Pangram review (2026): an honest verdict on the AI content detector — accuracy, false positives, image detection, pricing, limits, and who it's for.
Pangram is one of the more credible AI detectors available in 2026: an advertised 99%+ accuracy, a roughly 1-in-10,000 false-positive rate, segment-level highlighting, "humanizer" detection, and a new image detector, all reachable via web, a feed-labeling browser extension, Docs, an API, and LMS integrations. It's genuinely good at reading whether content looks AI-made — but it only reads content; it generates and publishes nothing, and even the best detector is an estimate, not a verdict.
Pangram is an AI content detector built by Max Spero and Bradley Emi, two Stanford AI graduates, and based in New York. It answers a single question — does this text or image look AI-generated? — and it answers it better than most. It's the detector Substack chose to power its reader-facing "Scan for AI text" feature, and in July 2026 it raised a $9 million round led by Menlo Ventures to keep scaling. I reviewed it as what it is: a detection tool, judged on how well it detects.
The short version is that it's strong. Pangram advertises accuracy above 99% and a false-positive rate around 1 in 10,000 human documents, numbers that have been examined by outside researchers (it cites work from the University of Chicago and the University of Maryland and a top placement on the COLING 2025 benchmark). It doesn't lean on watermarks or metadata; instead it uses a "synthetic mirror" method — generating a matched AI version of your text and analyzing the stylistic gap — and it catches text run through "humanizer" paraphrasers, which trips up weaker detectors. An image detector shipped as a research preview around July 29, 2026.
The honest caveats are about certainty and scope. No detector is infallible: a 1-in-10,000 false-positive rate is excellent but not zero, and distinctive human writing occasionally reads as AI, so a single score is an estimate, not proof — a real consideration when detection is used to make consequential judgments about someone's work. And Pangram's job stops at the score. It doesn't write, rewrite, design, or publish; it tells you how a finished piece reads.
This review scores Pangram as a detector. Where it's a strong fit and where its limits bite are both below, along with an honest note on where Kompozy fits — which is not as a detection rival but as the content engine that produces the work a tool like Pangram then checks.
Pangram is an AI-detection platform. Its text detector returns a probability that a passage is AI-generated or AI-assisted, with a segment-level view that highlights the AI-looking parts, and it flags text laundered through AI "humanizer" paraphrasers. The models are trained on tens of millions of known-human documents and, per the company, avoid metadata and watermarks in favor of a "synthetic mirror" approach that compares your text against a matched AI-written version. An image detector launched as a research preview around July 29, 2026, using pixel-level statistics to flag AI images across different generators, including AI elements composited into real photos. You reach it through a web app, a Chrome extension that labels AI-looking posts in your feed on X, LinkedIn, Substack, Reddit, and Medium, a Google Docs add-on, an API, and LMS integrations for education. It has a free tier and paid individual, professional, team, and enterprise plans. What it is not is a content or publishing tool — it produces no writing, video, images, or scheduling. It reads content and scores it.
Pangram fits anyone whose deliverable is a judgment about content: educators checking submissions, publishers and platforms verifying authorship (Substack being the marquee example), recruiters screening applications, and individual writers who want to self-scan a draft before publishing. It's also a fit for developers who need detection via API. It's a weak fit for creators who arrived hoping a detector would help them make content — it won't; once you have a score, the work of producing captioned video, carousels, posts, a blog, and a schedule is entirely ahead of you, and Pangram does none of it.
| Dimension | Score | Why |
|---|---|---|
| Text detection accuracy | 4.6 / 5 | Advertised 99%+ with outside-researcher scrutiny and a top COLING 2025 placement — strong for a hard problem, though still an estimate on any single input. |
| False-positive control | 4.4 / 5 | A claimed ~1-in-10,000 false-positive rate is excellent, but not zero — distinctive human writing can occasionally be flagged. |
| Humanizer / paraphraser detection | 4.4 / 5 | Catches text run through "humanizer" tools that many older detectors miss, thanks to the style-based method. |
| Image detection | 3.6 / 5 | A research preview using pixel-level statistics across generators — promising and provider-agnostic, but new and evolving. |
| Transparency of method & benchmarks | 4.2 / 5 | Explains the "synthetic mirror" approach and cites external benchmarks, though internal training details stay proprietary. |
| Access & integrations | 4.5 / 5 | Web app, feed-labeling browser extension, Google Docs add-on, API, and LMS integrations cover most real workflows. |
| Pricing & value | 4.2 / 5 | A usable free tier and a low ~$20/mo individual plan make it easy to adopt; heavy or API use scales the cost. |
| Reliability as a sole judgment | 3.5 / 5 | A score is an estimate, not proof — using it alone for high-stakes decisions about a person's work is the tool's main risk. |
Pangram's pricing is straightforward and, for what it is, fair. A free tier (roughly a couple thousand words a day plus a few image scans) lets most people try it with no commitment, and the individual plan sits around $20/month with an annual discount, raising word and image-scan limits and adding features like plagiarism detection and social-feed scanning. Above that, a professional plan (around $65/month) lifts limits further and bundles an API usage credit, with team, education, and enterprise options beyond it. Because detection tools frequently adjust their usage caps, reconcile any specific figure against pangram.com before budgeting.
