Pangram takes AI detection from text to images and video

Pangram has expanded its AI-detection product beyond text, releasing tools intended to assess uploaded images and video. Pangram Image 1.0 handles still images, while the faster Image Flash 1.0 powers video analysis, examining uploads frame by frame and identifying areas it detects as AI-generated.

The company’s image tool can generate a heatmap for mixed human and synthetic images, showing the regions that triggered its assessment. Pangram has also put Image 1.0 in its Chrome extension, where it can flag material while users browse social media, shopping sites, rental listings and other webpages. Its API now supports both image and video detection.

The release develops an image-detection research preview Pangram published on July 29. That earlier version was designed to identify outputs from image generators including GPT Image, Gemini’s Nano Banana, Midjourney, FLUX and Grok Imagine, as well as frames from several video-generation systems. It could not process images smaller than 512 by 512 pixels and was not designed to confidently identify deepfakes and face swaps.

Image 1.0 lowers the minimum accepted image size to 256 by 256 pixels. Pangram said it trained the model to accommodate compression, downscaling, blur, noise and other edits, but results for images below 512 pixels carry a reduced-confidence disclaimer. It also cautioned that even small generative changes can lead an image to be flagged more broadly, because contemporary editing tools may regenerate an entire picture.

Pangram reported 99.96% accuracy for Image 1.0 in an evaluation built from synthetic versions of 100 real photographs. The company used 12 image generators to create 1,200 synthetic images. Those are company-reported results from a purpose-built test, rather than an independent measure of performance in general use.

For video, Pangram said Image Flash 1.0 scored 98.7% on DF26, a benchmark of 2,691 real and synthetic clips of people speaking, including direct-to-camera recordings, statements and studio interviews. It said the score was 29 percentage points ahead of the next-best detector. Heise noted that Pangram’s comparative video figures had not been independently verified.

The tools assess visual characteristics associated with generation, rather than establishing a file’s origin, its editor or the sequence of edits. That differs from provenance systems such as Content Credentials, which create tamper-evident media-history records, and watermarking systems such as Google’s SynthID, which embeds an invisible signal in supported AI-generated images and video segments at creation.

Brookings said no current AI-detection tool performs well across all technical and policy considerations, and described detectors as one part of a wider credibility toolkit. It also warned that machine-learning watermark detectors can be defeated by small, imperceptible pixel changes.