Pangram Extends Its AI Detector to Images and Video
Pangram has released tools for flagging AI-generated images and video, extending a detector that until now has been known mainly for text. Pangram subscribers can upload a video and receive an assessment of whether AI was used to create or edit the clip, according to Engadget, which reported the launch on November 4, 2026.
The video tool separates clips that are partially AI-generated from those that are entirely AI-generated. It also includes a heatmap, a visual overlay that marks which parts of a video are likely produced by AI. Pangram says the model is particularly strong at detecting frontier AI video, and that its internal testing showed more than 98 percent accuracy in flagging AI content.
Image support arrives as Pangram Image 1.0, available through the company's Chrome extension. The extension can scan images automatically while a user browses and flag those likely to be AI-generated. Pangram says the image model flagged more than 99 percent of AI images in testing, with a 0.27 percent false positive rate, meaning the share of real images wrongly flagged as AI. Results for images below 512 x 512 pixels carry a disclaimer that confidence is reduced. That caveat matters in practice, since many images circulating in feeds and thumbnails are that small.
Access is tiered. The new image and video capabilities are available now to subscribers, while a free account can check a handful of images without a paid plan. The free text checker, which covers output from ChatGPT, Claude and Gemini, is unchanged. Pangram's API has also been updated, alongside the release of Flash 1.0, to support both image and video detection. The company says its image detection covers outputs from major providers including OpenAI's GPT Image, and that its video detection covers Kling, Seedance, Veo and Wan, according to its blog post on Pangram Image 1.0 and its image detection announcement.
Pangram was founded in Brooklyn, New York, in 2023. In late July, it announced $9 million in funding from Menlo Ventures to expand its detector from text to images, as described in a Business Wire release. The company says its classifier is trained to detect the latest models from OpenAI, Anthropic, Google, Meta, DeepSeek and xAI. It also claims that independent researchers ranked it the most accurate AI detector, and that researchers at the University of Chicago and the University of Maryland verified its accuracy. Its website says independent studies found its detection outperforms trained human readers at identifying AI-generated content.
In my view, the caveats deserve as much attention as the headline numbers. A classifier, a model trained to sort inputs into categories, is only as current as its training data. Its accuracy describes the generators it was built on. What matters for a newsroom, a trust-and-safety queue or a verification workflow is how it handles footage and images from generators released after that data was assembled. The company is positioning itself to keep retraining against new releases, which is the only workable approach in a field where generators change every few months.
The broader context here is a shift in where detection happens. Text detection has been a classroom and publishing concern for years. Image and video detection moves the same question into the browser, where an ordinary user can check a suspicious picture before sharing it. That is a practical gain for anyone who has struggled to tell a synthetic photograph from a real one at a glance. It also means that the reliability of these tools, including their error rates on small or compressed files, now shapes everyday judgments made by people with no interest in model architecture.
