Pangram AI Detector Review 2026: The Detector Substack Chose — and Why It Beat Turnitin in a 2026 Study

Pangram is the detector Substack chose in July 2026, when CEO Chris Best gave the problem it solves a name: "Claudefishing." Independent studies, not us, put it ahead of tools schools actually license: a Chicago Booth working paper reports a near-zero false-positive rate and robustness to StealthGPT, and a 2026 Springer article in the International Journal for Educational Integrity reports Pangram beating Turnitin, GPTZero, and Copyleaks. We have not tested Pangram ourselves, and our own detector has no Pangram label. Most universities still run Turnitin, not Pangram. If you are a student, that last sentence is the one that decides what you should check.

Research-standing, not a hands-on score

On published evidence: strong. We have not hands-on tested it.

That is the rating. We build SupWriter, which includes an AI detector. That detector produces one score from a single scoring model, then maps it onto twelve named labels (Turnitin, GPTZero and others) as a guide; it does not query those vendors, and Pangram is not among the labels. We have not run Pangram. We will not invent a Pangram-versus-SupWriter bypass table, and SupWriter has not published live-detector bypass tests for any detector (how we test).

Same discipline as our PaperBleach review: papers and journalism in one column, our own testing in another. That column is empty on purpose. If your assignment goes through an institutional license, what matters is the detector your university actually runs. A Substack headline is not a Canvas integration.

What Pangram is — and why Substack bought in

Pangram is a commercial AI-text detector. It sells the same job GPTZero and Originality.ai sell to publishers and platforms: given a passage, how likely is it that a language model wrote it? The difference, in the second half of 2026, is who picked it when the question stopped being "did a student write this essay?" and became "did a person write this post?"

On 21 July 2026, The Verge reported that Substack had added an AI detector powered by Pangram. Readers can scan a post longer than 100 words by choosing "Scan for AI text" from the post's menu. Chris Best, Substack's CEO, framed the problem this way: "The core problem is not people using AI, or the quality of its output. The problem is when there is a mismatch between a reader's expectation and reality, especially when they unwittingly invest their attention in something with no human thought on the other end. That's Claudefishing." Substack needed a classifier that could sit on a consumer publishing platform, not an LMS.

That choice is why Pangram jumped from a specialist name to a mainstream one. It is also easy to over-read. Substack is not a university. Notes is not Turnitin. A platform picking a vendor to answer "is there a person behind this?" is evidence about that product problem. It is not evidence that your faculty dashboard switched engines.

The Claudefishing idea is still useful if you never open Substack. Readers notice polished prose with nobody behind it, and the same rhetorical kit shows up in AI words to avoid in 2026: cadence, hedging, and the polished-nothing paragraph, not just a leftover "delve." The studies below are the part that is not a press cycle.

Pangram 4, $9 million, and the version-3 sunset

Eight days after the Substack news, the company raised money and shipped a numbered model.

On 29 July 2026, TechCrunch reported a $9 million raise, led by Menlo Ventures, aimed at detecting AI content as it flooded the open internet. The same window is when Pangram published Introducing Pangram 4. Version 3 is scheduled to deprecate on 30 September 2026. If you are integrating an older Pangram endpoint, that date is the operational fact. After 30 September, assume production traffic is Pangram 4 unless the vendor has told you otherwise.

We stop there on product claims. Pangram's launch post makes its own accuracy claims for Pangram 4; we have not verified them, so we are not reprinting them as findings. We will not invent a per-scan price or a language list either. If those numbers matter to a purchase, read pangram.com on the day you buy.

What the July cluster establishes without extra numbers: Pangram is capitalised, it is iterating on a numbered model, and a major consumer publisher already picked it. That is a different market position from a student-facing paste box — and from Turnitin, which still sits inside Canvas, Moodle, and Blackboard for most of the people reading this.

The Chicago Booth paper (Jabarian and Imas)

The research result that keeps getting attached to Pangram is Jabarian and Imas, a Chicago Booth / BFI working paper (August 2025). It compared four detectors (Pangram, Originality.ai, GPTZero and an open-source RoBERTa model); Turnitin was not part of it. Attribute what follows to that paper, not to SupWriter. We did not reproduce it.

