Pangram AI Detector: How It Works and What Scores Mean (2026)

Pangram is a newer AI text detector that claims 99.98% accuracy and a false positive rate of about 1 in 10,000. It says it spots AI writing from patterns in style and word choice, and it labels text as human, AI or AI-assisted. A 2025 working paper from University of Chicago Booth economists found its error rates very low. Even so, a Pangram result is still a statistical estimate about how text reads. It isn't proof of who wrote it.

How does the Pangram AI detector work?

According to Pangram's FAQ, the detector is a classifier trained on a large collection of human and AI writing. It looks at writing style, word choice, syntax and sentence structure. A few details the company states about itself, as of October 2026:

• It says it does not use perplexity or similar measures that correlate with AI writing.
• It says it does not compare your text to a database or to the web, so it isn't a plagiarism checker.
• It returns one of three labels: human, AI or AI-assisted.
• Its pages list per-model accuracy for 26 AI models, including Claude, GPT and Gemini versions. Those figures are Pangram's own, measured on its own benchmark.

The "pangram" in the name has nothing to do with the word puzzle in the news. It's a company, Pangram Labs, that happens to share the word.

Is Pangram accurate? What independent research says

Pangram's headline claim is "99.98% accuracy," and its FAQ puts the false positive rate at about 1 in 10,000 documents. Those are the vendor's numbers, so read them as claims.

The most-cited outside check is "Artificial Writing and Automated Detection," a September 2025 working paper by Brian Jabarian and Alex Imas (NBER 34223, from the University of Chicago Booth School of Business). They compared three commercial detectors, Pangram, Originality.ai and GPTZero, with one open-source model. According to the paper's abstract:

• Commercial detectors beat the open-source one.
• Pangram's error rates were very low and held up across AI models, very short passages (50 words or fewer) and text run through "humanizer" tools.
• Pangram was the only detector that met a strict policy cap of a false positive rate at or below 0.5% without losing its ability to catch AI text.

Two cautions. First, this is a working paper, not a final peer-reviewed result, and it tests a specific set of text types and tools at one point in time. Detectors and AI models both change quickly. Second, a 0.5% cap is not 1 in 10,000. The vendor figure and the study's measured rates are different measurements, and the study tested particular kinds of writing. Neither number tells you the chance that your essay or email is flagged.

For a wider view of why detectors disagree, see our comparison of GPTZero and ZeroGPT, which draws on a separate independent study.

What does a Pangram result mean?

A result is a statement about the text in front of the detector, not about you. Here is a sensible way to read each label:

Rules of thumb:

1. A score starts a conversation. It's evidence to discuss, not a finding of misconduct. Schools and employers that use detectors well treat them that way.
2. Short and formulaic text is harder. Lab reports, templates, product descriptions and very short answers leave less to go on.
3. Edited drafts blur the line. AI-assisted is a reasonable label for a draft that started as AI and was rewritten by a person.
4. Check the policy, not just the score. What matters is what your course, workplace or publisher allows, and whether you followed it.

Is Pangram free, and what does it cost?

As of October 2026, Pangram's FAQ says the free plan allows twenty checks a day after you sign up. Its pricing page lists an Individual plan at $20 a month with up to 300,000 words scanned per month, and higher-volume Professional and Team plans. Check the page before you buy, because detector plans change often.

What if Pangram flags your own writing?

Even a very low false positive rate means some human writing gets flagged. If yours is, here is what helps:

1. Don't delete anything. Keep every file and its history.
2. Ask for the report and the policy. Find out which tool, which label and which rule applies.
3. Show your process. Version history in Google Docs or Word, outlines, notes, sources and earlier drafts are the best evidence.
4. Explain any AI help honestly. If your course allows AI for brainstorming or grammar, say so. A short note works, for example: "I used an AI assistant to suggest an outline and fix grammar in my final draft. The arguments, sources and wording are mine."
5. Use the formal process. Schools and employers have appeal procedures. This post isn't legal advice, so for a serious case contact a student advocate or HR.

Our guide to finding out which AI detector your college uses covers what to ask your instructor and what a score can and can't prove.

Checking your own draft in Unflag

If you want a second read of your own writing before you send it, Unflag has an AI Detector on the web and in the iPhone and Android apps:

1. Open the AI Detector and sign in. The web tool needs at least 50 words. The free plan checks up to 600 characters per input, and Unflag Plus raises that to 10,000 characters.
2. Paste your text and press Analyze.
3. Read the AI score, the AI-generated vs human-generated split, and the verdict: "Likely AI-generated" above 50%, otherwise "Likely Human-written."
4. If a passage you wrote yourself reads as AI, rework it in your own words. The Rewrite tool has options like Improve Writing and Simplify, and you can then edit the result by hand to sound like you.
5. If the text states facts or figures, run Fact Check (Plus plan) to see each claim with a verdict and sources.

Unflag's detector is its own model. It can't tell you what Pangram, Turnitin or any other tool will say about the same text, and no tool can promise a particular result from another detector. That matters especially for humanizers: The Booth paper reports that Pangram held up against humanizer tools. So the honest use of any rewriting tool, Unflag included, is to make a draft clearer and more like your own voice, not to guarantee a score. As of October 2026, Unflag Plus is $8.99 a month or $45.99 a year on the web.

Honest caveats

Pangram's numbers come mostly from Pangram, and the strongest outside evidence is a single working paper from September 2025. Both could age quickly as AI models and detectors change. Treat any detector result, ours included, as one signal next to drafts, sources and a conversation with a real person.

How we put this together

We read Pangram's own site, FAQ and pricing page and the Jabarian and Imas working paper abstract on the NBER site in October 2026. We did not run Pangram on any text, and we are not reporting test scores of our own.

The short version: Pangram is one of the stronger detectors on the evidence so far, its label is a signal and not a verdict, and the best protection against a wrong flag is the trail that shows how you wrote the piece.

Frequently asked questions

How does the Pangram AI detector work?

Pangram says it uses a classifier trained on a large set of human and AI writing to spot patterns in style, word choice and sentence structure. It says it does not use perplexity and does not compare your text against a database or the web. It returns one of three labels: human, AI or AI-assisted.

Is the Pangram AI detector accurate?

Pangram claims over 99% accuracy and a false positive rate of about 1 in 10,000. A 2025 working paper by Booth economists Brian Jabarian and Alex Imas found it had very low error rates and was the only detector tested to meet a strict 0.5% false positive cap. Treat vendor numbers as claims, and a score as one signal.

Is the Pangram AI detector free?

Pangram's FAQ says the free plan includes twenty checks a day after you sign up. Paid plans add more volume and features. As of October 2026, the Individual plan is listed at $20 a month on Pangram's pricing page. Check that page before paying, because plans change often.

Can Pangram give a false positive on my own writing?

Yes. No detector is perfect, and even a very low rate means some human writing gets flagged, especially short, formulaic or heavily edited text. If it happens, keep your drafts and version history, ask to see the report and explain how you wrote the piece.

What does the AI-assisted label mean on Pangram?

Pangram reports three outcomes: human, AI, or AI-assisted. AI-assisted points to text that appears to mix human writing with AI help, such as an AI draft that a person edited. It describes how the text reads, and it can't show who wrote what or how much help was allowed.

Can I check my own writing with something other than Pangram?

Yes. Unflag's AI Detector gives an overall AI score, an AI-generated vs human-generated split and a Likely AI-generated or Likely Human-written verdict for your own draft. It is a different model, so it can't predict what Pangram or any other detector will say about the same text.