Judged as a detector, that's reasonable value: you're paying for accuracy and access, and the free and low tiers cover casual use well while API and education licensing serve institutions. The place to be careful is not the price but the interpretation — a score you pay for is still an estimate, and the cost of over-trusting it (a wrongly flagged student, applicant, or writer) isn't on the invoice.
The comparison people sometimes try to draw — Pangram versus a content engine like Kompozy — is a category error. Pangram prices detection volume; Kompozy prices content generated and published. One tells you how a piece reads; the other makes the piece. If your need is verification, Pangram's pricing is the relevant one. If your need is producing and shipping content, a detector's price sheet is answering the wrong question.
| Use case | Fit | Why |
|---|---|---|
| Checking whether a text passage was AI-written | Strong | This is the core job, and Pangram does it about as well as anything on the market. |
| Educators and platforms verifying authorship | Strong | Segment highlighting, LMS integrations, and a low false-positive rate suit verification workflows. |
| Self-scanning your own draft before publishing | Strong | The pre-publish check and Google Docs add-on make it easy to see what a reader would see. |
| Catching "humanized" / paraphrased AI text | Strong | The style-based method flags laundered AI writing that fools simpler detectors. |
| Detecting AI-generated images | OK | The image detector is capable and provider-agnostic, but it's a research preview, so treat results as evolving. |
| Making finished, on-brand content | Weak | Pangram generates nothing — it only scores content that already exists. |
| Publishing content across platforms | Weak | There is no scheduler or publisher; detection stops at the score. |
Kompozy is not an AI detector, and it would be dishonest to pitch it as a Pangram competitor — they sit at opposite ends of the workflow. Pangram reads a finished piece and tells you how AI it looks. Kompozy makes the piece: it's a content generation and publishing engine that turns a source into Blog Articles, Email Newsletters, Text Posts, Carousels, Quote Graphics, Photo Posts, and Persona/HeyGen avatar video, all governed by a Persona Brief and a banned-word filter, then reframes and publishes each across nine platforms from one review pipeline.
The reason the two belong in the same conversation is what Pangram's rise signals. Detection is becoming a standard reader-facing feature — Substack ships it, and more platforms will — and the losing response is to feed AI text into a "humanizer" and hope the score comes back green. That's an arms race with no end. The winning response, and the one Kompozy is built around, is to generate content in a genuine, disclosed brand voice, edit it into your own words, and diversify into formats a text detector can't even assess. Use Pangram as an honest proofing gate on your writing; use Kompozy to produce the writing — and the video, images, and carousels — that make a detector's verdict a footnote rather than a threat.
It's among the more accurate detectors available: it advertises 99%+ accuracy and a roughly 1-in-10,000 false-positive rate, with results examined by outside researchers including the University of Chicago and the University of Maryland and a top placement on the COLING 2025 benchmark. Those are strong numbers, but no detector is perfect — treat any single score as an estimate, and expect occasional false flags on distinctive human writing.
For detection, yes. If your job is verifying whether content is AI-generated — as an educator, publisher, recruiter, or a writer self-checking a draft — Pangram is accurate, well-integrated, and has a usable free tier plus a low ~$20/mo plan. It's not worth buying if what you actually need is to make content, because it generates and publishes nothing.
Per its co-founder, it avoids metadata and hidden watermarks. It uses a "synthetic mirror" method: it generates a same-topic, same-length, same-tone AI version of your text with a frontier model and analyzes the stylistic differences, drawing on training over tens of millions of known-human documents. That style-based approach is also why it can flag text run through AI "humanizer" paraphrasers.
Yes, via an image detector that shipped as a research preview around July 29, 2026. Instead of relying on a single provider's watermark, it analyzes pixel-level statistical patterns, which lets it flag AI images across generators and even detect AI elements composited into a real photo. As a preview, its image results should be treated as evolving.
For detection: Originality.ai for publishers and content teams, GPTZero for education, Copyleaks for enterprise, and Turnitin for academic integrity. If your real goal isn't detecting content but producing it, the tool you want is a content engine like Kompozy, which generates and publishes on-brand posts, video, carousels, and newsletters across platforms.
Yes. Even at a ~1-in-10,000 false-positive rate, genuine human writing — especially a distinctive or formulaic voice — can occasionally be flagged. That's why a score should be treated as a signal, not a verdict, and why platforms like Substack that use Pangram describe the result as an estimate and let writers report mistakes.
That's the natural pairing. Generate and edit a draft in Kompozy — where a Persona Brief and banned-word filter cut generic AI-tell at the source — then scan the finished text in Pangram, rewrite any high-scoring segments, disclose your process, and publish. Kompozy makes the content; Pangram checks it.
Educators and schools (via LMS integrations), publishers and platforms verifying authorship (Substack powers its AI scan with it), recruiters screening applications, developers integrating detection via API, and individual writers self-checking drafts. It fits poorly for anyone whose real need is creating and distributing finished content.