AI-generated photos and videos are now so realistic that a simple real-or-fake verdict is often not enough. A picture can be mostly genuine, with one face, object, or background swapped in, and a single score hides exactly where the problem lies. On November 4, 2026, Pangram launched two detectors aimed at that problem. Instead of stopping at a yes-or-no answer, they are designed to show where synthetic material appears, and, for video, when it appears.
What Pangram just launched
Pangram Image 1.0 is the company's new detector for AI-generated images. Pangram Image Flash 1.0 is a smaller, faster model for analyzing AI video frame by frame. The release also adds heatmap output, native video uploads, a Chrome extension, and API support for both images and video.
The heatmap is the point
Image 1.0's heatmap localizes synthetic material inside an image. That means it can point to an AI-generated element inside an otherwise real photograph, rather than only giving one score for the whole picture. Flash 1.0 takes a similar approach to video: it analyzes each frame of an uploaded video and flags which portions return an AI detection.
What Pangram claims
In an official launch post, Pangram said Image 1.0 identifies 99.2% of AI-generated images with a 0.27% false-positive rate. Those are Pangram's own claims, not independently verified results. On a company-created paired test, synthetic mirrors of 100 real photos made with 12 image generators, or 1,200 synthetic images in total, Pangram reports 99.96% accuracy. That figure should be read carefully: it is Pangram's own targeted test, not a universal accuracy rate for every image online. For video, Pangram says Flash 1.0 scores 98.7% on the DF26 benchmark, 29 points above the next-best detector. DF26 focuses on one person speaking direct-to-camera, and its authors reported that people and state-of-the-art detectors performed close to random chance on it.
The catches
The July 2026 research preview only accepted images 512x512 pixels or larger, and it could not confidently detect deepfakes and face swaps. Version 1.0 improves deepfakes, low-resolution images, and frontier models, according to Pangram. The minimum input is now 256x256, although images below 512 get a reduced-confidence disclaimer. The Chrome extension flags AI material across social media, shopping, and rental sites. Pangram also says it trains the model against compression, downscaling, blur, and noise, whether that distortion comes from social platforms or from evasion attempts. Detection is not the only approach:
watermarking can help when a mark is present, but it cannot rescue an image that was never marked in the first place.
Why it matters
Detection tools matter most when fakes mix real and synthetic parts, because that is exactly where a single yes-or-no verdict fails. A heatmap does not just say something is suspicious; it gives a reader, editor, or moderator a place to look. Other approaches, such as Google's
SynthID detector, show how the field is moving toward more specific evidence rather than a bare verdict. For anyone following
AI safety, the lesson is similar: a detector is only as trustworthy as its testing. Company benchmarks are a starting point, not proof, so treat heatmaps as strong evidence, not a verdict, and cross-check anything high-stakes. As generators improve, the detector-versus-generator race will keep escalating, and browser- and API-level detection is where most people will actually meet it.
Comments
Post a Comment