AI Detector Accuracy Statistics 2026
By Axis Intelligence Research and Sarah Mitchell | Last updated: June 16, 2026 | Next scheduled update: Q3 2026 (September) | License: CC BY 4.0
Quick Answer
A foundational 2023 Stanford study published in Patterns (Cell Press) found that seven widely-used AI detectors flagged 61.22% of genuine human-written TOEFL essays by non-native English speakers as AI-generated, compared to roughly 3.2% for native-English essays from the same testing framework — a roughly 19-to-1 disparity. According to Axis Intelligence’s cross-paper analysis, that same 91-essay benchmark, when re-tested by Pangram Labs in 2024, produced a 0% false positive rate, while an updated GPTZero model reached 7.7% — both dramatic improvements, though Pangram’s near-elimination of the gap remains an outlier rather than an industry norm. Separately, Ahrefs’ April 2025 analysis of 900,000 web pages found that 74.2% of newly published pages now contain some AI-generated content, with only 25.8% classified as purely human-written.
Key Findings
- Stanford’s 2023 study found a 61.22% false positive rate for non-native English TOEFL essays, versus approximately 3.2% for native-English essays — a roughly 19-to-1 disparity across seven widely-used AI detectors, published in Patterns (Cell Press).
- According to Axis Intelligence, this 19.1x Non-Native English Bias Multiplier dropped to approximately 2.4x for an updated GPTZero model and to 0x for Pangram, when both were tested on the identical 91-essay TOEFL benchmark in 2024 — the same test, one year later, with radically different outcomes depending on the detector.
- According to Axis Intelligence’s calculation, the same 2023 study’s own data reveals a Detector Disagreement Index of 78.02 percentage points: 97.8% of TOEFL essays were flagged by at least one of seven detectors, but only 19.78% were flagged by all seven — meaning detectors agreed with each other on fewer than one in five cases.
- OpenAI discontinued its own AI Text Classifier in July 2023, just six months after launch, after it correctly identified only 26% of AI-written text while mislabeling 9% of human-written text as AI-generated — an acknowledgment from the company that built the underlying models that reliable detection was not yet achievable.
- Saha and Feizi’s 2025 University of Maryland study found that minimally polishing AI-generated text with GPT-4o caused detection rates across different detectors to fall anywhere between 10% and 75%, while Ahrefs’ 2025 analysis of 900,000 web pages found 74.2% now contain some AI-generated content.
As AI-generated text has gone from novelty to majority share of new web content — content produced by the same data-center infrastructure Axis Intelligence tracks in its AI data center energy consumption statistics hub — a parallel industry has grown around detecting it, and a parallel body of academic research has grown around testing whether that detection actually works. The headline finding from the foundational study in this space has not aged well for the industry: in 2023, Stanford researchers found that seven widely-used AI detectors misclassified genuine human writing as “AI-generated” at a rate roughly 19 times higher for non-native English speakers than for native speakers. By 2024, at least one newer detector reduced that same benchmark’s false positive rate to zero.
This report compiles verified data from four peer-reviewed and preprint studies hosted on arXiv — Liang et al.’s foundational 2023 Stanford/Patterns study, Pangram Labs’ 2024 technical report, Saha and Feizi’s 2025 University of Maryland study on AI-polished text, and a December 2025 study on detector robustness against polished writing — alongside Ahrefs’ 2025 analysis of AI-generated web content at scale and the historical record of OpenAI’s own discontinued AI Text Classifier. It also introduces two original Axis Intelligence cross-source metrics: the Non-Native English Bias Multiplier and the Detector Disagreement Index.
The Foundational Bias Study: Stanford’s TOEFL Findings
The single most-cited study in AI detection research tested seven widely-used GPT detectors on two sets of genuine human writing: 91 TOEFL (Test of English as a Foreign Language) essays written by non-native English speakers, and 88 essays written by US 8th-graders, sourced from the Hewlett Foundation’s Automated Student Assessment Prize dataset. Published in Patterns by Liang, Yuksekgonul, Mao, Wu, and Zou, the study’s central finding was stark: the detectors demonstrated near-perfect accuracy on the US 8th-grade essays — a false positive rate of approximately 3.2% — but misclassified more than half of the TOEFL essays as AI-generated, an average false positive rate of 61.22% across the seven detectors.
