Word versions, and/or speaking to the student, should also be considered.

Myth 4: All AI detectors Agree on the Same Text

Students and teachers often believe that there is some 'true' AI% that detectors should all converge on. However, all AI detectors are independent classifiers and will each return a different (but estimated) value for any given input.

These estimates will vary with detector, length of the input, and subject matter.

Five causes explain most of the disagreement:

  • Training data: Each vendor trains on its own mix of human and AI samples, so a tool trained mostly on older model outputs may miss newer ones.

  • Model family: Some tools rely on perplexity and burstiness signals, while others use fine-tuned transformer classifiers, and they notice different patterns.

  • Flag thresholds: One tool may label 50% confidence as "likely AI," while another waits for a higher cutoff.

  • Document-level vs sentence-level scoring: A single overall score and a set of highlighted sentences are not measuring the same thing.

  • Short or mixed text: Accuracy falls on short passages and on human text mixed with AI, and some tools lose accuracy under about 350 words.

Published data shows the gap. Scribbr's 2023 test scored Copyleaks at 66% against its advertised 99%, and found AI text on specialist topics was detected 57% of the time, compared with 67% for general topics (Scribbr, 2023 test).

GPTZero and Copyleaks can land on very different scores for the same paragraph.

This is what we found in our spot check. Against the purely human ESL data, Phrasly gave 0%, Originality.ai 92%, and Copyleaks, Pangram, and ZeroGPT all 100% on the same passage.

The mixed human+AI data returned between 33% (ZeroGPT) and 100% (Copyleaks & Phrasly). But Phrasly clearly highlighted in red what it considered AI. Identical passages, wildly varying outputs.

If there is such variation, it suggests that the text sits somewhere in between what the detectors consider human vs. AI. Submitting the same passage to multiple detectors to obtain a "lower" score is meaningless.

For the fuller explanation, see why AI detectors give different results.

Is Turnitin's AI Detection Accurate?

Should we trust the results from Turnitin's AI detector? Turnitin claims to have minimal false positives, but research suggests their ability to flag AI writing is lower than advertised. The company also advises that the result should not be used as sole grounds to accuse a student. The score represents the proportion of text deemed likely to have been generated by AI.

Turnitin states a document-level false positive rate under 1% for documents with more than 20% AI writing, requires at least 300 words of prose, and says the AI indicator "should not be used as the sole basis for action" (Turnitin AI writing detection).

The <1% is a Turnitin statistic (at their defined confidence interval) and should therefore be quoted as a Turnitin statistic, not via their competitor.

Independent work is more skeptical.

Weber-Wulff et al. (2023), in the International Journal for Educational Integrity, tested a range of detectors and concluded they are "neither accurate nor reliable," with even the best scoring below 80% (Weber-Wulff et al., 2023).

Turnitin furthermore allows individual universities to toggle their AI detector on/off, meaning many students and lecturers will not have an AI % returned to them.

Below are publicly available accuracy rates for a range of detectors (as determined by):

  •  GPTZero: Publishes high vendor accuracy figures; treat these as vendor claims and compare them against independent tests and your own results.

GPTZero advertises accuracy above 99% on its own benchmarks; in the University of

Chicago study (Jabarian & Imas, 2025) it ranked a tier below Pangram, and in our spot

check it flagged both human samples, so weigh its accuracy claim against that false-

positive tendency.

  • Pangram Labs: A University of Chicago working paper by Jabarian and Imas (2025) found Pangram produced near-zero false positive and false negative rates on medium and long passages, ahead of GPTZero and Originality.ai (NBER working paper; Chicago Booth Review).

  • Originality.ai: Scribbr's 2023 test found Originality flagged human text that had been run through a paraphrasing tool in about 60% of cases, which means human writing passed through a paraphraser can be flagged (Scribbr, 2023 test).

For a closer comparison of the two tools students ask about most, see Turnitin vs GPTZero accuracy.

We do not rank Phrasly in this section or declare an overall winner.

What Your AI Score Means and What to Do After a Flag

Read a score as a band, not a verdict:

  • Low: The model found few AI-like patterns. This is not a guarantee of human authorship.

  • Mixed: The result is ambiguous. This band is common for edited, formulaic, short, or ESL writing.

  • High: The model found strong AI-like patterns. It is still an estimate that needs supporting evidence before anyone acts on it.

