Pangram adds AI video detection and brings its image detector to Chrome

Pangram, the Brooklyn start-up best known for spotting AI-written text, now checks video too. On 4 November 2026 it announced that subscribers can upload a clip and learn whether AI made or altered it, and that its flagship image model, Pangram Image 1.0, now also runs in the company's Chrome extension, which scans pictures as a person browses. Pangram calls the release state of the art for image and video detection, and its API now covers both.

The video detector is a separate model, Pangram Image Flash 1.0. It reads a clip frame by frame, tells partly AI footage from fully synthetic footage, and draws a heatmap over the stretches it thinks were generated. Pangram reports 98.7% on the academic DF26 benchmark, 29 points ahead of the next detector, and says it tested on 1,529 clips: 129 real social videos plus synthetic twins from 14 video generators. Unlike watermark checks such as SynthID or C2PA, which look for marks embedded by the maker, Pangram reads patterns in the pixels themselves. Engadget cites a figure of more than 98% from an internal test. All of these numbers come from the company itself.

“We’re getting new GPUs faster than new people are being born. If we do not actively discriminate in favor of human content, then we’re just gonna see more and more AI, and it’s just gonna drown out any human signal that we have.”

Max Spero, Pangram CEO, to TechCrunch in July · techcrunch.com ↗

For images, Engadget reports that the model flagged over 99% of AI pictures with a 0.27% false-positive rate, while Pangram's own post gives 99.96% on 1,200 synthetic images made from 100 real photos with 12 generators. The July preview put false positives at 0.16% on the ReLAION set, a different test, so the figures are not directly comparable. The minimum image size has dropped from 512 by 512 pixels to 256 by 256, with a reduced-confidence note below 512. A Pinterest screenshot Pangram shared showed "unsure" on two obviously fake photos, which Engadget put down to low resolution.

The hard part, according to CEO Max Spero, was that phone cameras apply AI processing to real photos, so the model had to separate edits from fully generated pictures. The launch was delayed to test several iPhone 18 Pro phones after the first version flagged their camera output as AI. Spero said on X that the team built the model to hold up under that post-processing while still distinguishing AI edits from wholly AI-generated visuals. Free accounts can check a handful of images; video upload and full access need a subscription at $20 a month, which both Engadget and TechCrunch report.

It is worth separating what is claimed from what has been checked outside the company. The 98.7% benchmark score and the internal accuracy figure above both come from Pangram, and Engadget presents the claim that the video model beats rival detectors as the company's own. The two image false-positive rates, 0.27% now and 0.16% in July, were measured on different test sets. In July the company claimed 99.5% accuracy for the image model against 98% for the next-best detector, while TechCrunch's own trial that month found it caught most AI images but labelled one photo of an AI image as human.

The image model began as a research preview on 29 July 2026, alongside Pangram 4 and a $9 million round led by Menlo Ventures. That preview covered GPT Image, Nano Banana, Midjourney, FLUX, Grok Imagine and video generators such as Kling, Seedance, Veo and Wan. Pangram says it will keep iterating on both models.