The study’s secondary findings are arguably more concerning for anyone relying on a single detector’s verdict. Of the 91 TOEFL essays — every one of them written entirely by a human — 19.78% were unanimously flagged as AI-generated by all seven detectors tested, while 97.80% were flagged by at least one. The researchers traced this to perplexity: TOEFL essays unanimously misclassified as AI-generated had significantly lower perplexity (statistical predictability) than other essays, with a p-value of 9.74×10⁻⁵ — meaning AI detectors are, in effect, partly measuring how “formulaic” or “simple” a piece of writing is, a pattern that correlates with second-language writing as much as it does with AI generation.
According to Axis Intelligence: The Non-Native English Bias Multiplier
Liang et al.’s study reports two false positive rates — 61.22% for non-native writers and approximately 3.2% for native writers — but does not itself express this as a ratio. Axis Intelligence’s Non-Native English Bias Multiplier does: dividing the non-native rate by the native rate produces a multiplier of approximately 19.1x, meaning a non-native English speaker’s genuine writing was roughly 19 times more likely to be falsely flagged as AI-generated than a native speaker’s, across the seven detectors tested in 2023.
What makes this Index genuinely cross-source is what happened next. Pangram Labs’ 2024 technical report deliberately re-ran the identical 91-essay TOEFL benchmark from Liang et al. — and reported a 0% false positive rate, attributing this to having held the benchmark essays out of its training data entirely. The same report notes that GPTZero published a response to the original finding (titled “ESL Bias in AI Detection is an Outdated Narrative”) alongside an updated model, which the Pangram report found still produced a 7.7% false positive rate on the same benchmark (or 1.1% if “Possible AI Content” classifications are counted as negative).
| Detector / Version | Year | FPR on 91-Essay TOEFL Benchmark | Axis Bias Multiplier (vs. ~3.2% native baseline) |
|---|---|---|---|
| 7-detector average (Liang et al.) | 2023 | 61.22% | 19.1x |
| GPTZero (updated, post-ESL-debiasing) | 2024 | 7.7% | 2.4x |
| Pangram | 2024 | 0% | 0x |
According to Axis Intelligence’s cross-paper analysis, the same 91-essay benchmark — held constant across two independent papers published roughly a year apart — produced bias multipliers ranging from 19.1x down to 0x, depending entirely on which detector and which model version was tested. This is simultaneously good news (the bias is not an immutable property of “AI detection” as a category) and a caution (most production detectors in 2024-2026 have not published equivalent zero-bias results on this benchmark, and the 7.7% GPTZero figure — itself a major improvement over the 2023 average — is still 2.4 times the native-speaker baseline).
Methodology and stated assumptions:
- Native baseline held constant. The “~3.2%” native-English baseline comes from Liang et al.’s 2023 testing only; neither the Pangram nor updated-GPTZero figures include a re-tested native-English baseline from 2024. The multiplier therefore assumes the native baseline did not change materially — a reasonable assumption given native-English FPRs were already near zero and have limited room to fall further, but not independently confirmed for 2024.
- Single benchmark, not a general claim. Both 2024 figures come from one 91-essay benchmark. Pangram’s own report frames this as a best-case result partly attributable to training-data exclusion, not as a claim that Pangram achieves 0% FPR on all non-native writing universally.
- “7-detector average” vs. individual detectors. The 2023 figure of 61.22% is an average across seven detectors of varying quality; individual 2023 detectors likely ranged both above and below this figure, a distribution not preserved in this Index.
According to Axis Intelligence: The Detector Disagreement Index
Liang et al.’s 2023 study reports two figures about the same 91 TOEFL essays that, read together, describe something the paper itself does not name: how often the seven detectors agreed with each other. 97.80% of the essays were flagged as AI-generated by at least one detector; only 19.78% were flagged by all seven. Axis Intelligence’s Detector Disagreement Index is the gap between these two figures: 97.80% − 19.78% = 78.02 percentage points.
