Quick answer: Pangram AI Detector reports very low error rates in its 2026 tests, but its accuracy depends on the model version and the writing being checked. Pangram 4 reported a 0.0041% false positive rate on one million English human documents. Pangram 4.1 reported 0.0090% on a different, filtered English sample. Neither result proves who wrote an individual document.
The most useful way to judge Pangram is to examine the testing data, understand what its scores mean, and consider the limitations that apply to your own text. Pangram 4.1, released on October 19, 2026, also changes how uncertain results are communicated.
| Quick question | Direct answer |
|---|---|
| Is Pangram accurate? | Its developer reports low benchmark error rates, not guaranteed accuracy for every document. |
| Latest version covered | Pangram 4.1, released October 19, 2026 |
| Minimum supported length | 50 words of mainly natural-language prose |
| Can it distinguish AI editing? | It labels passages Human, AI-Assisted, or AI-Generated. |
| Does a flagged score prove AI use? | No. A score is a model assessment, not authorship evidence. |
How Accurate Is Pangram AI Detector in 2026?
Pangram’s reported accuracy is strong on its defined evaluation sets, particularly for long-form English writing. The latest published model card gives Pangram 4.1 an English false positive rate of 0.0090% and a false negative rate of 0.3652%. Those numbers come from different evaluation populations, so they cannot be combined into one universal accuracy score.
Pangram 4’s earlier English results were 0.0041% false positives among one million known human documents and 0.3396% false negatives among 519,993 AI documents. The newer false positive evaluation excludes certain material outside the detector’s intended scope. Comparing those headline rates without their sample definitions would be misleading.
| Published English Test | Pangram 4 | Pangram 4.1 | Interpretation |
|---|---|---|---|
| Original Pangram 4 human sample, 1,000,000 documents | 0.0041% false positives | Not reported on this identical original sample | Earlier benchmark only |
| Matched filtered human sample, 781,981 documents | 0.0164% false positives | 0.0090% false positives | Comparable within this filtered set |
| English AI generations, approximately 520,000 documents | 0.3396% false negatives | 0.3652% false negatives | Slightly different sample counts |
All percentages above are developer-reported. The original and filtered English human datasets must remain separately labeled, even though they both concern false positives.
What Do False Positives and False Negatives Mean?
A false positive wrongly flags human writing as AI, while a false negative misses text known to be AI-generated. These measures use different groups of documents as their denominators. For someone accused of using AI, the false positive risk is particularly important. A low false positive rate is not the probability that a specific flagged document was written by AI.
In Pangram 4’s original English human sample, 41 of one million documents were incorrectly classified. That observation describes the benchmark population. It does not tell a university, publisher, or employer how to decide the authorship of a particular submission without other evidence.
Is Pangram 4.1 More Accurate Than Pangram 4?
Pangram 4.1 improved two reported measures but did not outperform Pangram 4 on every metric. On the matched filtered English human dataset, its false positive rate fell from 0.0164% to 0.0090%. Its English false negative rate, however, rose from 0.3396% to 0.3652%. These differences should not be condensed into a claim of universally better detection.
According to the official Pangram 4.1 model card, the update also adds probabilities for the three segment labels and changes the dashboard’s presentation of uncertain text. The developer published these results on October 19, 2026. They are model evaluation evidence, not an independent guarantee for new material.
Does Pangram 4.1 Produce More Stable Scores After Editing?
Yes, in Pangram’s paired-edit test, unintended changes to the overall verdict became less frequent. Across 29,103 paired edits, the reported verdict flip rate decreased from 1.08% in Pangram 4 to 0.15% in Pangram 4.1. The newer interface also communicates less certainty for documents below 150 words.
A stable prediction is not necessarily a correct prediction. Stability describes whether small changes in input produce unexpected changes in the result. Accuracy describes whether the result matches the document’s known origin. Reviewers should assess those properties separately.
What Does a Pangram AI Detection Score Actually Mean?
A Pangram score describes how the detector classified the submitted writing, not the probability that an author used AI. Pangram assigns Human, AI-Assisted, or AI-Generated labels to segments of text. Its document fractions summarize the share of analyzed characters assigned to each category. Pangram 4.1 also provides model probabilities at the segment level.
| Displayed Result | What It Indicates | What It Does Not Establish |
|---|---|---|
| 95% AI-Generated fraction | Approximately 95% of analyzed characters received that category. | 95% certainty that the author used AI |
| AI-Assisted passage | Text patterns match the detector’s assisted-writing category. | The precise editing tool or action taken |
| Human document label | The detector favored a human classification. | That no AI assistance occurred |
| Mixed document label | Different passages received different classifications. | A verified history of how the document was produced |
Scores from separate detectors should not be treated as identical measurements. For instance, the Grammarly AI Checker guide explains how another detection product presents percentages and why results from different systems can disagree.
