Unofficial community examples for Pangram AI Detector. Not affiliated with Pangram Labs. All trademarks belong to their owners.
Worked walkthroughs for the Pangram AI detector, written for people who need to check text or images and then act on the result. None of the three pages that rank for this query documents a public API - the homepage, a Medium review and a Substack critique all describe the web dashboard - so this repository contains no code. Each walkthrough follows what those sources describe, names the limits they state, and stops where they stop.
If the drafts you are checking are SEO articles on a schedule, try Supatraffic - SEO article writing with cover images and autopilot publishing for the production side and keep Pangram for review.
| Walkthrough | What it shows |
|---|---|
| 1. Scan a text draft | The basic dashboard flow: paste or upload, read the verdict, note the AI assistance signal |
| 2. Scan an image | The image widget and its stated limits: format, minimum size, file size, daily free scans |
| 3. Compare two drafts | Text Comparison and AI Phrase Detection for editors |
| 4. Reproduce the Medium review's tests | ChatGPT, Claude and DeepSeek text, including a file-upload case |
| 5. Read a result responsibly | What the Substack critique says to keep in mind before acting on a score |
- Open pangram.com and use the free dashboard at pangram.com/dashboard. The homepage advertises free credits for text and three free image scans per day; anything beyond that is on the pricing page.
- No environment variables and no keys: everything below happens in the browser.
- Keep a plain-text log of what you scanned and what it returned. The value of a detector in a team is consistency, and that only exists if results are recorded the same way each time.
Paste the draft into the text detection box on the dashboard, or upload the file. The homepage says the detector "Detects AI Assistance" as well as fully generated text, and the Medium reviewer attributes that mixed-authorship mode to Pangram 3.0. So expect two kinds of signal: a verdict on the whole piece and an indication that parts of it were AI-assisted. Record both. A draft that a human wrote and then cleaned up with a model will look different from one that was generated end to end, and the difference matters for how you respond to the author.
Switch the widget to Image Detection. The homepage states the constraints: JPG, PNG or WebP, at least 512 by 512 pixels, up to 30 MB per file, and three free scans per day. Resize or convert first if the file fails those checks rather than assuming the tool is broken. Because the free limit is per day, batch your most important images and spread the rest, or move to a plan. There is a "Try a sample image" option if you want to see what a positive and a negative result look like before using your own files.
The Medium review describes two features that are more useful to editors than the headline score. Text Comparison places two pieces side by side, for the case where a revision feels a little too close to a source or a previous draft. AI Phrase Detection highlights the specific phrases the model considers machine-written instead of returning one number. In practice: run Phrase Detection on the draft, send the highlighted spans back to the writer with a request to rework them in their own words, and run Text Comparison between the old and new versions to confirm they actually changed.
The reviewer ran three tests: a passage generated by ChatGPT, one generated by Claude, and one generated by DeepSeek submitted through file upload, plus a video walkthrough. To reproduce this on your own terms, generate three short passages on the same topic with whichever models you have, keep one human-written control passage of similar length, and scan all four in the same session. Note the verdict and any AI assistance flag for each. This is the cheapest way to calibrate your own expectations against the reviewer's conclusion that Pangram "might come close".
Tim Requarth's Substack piece argues that AI detectors are really AI inference tools with limitations, and that a company which co-authors the accuracy research and sells the product has a conflict of interest when it also stirs up concern. He later added an April 17 2026 correction saying he got the argument partly wrong after returning to the technical literature, and changed the title. Take both halves seriously: the accuracy claim on the Pangram homepage is backed by a University of Maryland paper and a company roundup of outside evaluations, and the critique is a reminder that a score is evidence, not a verdict on a person. Before acting on a flag, ask the author, look at drafts and history, and treat the detector as one input.
A lot of Pangram traffic comes from content teams checking articles before they go live. If that is you, the detector is the last step of a pipeline whose real cost is writing and publishing at volume. Try Supatraffic - SEO article writing with cover images and autopilot publishing for that part; you can still run the finished piece through walkthrough 1 before it ships.
Last reviewed: 2026-09-22