According to Axis Intelligence’s calculation, for more than three-quarters of the essays in this benchmark, whether a given essay was “flagged as AI” depended on which specific detector was used — not on any property of the essay that all seven detectors agreed on. An instructor or platform using Detector A might flag an essay that Detector B clears, with no principled way to know which detector is “right” for a given case, since all seven were testing genuine human writing.
This disagreement pattern is not unique to 2023 or to non-native English specifically. The December 2025 study on detector robustness against polished writing, discussed below, found that detectors tuned to flag one AI model’s output respond very differently to text polished by a different model — meaning the same underlying inconsistency (detectors disagreeing based on stylistic artifacts rather than a shared, principled definition of “AI-generated”) persists as a structural feature of the field, even as bias against any single population (such as non-native English writers) has, per the Non-Native English Bias Multiplier above, been addressed by at least one detector.
Explore the Data
From OpenAI’s Discontinued Classifier to Today’s Landscape
Context for how far the field has come — and how far it has not — starts with the company that builds the models being detected. OpenAI launched an AI Text Classifier on January 31, 2023, designed to assign a probability rating indicating whether a passage of at least 1,000 characters was AI-written. By its own published evaluation, the classifier correctly identified only 26% of AI-written text as “likely AI-written” while mislabeling 9% of human-written text as AI-generated. Less than six months later, on July 20, 2023, OpenAI quietly discontinued the tool, adding a note to its original announcement stating it was “no longer available due to its low rate of accuracy” — even as the same company’s ChatGPT, tracked in Axis Intelligence’s ChatGPT statistics hub, was already becoming the dominant source of the AI-generated text such classifiers were built to catch.
The significance of this episode is not that OpenAI’s specific classifier was bad — early detectors generally were — but that the company with the deepest access to how its own models generate text concluded that reliable detection, at least with the techniques available in early 2023, was not achievable enough to keep a product live. The detectors that remained in the market (GPTZero, Originality.ai, Copyleaks, Turnitin, and others) continued operating and improving, and the 2024 Pangram benchmark above shows that meaningful improvement on at least one well-known bias problem is achievable. But OpenAI’s withdrawal — from a company whose business model would benefit from credible AI-content provenance tools — remains a useful anchor for how difficult this problem was understood to be at the outset of the current detection landscape.
Evasion and Robustness: What Happens When AI Text Is Polished
The bias findings above concern detectors evaluated on unedited text — either genuinely human-written (TOEFL essays) or, in most detector-accuracy studies, raw AI output. A separate and growing body of research asks a different question: what happens to detection accuracy when AI-generated text is lightly edited, or “polished,” before being submitted?
Saha and Feizi’s 2025 University of Maryland study, “Almost AI, Almost Human: The Challenge of Detecting AI-Polished Writing,” found that detectors frequently flag even minimally polished text as AI-generated and struggle to differentiate between degrees of AI involvement — but the more striking finding for evasion purposes is the inverse: minimally polishing AI-generated text with GPT-4o caused detection rates across different detectors to fall anywhere between 10% and 75%, a 65-percentage-point spread depending entirely on which detector was used.
A December 2025 follow-on study testing detector robustness against text polished by multiple different models — GPT-4o, DeepSeek 3.1, Mistral, Kimi K2, LLaMA-4, and Claude-4-Sonnet — found that a detector tuned for GPT-4o output maintained 99.2% accuracy on unpolished GPT-4o text, but dropped to 95.26% once that text was lightly polished (10%) using LLaMA-3-70B — a 3.94-percentage-point accuracy decline from a single round of light editing using a different model than the one the detector was tuned for. Separately, a DeepSeek 3.1-targeted detector showed a baseline false positive rate of 3.13% on unedited text, broadly consistent with the “near-zero for native/unedited content” pattern established by the Pangram benchmark above — but the same study found this detector’s performance degraded substantially once content was polished using cross-model inputs, reinforcing the Detector Disagreement Index’s core finding: detector performance is highly sensitive to which specific model produced or touched the text, in ways that are not yet standardized across the industry.
