Pangram AI Detector Review 2026: The Tool Substack Uses

Pangram spent two years as a specialist AI detection startup that most writers had never heard of. Then, on July 21, 2026, Substack wired it directly into the reading experience, letting anyone scan a post to see an estimate of how much of it a human actually wrote. A week later, Pangram closed a $9 million round led by Menlo Ventures. Overnight, the Pangram AI detector went from an industry tool to the score millions of newsletter readers can pull up next to your byline.

Here is the short verdict: Pangram's classifier is one of the strongest in the consumer market right now, with accuracy claims that have held up better under outside scrutiny than most of its rivals. It is also not magic. Its hardest case is exactly the kind of writing most professionals produce in 2026, a human draft with AI edits woven in, and its verdicts now arrive in public. If you publish anywhere Pangram scans, you want to know your score before your readers do.

Key takeaways

• Pangram is the AI detector behind Substack's reader-facing scans, which cover posts and notes published on or after July 21, 2026.

• The company claims 99.9%+ accuracy and a false positive rate around 1 in 10,000; outside evaluations have been favorable, but mixed human-AI writing remains its weakest area.

• Pricing runs from a free tier (about 2,000 words per day) to $20 per month for individuals and $65 per month for professionals, plus a pay-as-you-go API.

• Writers can self-check drafts before publishing with a free detector such as WriteHuman's AI detector, then revise anything that scores high before an audience sees a public verdict.

What is Pangram?

Pangram is an AI text detection company founded in New York in 2023 by Max Spero and Bradley Emi, two Stanford-trained machine learning engineers whose résumés run through Google and Tesla's Autopilot team. In late July 2026 the company raised $9 million led by Menlo Ventures, bringing its total funding to roughly $13 million.

The technical approach is what separates it from the last generation of detectors. Instead of leaning on watermarks or metadata, Pangram trained its models on tens of millions of human-written documents, then had frontier language models produce what the company calls synthetic mirrors: AI versions of each document matched for topic, length, and tone. The classifier learns the stylistic tells that models produce consistently, and the current text model, Pangram 4, claims over 99% accuracy at spotting not just fully generated text but AI-assisted and mixed human-AI writing.

Output is more granular than a single percentage. Pangram returns an overall AI likelihood, a segment-by-segment breakdown showing which passages look generated, and a distinction between full generation and lighter AI assistance. Beyond Substack, its customers include Quora, NewsGuard, Wiki Education, publishers, and recruiters screening application materials, and a Chrome extension labels posts across X, LinkedIn, Reddit, Medium, and Substack. An image detection model, Pangram Image, launched in research preview in July 2026.

Does Substack use Pangram?

Yes, and the integration is the reason searches for the Pangram AI detector spiked this summer. Substack announced the partnership on July 22, 2026, and the mechanics matter if you publish there:

• Readers can scan almost anything. Posts, notes, comments, and replies published on or after July 21, 2026 are eligible, on web and iOS first with Android to follow. Very short items are excluded; the scan needs a real chunk of text to work with (Substack's help documentation puts the floor at around 100 words).

• The result is a percentage estimate. A scan shows how much of the text Pangram believes was human-written versus AI-generated or AI-assisted.

• Writers get some control. Creators can run Pangram on their own drafts before publishing, report scans they believe are wrong, disable detection on their posts, and attach an optional "How I make this" note disclosing how AI figures into their process.

Substack has framed this as transparency rather than punishment. There is no penalty for a high AI score, and CEO Chris Best has been careful to describe disclosed, thoughtful AI use as legitimate. But the practical effect is obvious: for the first time on a major publishing platform, a detector verdict sits one tap away from every paid subscription decision. We cover what that shift means for newsletter writers specifically in our companion piece on Substack's AI detection rollout.

How accurate is Pangram?

Pangram's own numbers are aggressive: 99.9%+ accuracy across major models (ChatGPT, Claude, Gemini, Grok, Llama, DeepSeek) and a false positive rate of roughly 1 in 10,000 on aggregated public datasets. Unusually for this industry, some of those claims have outside support. A University of Chicago evaluation measured Pangram's false positive rate on product reviews at 0.5%, the lowest among the detectors tested, and University of Maryland researchers have also validated its performance. TechCrunch's own spot-testing found the detector reliably caught fully AI-generated articles.

The same testing surfaced the weak point: partially edited content. When TechCrunch fed it human writing that had been reworked with AI in the mix, Pangram occasionally tagged genuinely human sentences as generated. That matters because blended drafting is now the normal way professionals write, and it is precisely the gray zone where a percentage score is most likely to mislead.

