Testing Pangram, an AI detector for text, images and video

AI-generated content is spreading across social media. One estimate puts it at more than half of LinkedIn posts and as much as 45% of TikTok content, while imperfect labeling tools often fail to flag it. Pangram says its detector can help users tell authentic material from AI-made content. I tried it on several kinds of media to see how well it holds up.

Pangram checks text, still images and video, pitching itself as a broad tool for spotting AI-generated material. That promise matters: earlier AI detectors have sometimes labeled human-made work as machine-generated, which can make their results hard to trust.

The company says its false-positive rate is just 0.01% and that independent researchers have ranked the service first. I put those claims to a practical test using real images, AI-altered images, generated media, my own writing and text produced by AI.

Text test

I started with a short poem about daffodils written by Grok. Pangram marked it as “possibly AI” with 85% certainty, adding that a longer sample could help it reach a firmer conclusion.

That’s a reasonable caveat, though 85% still leaves some uncertainty. Short social media posts may not give a detector much to work with.

Next, I gave the same prompt to ChatGPT, Claude, Gemini and Siri. Pangram identified all four responses as AI-generated, each with 100% certainty.

For comparison, I submitted the first three paragraphs of this article, which I had written myself. Pangram labeled them human-written with 100% certainty. That was a strong result.

Image test

I then checked a real photo of a woman in an autumn setting. Pangram correctly reported that it had not detected AI use.

I also submitted a slightly altered version of the photo. This time, the tool flagged AI involvement with high certainty. It displayed a heat map highlighting areas it identified as signs of AI editing.

Video test

For the video trial, I asked Gemini to generate a clip of a boat moving along the East River beneath the Brooklyn Bridge. Pangram identified it as AI-generated with 100% certainty, which was correct.

Verdict

AI detectors have a history of false alarms, and those errors can have real consequences. Students, for example, have had to push back against claims that they used AI to complete their work when they say they did not.

In my tests, Pangram performed well. It recognized the AI-generated examples and correctly identified the human-written text and unaltered photo. The main uncertainty came with Grok’s short poem.

That limitation may matter when checking brief posts, which don’t always provide enough text for a confident assessment. Pangram says it is designed to detect writing from ChatGPT, Claude and Gemini, so identifying the Grok sample as potentially AI-written was still a useful result.

Using Pangram

Some features can be tried without an account. Uploading files and scanning URLs require signing in.

During my test, scans used credits: one for every 100 words, five for an image and one for each second of video. Those rates appeared to differ slightly from the figures on Pangram’s support page, so they may change. The daily allowance was 20 credits, which could go quickly if you have a lot to check.

For occasional checks, Pangram seems worth having available. It won’t settle every case, especially when the sample is short, but it handled most of my test material convincingly.