From text to images and video
Pangram is a New York startup whose main business is detecting AI-generated text. On November 4, it announced two visual detection models on its official blog: Pangram Image 1.0 handles images and can output a heatmap marking where AI edits or AI content appear inside an otherwise real photo; Pangram Image Flash 1.0, smaller and faster, focuses on video and reports frame by frame which segments are judged to be AI. Image detection had launched earlier as a research preview; this is the full release, alongside native video upload, image detection in the Chrome extension and an API that supports both images and video. According to Engadget, the features are available to paid subscribers, and free accounts can check a handful of images.
Where 99.96% comes from
The headline number in the blog is 99.96%, and the setup matters: Pangram took 100 real photos and used 12 frontier image generators to create a "synthetic mirror" of each with prompts closely matching the original, for 1,200 synthetic images; Image 1.0 reached 99.96% accuracy on this self-built evaluation set. Engadget reports another company figure: in testing it flagged more than 99% of AI images with a false positive rate of 0.27%. The minimum input size drops from 512×512 to 256×256, but results below 512 come with a reduced-confidence disclaimer.
On video, Pangram cites the DF26 academic study, which found humans are barely better than random at telling real talking-person videos from AI ones. The company says Flash 1.0 scores 98.7% on DF26, 29 points ahead of the next-best detector. It also built its own comparison: starting from the first frame of 129 real social media videos plus a description written by a vision-language model, it had 14 video generators create 100 mirrors each, for 1,529 videos in total, and compared its model with several commercial detection services.
The hard part is the phone camera
CEO Max Spero explained that most phone cameras, including iPhones, apply some AI post-processing to add detail to real photos, so the model has to tell that apart from AI edits and fully AI-generated images. Pangram also delayed the release because the model initially flagged photos taken with the iPhone 18 Pro as AI.
The test sets and comparisons above were all designed by Pangram, and no independent evaluation has appeared yet. For content moderation, fact-checking, e-commerce and rental platforms, tools like this can serve as one screening signal, but a detection result should not be treated as the only evidence, especially for low-resolution or compressed images.
via: Pangram: Introducing Pangram Image 1.0, Engadget: Pangram expands its AI detection tools to video and images