Near-zero false-positive rate. In their evaluation, Pangram almost never called human text "AI." That is the number educators care about first, because a false flag is an accusation. Turnitin markets the same priority. A near-zero FPR, if it holds outside the paper's sample, is the feature that makes a detector usable in a hearing.
Robust to StealthGPT. The authors ran every AI passage through StealthGPT's default rewrite and re-scored it; Pangram's performance was "largely robust to the humanizer." That is the claim competitors most want to argue with, and the one we are least willing to launder into a line of our own.
GPTZero lost much of its detection on humanized text; Originality.ai held up better. Same paper, same condition. After the StealthGPT pass, GPTZero's miss rate was around 50% and above across most genres and models, while Originality.ai's rose to about 5% on longer text and up to 21% on shorter text. That is a finding about those tools in that setup. It is not our Originality.ai review, and it is not a SupWriter bypass percentage.

We disclose the conflict: we sell a humanizer. A paper that finds a widely used detector missing around half of StealthGPT-processed text is a paper about detector fragility, not a coupon. It also does not say "every humanizer beats every detector." StealthGPT is one product. Pangram is one detector. The accuracy landscape is still messy in both directions — missed AI and false positives on human writers, especially ESL students.

What we will not do with Booth: invent a "Pangram vs Turnitin" percentage (Turnitin was not in the study), or pretend we reran StealthGPT last week. If you need the tables, open the PDF.

The Springer 2026 paper

The academic-integrity citation is a 2026 article in Springer's International Journal for Educational Integrity, "Who wrote this? Evaluating the reliability of AI detection tools in higher education" (doi:10.1007/s40979-026-00226-w). It compared four detectors (Turnitin, Pangram, Copyleaks and GPTZero) on 160 synthetic English academic papers of at least 4,000 words each: 40 fully human, 40 fully AI-generated, 40 hybrid and 40 humanized. It also analyzed 1,163 real master's theses. We are reporting only findings stated in the article.

The Springer International Journal for Educational Integrity article reports:

Pangram beat Turnitin, GPTZero, and Copyleaks on the evaluation they ran. On the fully AI-generated papers, "Pangram was the only tool that performed well," and on hybrid and humanized papers it identified AI content in the expected range in 37 of 40 cases.
Turnitin classified 100% of the fully AI-generated papers as false negatives, with scores between 0 and 20%. For context, Turnitin displays any score from 1% to 19% only as an asterisk (Turnitin FAQ).

Those two sentences are already enough to change how you read a Turnitin score, and already enough to over-claim if you stop being careful. A 100% false-negative result on fully AI-generated papers is a finding about that study's materials and that Turnitin configuration. It is not a promise that your 2,000-word essay will sail through. Turnitin's own product has moved in 2026 — a single combined score as of 4 August, Arabic AI detection from 18 August, bypasser detection that actually launched in August 2025. A paper's Turnitin column is a snapshot.

We are not adding extra percentages from secondary blogs. If a ranking page tells you "Pangram is 99.x% accurate" without quoting this article's table, treat it as marketing until you can open the PDF.

The honest synthesis of Booth plus Springer, still not our testing: on the published record, Pangram is the detector that looks strongest and most resistant to at least one named humanizer, and Turnitin — the tool your school probably runs — can miss fully generated academic papers in a published test. That pair of facts is why Pangram is getting cited. It is also why a student who only "beats Pangram" in a browser tab may still be looking at the wrong exam.

What we have not done

Pangram is not one of our detector's labels. SupWriter's detector produces one score and maps it onto twelve named labels (Turnitin, GPTZero, OpenAI, Claude, Gemini, Grok, Grammarly, Crossplag, Copyleaks, QuillBot, Sapling and ZeroGPT) as a guide. It does not query any of those vendors, and there is no Pangram label. How the score is produced.
We have not run Pangram. No sample set, no table. There is no SupWriter-versus-Pangram number in this company. If you see one attached to our name, it was not us.
We have not measured our humanizer against Pangram, or published live tests against any other detector. Our methodology page says so plainly.

This is the same discipline we used when a model launched last week and everyone wanted a percentage. GPT-6 Astra has no honest bypass table yet. Pangram has honest other people's tables. We will cite those. We will not impersonate them.

Why students should still care about Turnitin more than Pangram

Most universities did not rip out Turnitin in July because Substack made a vendor choice. The institutional stack is inertia plus a plagiarism license that already includes an AI indicator. Instructors see that indicator. Students usually do not. Policy, when it exists, is written in Turnitin percentages — or, more often, in the absence of a published number, which we covered in what percentage of AI is acceptable on Turnitin.