Does a 100% AI Score Prove Someone Used ChatGPT?
No. A 100% AI-Generated classification means Pangram labeled all analyzed text as generated, not that it observed ChatGPT being used. Classification does not identify a particular person, establish a writing process, or reliably pinpoint the model involved. The result may justify closer examination, but supporting records matter when the stakes are high.
Can Pangram Mistake Human Writing for AI?
Yes. Pangram can produce false positives even when its published benchmark error rates are low. Writing style, genre, sample length, formatting, and the detector version can influence a classification. A flagged essay or article should therefore be reviewed with its drafts and context rather than treated as conclusive evidence of misconduct.
Pangram 4 reported a 0.019% false positive rate for 62,971 English academic documents and 0.078% for 12,769 poetry samples. These are distinct evaluation sets, which helps explain why the original 0.0041% English headline rate cannot represent every kind of writing equally well.
Does Pangram Accuracy Vary Across Languages?
Yes. Pangram supports multiple languages, but its published tests show different error rates across the evaluated material. In Pangram 4’s AI-generated datasets, reported false negatives included 0.8867% for Spanish, 1.3682% for Hindi, 3.1715% for Persian, and 5.3169% for Urdu. These were separate test sets rather than matched translations of identical documents.
Pangram 4.1 reports support for English, Arabic, Chinese, French, German, Hindi, Urdu and many other languages. Organizations reviewing multilingual writing should examine results from the relevant language and version before choosing a detection threshold or policy.
Is Pangram Reliable for Non-Native English Writers?
Pangram has published encouraging results for non-native English writing, but the available sample is too small to settle fairness questions. Pangram 4.1 reported no false positives on a third-party benchmark containing 91 human-authored non-native English documents. That finding is useful, yet it cannot guarantee the same result for writers with different proficiency levels or writing styles.
English written by a second-language learner is not the same evaluation problem as a document written entirely in Urdu or another language. Institutions should avoid drawing one conclusion from unrelated language datasets.
Can Pangram Detect AI-Assisted and Humanized Writing?
Pangram is designed to identify mixed human and AI writing as well as text modified by some AI rewriting tools. Its segment analysis separates Human, AI-Assisted, and AI-Generated passages. This gives reviewers more information than a single document label, but the tool cannot reconstruct every action taken during drafting, proofreading, or rewriting.
Minor spelling corrections, a rewritten introduction, and a fully generated paragraph represent different types of assistance. Pangram’s predictions may respond differently to each pattern. A report should be interpreted in light of what the relevant academic or editorial policy actually permits.
Why Can Pangram Scores Change After Editing?
Editing changes the text patterns and context supplied to the classifier, so a new scan may return different segment labels. Moving a paragraph into a longer document can also change its surrounding context. To compare results fairly, record the same detector version, input format, and text selection rather than assuming every difference reflects a change in true authorship.
What Does Independent Research Say About AI Detector Accuracy?
Peer-reviewed research shows why detector performance should be tested under consistent conditions rather than judged by marketing percentages. A 2025 Association for Computational Linguistics study by Jenna Russell, Marzena Karpinska, and Mohit Iyyer compared automated tools with human readers on 300 nonfiction English articles. The research tested earlier detector versions, not Pangram 4.1.
The study reported that the majority vote of five frequent users of language models misclassified only one of the 300 articles. That result concerns this particular study design and participant group. It does not show that human readers always outperform automated systems, or that current Pangram models would produce the same outcomes.
How Can You Test Pangram Accuracy Fairly?
A meaningful test needs writing with verified origins, one recorded detector version, and separate counts for each kind of error. Include human work, unedited AI output, extensively revised passages, and mixed documents. Choose samples that resemble the writing you expect to check, then document genre, length, language, and model source before interpreting the results.
- Human sample: Use documents supported by reliable drafts or writing records.
- AI sample: Save original generated text and identify the generator used.
- Mixed sample: Document which sections were drafted or rewritten with assistance.
- Consistent test: Keep the detector version and evaluation settings fixed.
- Separate outcomes: Count false positives and false negatives rather than reporting only overall accuracy.
Researchers planning a multi-model test can consult Primingo’s ChatGPT, Claude and Gemini comparison for background on different AI writing systems. It is a product comparison, not a scientific detector benchmark.
Does Text Length or File Format Affect Pangram Accuracy?
Yes. Pangram 4.1 is intended mainly for natural-language passages of at least 50 words, preferably written in complete sentences. Text under 150 words receives a more cautious uncertainty presentation in the newer interface. Code, markup, reference lists, templates, and text dominated by mathematics fall outside the detector’s intended use or may produce less reliable results.
| Input Type | Practical Guidance |
|---|---|
| Complete essay or article | Suitable when it contains representative natural-language prose |
| Fewer than 50 words | Below the model’s stated minimum |
| 50 to 149 words | Supported but displayed with increased uncertainty |
| Code, formulas, or reference lists | Outside the intended scope or less reliable |
| Check for text extraction errors | |
| DOCX or clean plain text | Preferred where available |
Before comparing results from different files, verify that the detector received the same words in the same order. Extraction errors can create apparent inconsistencies that have nothing to do with the underlying model.