The Scale of the Problem: How Much Web Content Is AI-Generated
The accuracy and bias findings above matter in proportion to how much content is actually being evaluated. Ahrefs analyzed 900,000 newly-published English-language web pages in April 2025 using its own AI content classifier, finding that 74.2% of these pages contained some AI-generated content — 2.5% were classified as “pure AI,” 71.7% as a human-AI mix, and only 25.8% as “pure human.”
Read against the bias findings above, this scale creates a genuine tension. If three-quarters of new web content contains AI involvement, and detector verdicts on genuinely human-written text vary by up to 78.02 percentage points depending on which detector is used (the Detector Disagreement Index), then any platform, publisher, or institution applying a single detector’s threshold at scale is making a large number of decisions — about real people’s writing — using a signal that two equally “leading” detectors might score differently for the same piece of text. An increasing share of the AI side of that 74.2% is no longer a person prompting a chatbot document-by-document: Axis Intelligence’s AI agents statistics hub tracks the rise of autonomous systems generating content at scale, which removes even the loose signal of “a person sat down and wrote a prompt for this.” This connects directly to Axis Intelligence’s generative AI statistics hub, which tracks the broader adoption curve this detection challenge sits downstream of, and to the AI governance statistics hub, which tracks policy responses — including, in some jurisdictions, restrictions on using AI-detection scores as the sole basis for academic or employment decisions.
Methodology
This report compiles data from six primary sources, four of which are peer-reviewed or preprint papers hosted on arXiv: Liang et al.’s 2023 study published in Patterns (Cell Press), testing seven AI detectors on TOEFL and US 8th-grade essays; Pangram Labs’ 2024 technical report, which re-tested the identical 91-essay TOEFL benchmark; Saha and Feizi’s 2025 University of Maryland study on detecting AI-polished writing; and a December 2025 study testing detector robustness against text polished by six different AI models. These are supplemented by Ahrefs’ April 2025 analysis of 900,000 web pages for AI-content prevalence, and the historical record of OpenAI’s AI Text Classifier (January-July 2023). All figures were live-verified via direct source retrieval on June 15, 2026.
The two original Axis Intelligence elements — the Non-Native English Bias Multiplier and the Detector Disagreement Index — are both derived from figures explicitly reported in Liang et al.’s 2023 paper (and, for the Multiplier’s 2024 data points, Pangram Labs’ 2024 report testing the same benchmark), recombined into ratios and gaps that neither paper computed directly. Each element’s calculation section above states the specific assumptions made. Given how rapidly individual detectors update their models — Pangram’s report documents GPTZero updating its model specifically in response to the 2023 finding — both elements should be read as benchmarks of specific model versions at specific points in time, not permanent properties of “AI detection” as a category.
About This Dataset
The complete dataset behind this report — 27 individual data points spanning bias/fairness findings, detector performance benchmarks, evasion/robustness data, web-content-scale statistics, and both original Axis Intelligence indices — is available for download as a CSV file under a CC BY 4.0 license, free to use, share, and adapt with attribution to Axis Intelligence. Each row includes the metric, value, unit, scope, time period, source organization, and source URL for independent verification.
This dataset will be refreshed as new benchmark studies are published. The 91-essay TOEFL benchmark from Liang et al. (2023) has already been re-tested once (Pangram Labs, 2024); Axis Intelligence will track any future re-tests of this same benchmark by other detectors, as each provides a directly-comparable data point for the Non-Native English Bias Multiplier. Ahrefs’ AI-content-prevalence analysis is conducted periodically (most recently April 2025); the next available edition will update the Scale/Adoption figures. Both Axis original indices will be recalculated whenever a new detector publishes results on the same 91-essay benchmark.