The risk model for a marked assignment is still: does this course run Turnitin (or Copyleaks, or a GPTZero teacher account)? What does that tool say about the file you will submit? What does process evidence say — Docs history, a viva, a declaration rule?

Pangram can be "better" in a paper and still be irrelevant to the first question. A Springer result in which Turnitin missed fully AI-generated academic papers is a reason for faculty to be less smug about a 0% AI score. It is not a reason for a student to ignore the detector their LMS actually calls.

What we can offer is a quick pre-check, not a verdict. SupWriter's AI detector gives one score with sentence highlights, based on a single scoring model; it is not Turnitin, and it is not Pangram. Free evaluation: 300 words in total, one-time, with a free account and no card, on the free humanizer. Paid list prices: Basic $9.99/month for 5,000 words, Pro $19.99 for 15,000, Ultra $39.99 for 30,000.

Institutional-detector deep dives: GPTZero vs Turnitin, does Turnitin detect AI writing, Originality.ai review. Those are the products a hearing will name.

Who should use Pangram versus our detector

Use Pangram if you are a publisher, platform, or researcher trying to match the stack Substack chose, or you want to read the Booth and Springer evaluations against the same vendor those papers used. If you are buying detection as a product — API, moderation queue, editorial workflow — Pangram is now a shortlist name on the public record. We cannot tell you what it will cost, because we will not invent a price.

Use a school-facing detector if you are submitting coursework. Check what your course actually runs, not last month's tech news. Our AI detector can give you a quick pre-check, but its Turnitin and GPTZero labels are a guide derived from one score, not results from those tools. It will not return a Pangram score either. That is a limitation we are stating, not hiding.

Use a humanizer only where policy allows a draft, and only as a voice pass. A humanizer does not replace a declaration (Malaysia is the blunt version of that sentence). It does not erase version history. It does not "beat Pangram" in any number we will quote. If you are comparing humanizers, see best AI humanizers 2026 and best AI humanizer for students. For Undetectable AI specifically: review, alternatives, head-to-head.

Do not treat this review as academic-misconduct insurance. If your school ever licensed Pangram, the Booth robustness finding is not a comfort. Until they do, your problem is still the blue percentage in the LMS.

Frequently Asked Questions

Have you tested Pangram yourselves?

No. It is not one of our detector's labels, and we have not run it. This review is research-standing: Substack / The Verge (21 July 2026), the $9 million raise and Pangram 4 (29 July 2026), Pangram 3's 30 September 2026 sunset, the Chicago Booth working paper, and the Springer International Journal for Educational Integrity article. No samples of our own.

Is Pangram more accurate than Turnitin?

On the published academic record we can cite, the Springer article reports Pangram beating Turnitin, GPTZero, and Copyleaks, and it reports Turnitin classifying 100% of fully AI-generated papers as false negatives. The Booth paper, which did not include Turnitin, reports a near-zero false-positive rate for Pangram and robustness to StealthGPT. We have not reproduced either result. Most campuses still score you with Turnitin.

Can SupWriter bypass Pangram?

We do not have that number, and we will not invent one. SupWriter has not published live-detector bypass rates for Pangram or any other detector. If Pangram is what your reader uses, the only honest test is your own text on Pangram.

Why did Substack choose Pangram?

The Verge's 21 July 2026 piece is the source: Chris Best wanted a response to what he calls "Claudefishing," a mismatch between a reader's expectation and reality, when readers invest attention in something with no human thought on the other end. That is a platform moderation decision, not a university procurement story.

Should I check my essay with Pangram before I submit?

Only if Pangram is what your marker uses. If the course runs Turnitin or Copyleaks, a Pangram pass is the wrong rehearsal. Our AI detector can give you a quick pre-check, but it is not Turnitin or Copyleaks either.

Is Pangram 3 still usable?

Pangram 3 is scheduled to deprecate on 30 September 2026. After that date, treat Pangram 4 as the current model unless the vendor says otherwise. We have not benchmarked either version.

Pangram looks strong in other people's papers. Your school still probably runs Turnitin. For a quick pre-check before you submit, try our AI detector (a single-model guide, not Turnitin), and if you are editing voice on a policy-allowed passage, 300 words free on the AI humanizer, no card.