Pangram AI
Pangram is a research-grade AI detector for text and images, built around the lowest independently verified false-positive rate in the category. Strong on raw AI writing, weaker on edited text and strippable image labels.
The best-documented false-positive rate in AI detection, let down by an image check you can strip in under a minute and a free tier that barely lets you test it.
Should you use Pangram?
Integrity offices, publishers and editors who have to defend a flag to a student or author and want the lowest independently verified false-positive rate.
You mostly check paraphrased or humanized text , or you want a generous free plan. Accuracy slips on reworked writing and free is capped per day.
Free scanning is a daily allowance. Real use starts at Individual $20/mo billed yearly , which covers 300,000 words a month.
Image detection often reads a hidden file label rather than the picture. Strip that label through any editor and an AI image passes as human.
Key features & how they perform
Each feature rated from hands-on testing, independent benchmarks and aggregated review sentiment.
Model-level AI detection
Flags writing from ChatGPT, Claude, Gemini, Grok and Llama, with a vendor-reported 99.7% across 26 benchmarked models. Strongest on raw, unedited output.
Low false-positive design
A roughly 1 in 10,000 false-flag rate, checked by University of Chicago Booth and University of Maryland researchers. The number that matters in integrity cases.
Segment & AI-assist analysis
Marks which sentences read as AI and which look AI-edited rather than returning one blunt score, so a reviewer can see where the suspicion sits.
AI image detection
Reads embedded content credentials first and runs its own model only when none exist. Solid on signed files, beatable once the label is gone.
Integrations & workflow
Google Docs, Gmail, a Chrome extension, an MCP server and an API, plus LMS hooks for Canvas, Moodle, Google Classroom and Brightspace.
Multilingual + similarity
Covers 20+ languages and adds a one-click plagiarism check on paid plans, though confidence drops on short or non-English passages.
Feature ratings blended from hands-on testing, independent research (University of Chicago Booth, University of Maryland) & aggregated review sentiment.
Six checks through Pangram's detector, text and image
From a plain block of AI writing to a C2PA-signed picture, here is what the no-cost plan actually caught.
Raw AI text, caught in seconds
Pangram's core job is the plain paste-and-scan check. You drop text into the box, it reads the patterns in the writing, and it returns a score plus sentence-level highlights.
For a clean baseline I wanted unedited AI with nothing hidden. I asked ChatGPT for a short blog intro and pasted the answer in with no changes.
"Write a 180-word blog intro about why morning walks help busy professionals focus."
The AI text I pasted (left) and Pangram's 100% AI result (right).
A 2016 article stayed human
A detector is only trustworthy if it leaves real human writing alone. The number that protects an innocent writer is the false-positive rate, so I needed text written well before modern AI tools existed.
I pulled a published article from 2016 about academic publishing and ran 165 words of it through the same free check.
A 2016 article scored 100% human. The similarity report stays locked on the free plan.
Spanish worked, but short text made it hedge
Pangram lists support for more than 20 languages, so English alone would not test it. I generated a short Spanish paragraph of AI text and ran that next.
"Escribe un párrafo de 120 palabras sobre los beneficios de caminar por la mañana."
Spanish input (left) and the hedged "Possibly AI" verdict (right).
AI images and content credentials
Image detection runs on a separate, newer model. To start I uploaded an AI image that still carried its content credentials, a hidden record called C2PA that some generators sign into the file.
The result came back fast, but the small print mattered more than the verdict.
The AI image was flagged from its C2PA label, and Pangram's own model did not run.
One trip through Canva broke the image check
If the first image passed on its label alone, the obvious question was what happens when the label disappears. I took the same AI image, opened it in Canva, exported it with no edits, and uploaded the new file.
The content credentials were gone, so this time Pangram had to judge the picture itself.
The upload screen caps the free plan at 3 image scans (left). An AI image routed through Canva passed as "No AI" (right).
What the free plan actually gives you
After six checks the shape of the free plan was clear, and it is tight.
Logged out, I began with 20 text credits at 1 credit per 100 words, which comes to roughly 2,000 words, plus 3 image scans. Text runs on Pangram 4.1 and the image check on Pangram Image 0.2.
Pangram pricing
Figures taken from the official Pangram pricing page. Prices shown monthly, with annual billing saving the amounts noted.
| Plan | Price | What's included |
|---|---|---|
| Free | $0 | 20 text checks and 3 image scans per day · no account needed · Pangram 4.1 text model · similarity report locked |
| Individual Recommended | $20 /mo | 300,000 words/mo · 100 image scans/mo · similarity on every scan · MCP, OCR, Google Docs & Gmail · $10/mo API credits · 7-day free trial · save $60/yr |
| Professional | $65 /mo | Everything in Individual, plus 5x words (~1.5M/mo), 5x image scans (~500/mo) and $200/mo API credits · save $240/yr |
| Team | $20 /seat/mo | $40/mo for 2 seats, add more anytime · everything in Individual, plus admin & member dashboard and unified billing · save $120/yr |
| Organizational | Contact sales | Custom institutional pricing for universities and enterprises, listed under a separate Organizational tab rather than published online |
Pricing is accurate as of 30 October 2026 and was captured from the official Pangram website. Plans, prices and scan limits change often, so check the official Pangram pricing page for the current figures before you subscribe.
Pros & cons
What real users praise and complain about, drawn from public reviews, not from our testing.