How to Cite This Research
APA: Axis Intelligence Research & Mitchell, S. (2026). AI Detector Accuracy Statistics 2026: False Positive Rates, Bias, and Detection Reliability Data. Axis Intelligence. https://axis-intelligence.com/ai-detector-accuracy-statistics/
MLA: Axis Intelligence Research, and Sarah Mitchell. “AI Detector Accuracy Statistics 2026: False Positive Rates, Bias, and Detection Reliability Data.” Axis Intelligence, 2026, axis-intelligence.com/ai-detector-accuracy-statistics/.
Chicago: Axis Intelligence Research, and Sarah Mitchell. “AI Detector Accuracy Statistics 2026: False Positive Rates, Bias, and Detection Reliability Data.” Axis Intelligence. 2026. https://axis-intelligence.com/ai-detector-accuracy-statistics/.
Frequently Asked Questions
What is the false positive rate for AI detectors on non-native English writing?
A 2023 Stanford study published in Patterns found that seven widely-used AI detectors flagged 61.22% of genuine human-written TOEFL essays by non-native English speakers as AI-generated, compared to approximately 3.2% for native-English essays from US 8th-graders.
What is the Axis Non-Native English Bias Multiplier?
It is an Axis Intelligence original metric dividing the non-native false positive rate by the native false positive rate from the same 2023 study (61.22% / 3.2% ≈ 19.1x). When the identical 91-essay benchmark was re-tested in 2024, this multiplier fell to approximately 2.4x for an updated GPTZero model and to 0x for Pangram.
How much did AI detector bias improve from 2023 to 2024?
On the identical 91-essay TOEFL benchmark used in the original 2023 study, Pangram Labs reported a 0% false positive rate in its 2024 technical report, and an updated GPTZero model reported 7.7% (or 1.1% under generous counting) — both substantial reductions from the 2023 seven-detector average of 61.22%.
What is the Detector Disagreement Index?
It is an Axis Intelligence original metric calculated from the 2023 Stanford study’s own data: 97.80% of TOEFL essays were flagged as AI-generated by at least one of seven detectors, but only 19.78% were flagged by all seven — a gap of 78.02 percentage points, indicating that for most essays, the “AI-generated” verdict depended on which specific detector was used.
Why did OpenAI discontinue its AI text classifier?
OpenAI launched its AI Text Classifier on January 31, 2023, and discontinued it on July 20, 2023, after it correctly identified only 26% of AI-written text while mislabeling 9% of human-written text as AI-generated — a low accuracy rate that OpenAI itself cited when withdrawing the tool.
What happens to AI detection accuracy when text is polished?
Saha and Feizi’s 2025 University of Maryland study found that minimally polishing AI-generated text with GPT-4o caused detection rates across different detectors to fall anywhere between 10% and 75%. A December 2025 follow-on study found a GPT-4o-tuned detector’s accuracy dropped from 99.2% on unpolished text to 95.26% after light (10%) polishing using a different model (LLaMA-3-70B).
How much web content is AI-generated in 2025?
Ahrefs’ analysis of 900,000 newly-published web pages in April 2025 found that 74.2% contained some AI-generated content — 2.5% were classified as “pure AI,” 71.7% as a human-AI mix, and 25.8% as “pure human.”
Are AI detectors reliable enough for academic discipline decisions?
The data in this report suggests caution: the Detector Disagreement Index shows detectors agreed with each other on fewer than one in five cases for genuine human writing in the foundational 2023 study, and the Non-Native English Bias Multiplier shows historical bias against non-native English speakers, though at least one 2024 detector (Pangram) reported eliminating this specific bias on the benchmark tested.
Which AI detector has the lowest false positive rate on record?
Pangram Labs reported a 0% false positive rate on the 91-essay TOEFL benchmark from Liang et al. (2023) in its 2024 technical report, the lowest figure in the data compiled for this report, attributed to excluding the benchmark essays from its training data.
What causes AI detectors to flag human writing as AI-generated?
The 2023 Stanford study found that essays unanimously misclassified as AI-generated had significantly lower perplexity (statistical predictability) than other essays (p = 9.74×10⁻⁵) — meaning detectors partly measure how “formulaic” writing is, a pattern that correlates with simpler vocabulary and sentence structures common in second-language writing, independent of whether AI was involved.