How Does Pangram Compare With GPTZero and Turnitin?
No current evidence establishes one universally more accurate detector across every language, document type, and model version. Pangram emphasizes detailed mixed authorship analysis. GPTZero focuses on dedicated AI detection workflows, while Turnitin is commonly used within institutional academic integrity systems. A meaningful accuracy ranking requires current versions tested against the same known-origin documents.
| Tool | Main Workflow | What to Check |
|---|---|---|
| Pangram | Mixed authorship and detailed passage classification | Detector version, language, and false positive performance |
| GPTZero | Dedicated AI detection for individual and education use | Document length, score interpretation, and test coverage |
| Turnitin | Academic integrity workflows within institutions | Institutional policy, report scope, and review process |
| Grammarly | Writing assistance with AI detection options | Availability of detection features in the selected plan |
For a deeper tool-by-tool discussion, the Pangram vs Other AI Detectors article examines use cases and practical differences. Those comparisons still need careful version matching before treating numerical scores as directly comparable.
What Should You Do If Pangram Incorrectly Flags Your Writing?
Preserve the detection report and evidence of your writing process before challenging a suspected false positive. Check that the scanned text matches the document you submitted, then gather earlier drafts, research notes, and revision history. Where an academic or professional policy applies, request human review under that policy rather than arguing from a single percentage.
- Save the original Pangram report and identify the detector version when available.
- Preserve dated drafts, outlines, sources, and document revision history.
- Check whether file conversion removed or altered text.
- Record any proofreading, translation, or AI rewriting that was permitted.
- Ask the appropriate reviewer to consider the work and process evidence together.
- Report a suspected false positive to Pangram when doing so is appropriate.
A lower score after rewriting is not proof that the original accusation was correct. Revision changes the classifier’s input. The meaningful question is whether the person followed the relevant authorship or AI assistance rules.
Is Pangram Worth Using for Regular AI Detection?
Pangram is worth evaluating when you need repeatable text checks, detailed passage labels, multilingual coverage, or an established review workflow. Its Free plan provides up to 2,000 words per day. Direct Individual and Professional plans are listed at $20 and $65 per month respectively. Those prices describe official plan tiers, not independent retailer offers.
Scanning volume and associated features matter more than a single impressive benchmark percentage. The Pangram Pricing guide explains standard word allowances, image checks, and the distinction between personal and professional capacity.
For buyers comparing retail access separately from Pangram’s direct plans, get Pangram Professional access on Primingo after reviewing the selected duration, account terms, and delivery conditions. Primingo is an independent retailer; the offer should not be mistaken for a benchmark endorsement or for identical direct-provider entitlements.
Readers comparing software subscriptions for their writing workflow can explore Primingo. The purchase decision should remain separate from the question of whether a detector result is supported by suitable evidence.
Frequently Asked Questions About Pangram AI Detector Accuracy
Can Pangram Detect Writing From ChatGPT, Claude, and Gemini?
Yes. Pangram has published evaluations involving major text generator families, including systems from OpenAI, Anthropic, and Google. Successful detection varies with the specific generator, the content, language, and amount of editing. Its ability to flag AI-like writing does not mean the detector can reliably name the exact model that created a passage.
Does a Human Result Guarantee That No AI Was Used?
No. A Human result means the detector assigned a human classification to the analyzed material. AI-generated or edited content may occasionally escape detection, and individual documents do not inherit guarantees from aggregate benchmark rates. Where a decision matters, writing history and other evidence provide context that a classifier cannot supply.
Is Pangram AI Detection the Same as Plagiarism Detection?
No. AI detection classifies patterns that resemble human or machine writing, while plagiarism checks look for overlap with existing material. Newly generated AI prose can be original in wording, and a manually copied human paragraph can be plagiarized. These checks answer separate questions and should not be interpreted as interchangeable scores.
Can Teachers Use Pangram Alone to Prove a Student Used AI?
No. A Pangram classification alone cannot establish academic misconduct or reconstruct how an assignment was produced. Teachers should consider drafts, revision history, permitted editing tools, relevant course rules, and the student’s explanation. Institutions should use a fair review process before making decisions that affect grades or disciplinary records.
Can Pangram Reliably Detect Very Short Answers or Source Code?
Pangram 4.1 is not designed primarily for source code or very short factual responses. It requires at least 50 words of suitable input, and the dashboard reflects greater uncertainty below 150 words. For code, mathematical notation, or documents dominated by reference lists, its reported text detection benchmark figures should not be assumed to apply.