Two more caveats belong in any honest review. First, scale changes what small error rates mean. One false positive in 10,000 sounds negligible until a platform runs millions of scans, at which point real writers get wrongly labeled every week, and critics raised exactly this concern when Substack switched the feature on. Research on earlier detectors found they flagged neurodivergent writers more often than neurotypical ones (that study did not test Pangram specifically), and non-native English professionals have long reported similar problems across the category. The Atlantic's Matteo Wong warned that "AI accusations could very quickly spiral into a witch hunt." Second, the detection industry has a history of publishing error rates that later proved optimistic, so treat every vendor's headline number, Pangram's included, as a claim rather than a law of physics.

Is Pangram free? Pricing breakdown

Pangram has a genuinely usable free tier, which is more than most rivals offer. Published pricing as of August 2026:

For a freelancer or newsletter writer who just wants to check drafts occasionally, the free 2,000 words a day covers most use. Agencies and publishers running volume will land on the $20 or $65 tiers. Note that if you only care about Substack, you do not need any plan at all: the reader-facing scan there is free by design.

Where Pangram falls short

The honest limits, drawn from public reporting rather than our own lab claims:

• Mixed writing is the hard case. Fully generated text gets caught; human drafts with AI edits produce the misfires, in both directions.

• It cannot see intent. Pangram detects whether AI generated the words, not whether AI did the research, outlining, or thinking behind them. A deeply reported piece drafted with an assistant and a lazy paste job can score identically.

• The baseline is aging. Reporting on the Substack rollout noted Pangram's human baseline leans on pre-2021 writing. Human style keeps drifting toward AI phrasing as people read more of it, which makes the boundary blurrier every year.

• Verdicts are public before they are appealable. On Substack, a reader sees your score the moment they scan. The report-a-bad-scan flow exists, but the first impression has already happened.

What writers should do before publishing

The rational response to public detection is a private pre-flight check. Before a post goes out, paste it into a detector yourself and see what a classifier sees. WriteHuman's AI detector is free and gives you a score plus a read on how AI-like the text feels, so a surprise verdict happens on your screen instead of under your byline.

If sections come back with high AI probability and they read stiff to you too, that is a revision signal, not a verdict on your integrity. Rework those passages by hand, or run them through the AI humanizer on WriteHuman's homepage, which rewrites machine-drafted phrasing toward the varied rhythm and word choice of natural human writing. The free plan covers 3 humanizations of 250 words each per month, and paid plans start at $20 per month. To be clear about what any of these tools can and cannot do: no humanizer guarantees a specific score from Pangram or anyone else, and detectors frequently disagree with each other. What a self-check gives you is information before publication instead of after. And since Pangram is pushing into image detection too, the same logic applies to your artwork: WriteHuman's free AI image detector will tell you how a generated-looking header image reads before your audience weighs in.

Frequently asked questions

What is Pangram?

Pangram is an AI text detection company founded in 2023 by ex-Google and ex-Tesla machine learning engineers Max Spero and Bradley Emi. Its classifier estimates how likely a piece of text is to be AI-generated or AI-assisted, with segment-level analysis. It powers Substack's reader-facing AI scans and is also used by Quora, NewsGuard, publishers, and recruiters.

How accurate is Pangram?

Pangram claims 99.9%+ accuracy and a false positive rate of about 1 in 10,000, and a University of Chicago evaluation found it had the lowest false positive rate of the detectors tested. Independent spot-testing shows it reliably catches fully AI-generated text but sometimes mislabels human sentences inside partially edited, mixed human-AI drafts. Treat any single score as an estimate, not proof.

Does Substack use Pangram?

Yes. Since July 21, 2026, Substack readers can scan posts, notes, comments, and replies with Pangram to see an estimated split between human-written and AI-generated text. Creators can pre-scan drafts, report scans they believe are wrong, disable detection on their posts, and add an optional note disclosing how they use AI.

Is Pangram free?

Partly. A free tier covers about 2,000 words of scanning per day plus 3 image scans, which is enough for occasional self-checks. Paid plans start at $20 per month for 300,000 words, with a $65 professional tier and a pay-as-you-go API above that. Scanning posts on Substack itself costs readers nothing.

The bottom line

Pangram is a technically serious detector with better outside validation than the category norm, and the Substack deal made it the first AI detector most readers will ever actually use. But its scores are estimates with known blind spots, especially on the blended human-AI drafts that define modern professional writing. The writers who handle this well are the ones who check their own work before publishing, instead of finding out from a reader's screenshot.