A percentage such as 32% will represent either the likelihood that the text has been AI-generated or the proportion of the text highlighted by the tool - but not the amount of "cheating". You should check what the % refers to.

The sentences highlighted will indicate portions of the writing that were considered low variance or predictable by the model. It does not mean they were plagiarized, nor does it indicate the origin.

Small error rates still matter at scale. A 1% error rate across 10,000 students means about 100 people wrongly flagged.

For where a reasonable line sits, see how much AI detection is acceptable.

Save Your Drafts and Version History

If your work has been detected as AI-written, despite it being 100% your own, the best

thing to have on hand is process evidence. Further AI detection is unlikely to be helpful.

Make sure you have a copy (export/screenshot) of your Google Docs/Word document history. Assemble the following:

  • Version history and earlier drafts

  • Research notes and sources

  • Outlines and file timestamps

  • Prior writing samples in your own voice

Ask Which Tool and Threshold Were Used

Ask (politely) which detector was used, what reading was returned, and if the professor received a % or flagged individual sentences. You can then find the false positive rate of that tool as listed above.

If you are a non-native speaker of English, you may also want to mention the paper by Stanford (2023) demonstrating detector bias towards non-natives, again politely.

Request a Conversation, Not a Verdict

Request a sit-down to go through the various drafts and verbally defend your work. Refer to your university's policy on academic misconduct to be sure you are following the correct procedure.

Nobody can guarantee an appeal's outcome, but having a procedure to follow is the best position to be in.

When marking, view a low grade as an opening to discuss the student's work. Inspect drafts and question the student orally regarding their line of reasoning. Then proceed in line with the university's academic misconduct policy.

Check Your Own Draft Before You Submit

Phrasly has an AI detector, which has been used (and disclosed) in the tests above, and we have reported the outcome with the other 5 detectors.

You can run your writing through this and identify any low-perplexity lines, reword these yourself, and continue to build up your writing samples.

On our spot-check, Phrasly passed 2/2 human and detected 4/4 AI articles. The only detector that didn't make a false positive or false negative.

Of course, with just 1 sample in each group, that doesn't make it definitively the best performing - refer to the 3rd party testing above.

Reading a score well also means knowing where a reasonable line sits, so see how much AI detection is acceptable before you act on any number.

Phrasly’s AI detector is free to use, and please see Phrasly’s pricing page (phrasly.ai/pricing).



FAQs

Why don't AI detectors work?

They rely on probabilistic similarities between human and machine-generated writing. In July 2023, OpenAI retired its AI Text Detector due to poor performance, and at least one University (Vanderbilt) has chosen to disable this function.

Any detector can only provide a likelihood that the text was AI generated.

Does paraphrasing AI text get past AI detectors?

The extent to which the AI score can be reduced depends on the tool and degree of rephrasing.

From a quick test of our samples, rephrasing an AI-generated sample was not enough to evade detection, with all tools flagging it as >96% AI. Our sample of ~50% human / 50% AI text was not detected as AI by Originality.ai, Pangram, and ZeroGPT (<50% AI).

Hence, combining human and AI-written text was successful in passing these AI detectors.

How often are AI detectors wrong?

Depends on the text. Also, need to distinguish between false positives (flagging human-written) and false negatives (letting AI-written through). Frequently enough to preclude a single tool from being determinative.

As can be seen from our brief sample, our false positives occurred against the ESL and human-authored passages (5/6 and 1/6, respectively), whereas the majority of our false negatives were against our human+AI passage (3/6).

These rates will increase further for ESL, human+AI, shorter or heavily-edited text - all common to students.

Can AI detectors wrongly flag my writing as AI?

Yes! Most at risk are non-native English writers, plus anyone using formulaic, template-like, or heavily grammar-corrected prose.

Liang et al. (2023) documented this bias against non-native writers (paper).

Keep your drafts and version history, and see why your essay was detected as AI.

Are AI detectors accurate or reliable?

Here are some further questions on this. As noted above, accuracy describes how frequently a detector correctly identifies a dataset. Consistency refers to whether different detectors (or the same detector with different versions/lengths) return the same result.

Therefore, a detector may be accurate but not consistent for a given use-case, and should not be used to inform a high-impact decision.

If English is your second language, you can protect yourself from AI detector false positives by keeping drafts.

Written by

Uzair Khan

Uzair Khan, Senior SEO Strategist at Phrasly AI, develops content and SEO strategies that make AI detection and humanization technology accessible to students, educators, and content creators.

Share this article