✅ Pros
- Catches AI-assisted edits from Grammarly and Copilot, not just fully generated text
- Classification reports are clear and come back in seconds
- Chrome extension and batch file scanning speed up repeat checking
- Support follows up on questions about a disputed flag
⛔ Cons
- Human-written dissertations and research proposals flagged as 100% AI
- Substack writers report their own newsletter posts wrongly flagged
- Re-running the same passage can return a different verdict
- Students required to use it by their school dispute its results
- Some signups and queries went unanswered during busy spells
Synthesized from real user reviews on Trustpilot and Product Hunt, with little presence yet on G2 or the app stores · paraphrased, not quoted
Pangram alternatives
| PangramReviewed | Scribbr AI | Grammarly AI | GPTZero AI | Copyleaks AI | |
|---|---|---|---|---|---|
| Our Score | 7.2/10 | 7.0/10 | 6.4/10 | 6.2/10 | 5.9/10 |
| Entry paid plan | $20/mo (billed yearly) | Detector free; plagiarism from $19.95/doc | $12/mo (Pro) | $10/mo (Essential) | $7.99/mo (AI detector) |
| Free tier | 20 checks a day, no card | Unlimited scans, 1,200 words each | Limited free checks | 10,000 words a month | ~2,000-word trial |
| AI image detection | Yes, separate model | No, text only | No, text only | No, text only | Yes, image and deepfake |
| Plagiarism check | On paid plans | Separate paid product | Included in Pro | On Premium and up | Separate add-on plan |
| Best for | Integrity teams defending a flag | Students wanting a free quick check | Writers already inside Grammarly | Educators on a low budget | Enterprises needing many languages |
| Standout strength | Lowest verified false-positive rate | Unlimited no-signup scans | Built into an everyday writing tool | Permanent free plan for light use | Detection across 30+ languages |
Who Should Use Pangram, and Who Should Avoid It?
Pangram is the most defensible way to flag AI text, but its image blind spot and single-purpose design rule it out for some jobs.
Who Should Use Pangram?
Best suited for these workflows
Integrity Offices and Editors
Anyone who must defend a flag to a student or author gets the lowest independently verified false-positive rate in the category.
Publishers Checking at Scale
Batch uploads, segment analysis and AI-assist detection suit newsrooms and agencies vetting large volumes of submitted content.
Teams Wanting AI and Plagiarism Together
A single scan on a paid plan returns both an AI verdict and a similarity check, so reviewers see both signals in one report.
Developers and Institutions
An API, an MCP server and LMS hooks for Canvas, Moodle, Google Classroom and Brightspace fit detection into systems you already run.
Who Should Avoid Pangram?
Consider an alternative if you need
Proof Rather Than a Signal
Even the strongest detector can misjudge, and real writers report human work flagged as AI, so a score is evidence, never a verdict.
To Catch Heavily Reworked Text
Paraphrased, humanized or lightly edited AI slips through, a limit independent studies find across every tool in this category.
Bulletproof Image Checking
Image detection is a newer, separate model and leans on embedded labels, so a stripped or edited file is far less certain.
One Tool for Writing Too
Pangram only detects. It has no grammar help, paraphrasing or drafting features, so it sits alongside your writing tools, not inside them.
Our Recommendation
Pangram is the pick when a false accusation carries real cost and you can treat its score as evidence rather than proof. For casual spot-checks or for image authenticity that has to survive an edit, it is either overkill or outright beatable. Before you commit, run a few pieces of your own pre-2023 writing through it to gauge false positives, and confirm the current word and image limits for your volume.
Pangram scorecard
Rated against what an AI content detector is actually built to do.
| Dimension | Verdict | Score |
|---|---|---|
| Detection accuracy Catching unedited ChatGPT, Claude, Gemini | Excellent |
9.0
|
| False-positive control How rarely it flags real human writing | Excellent |
8.5
|
| Independent validation Third-party evaluation and open research | Excellent |
8.5
|
| Integrations & workflow Google Docs, Gmail, Chrome, API, LMS | Good |
8.0
|
| Ease of use & reporting Segment analysis and readable results | Good |
8.0
|
| Similarity checking Built-in plagiarism check on paid plans | Good |
7.0
|
| Evasion resistance Holding up on paraphrased or humanized text | Average |
6.5
|
| Multilingual & short text 20+ languages, confidence on brief passages | Average |
6.0
|
| Value & free tier Entry price and free allowance | Average |
5.5
|
| Image detection Spotting AI images once the label is gone | Weak |
5.0
|
| Scrutool Score Equal-weight average of all 10 dimensions |
7.2
|
Strong enough to defend a flag, not strong enough to trust blindly
Pangram does the hardest part of AI detection better than almost anyone. Its accuracy on raw AI text is excellent, and its false-positive rate is the one figure in this category backed by outside researchers rather than the vendor's own dataset, which is exactly what you need when a flag could cost a student or a writer their reputation. The holes are real and worth knowing before you pay. Image detection often reads a hidden file label instead of the picture, so routing an AI image through any editor strips that label and lets it pass as human. Confidence drops on short and non-English passages, paraphrased text slips through as it does with every detector, and the free tier is a small daily allowance that only lets you sample the engine. Public sentiment is mixed too, with genuine human writers reporting false flags. Use Pangram where a false accusation carries weight and you treat its score as evidence to investigate, not a verdict to act on. For casual checks, tight budgets, or image authenticity that has to survive an edit, look elsewhere.
Discussion
Join the discussion and share your perspective.