Pangram adds AI video checks and brings image detection to Chrome

Pangram users can now check uploaded videos for AI involvement and scan images as they browse in Chrome. The company released Image 1.0 and Image Flash 1.0 on November 4, 2026, adding visual detection tools alongside its text detector. Pangram says the video tool can distinguish AI-edited clips from those generated entirely by AI.

From browser feeds to uploaded clips

The image model follows an earlier research preview. Its integration into Pangram’s Chrome extension lets it flag likely AI-generated images on social feeds, shopping pages and rental listings. The image and video features are available to subscribers; according to Engadget, people with free accounts can also check a limited number of images without paying for a plan.

For video, Pangram now accepts uploads directly. The company says its smaller, faster Image Flash 1.0 model examines clips frame by frame to identify portions it detects as AI-generated. The video tool also provides a heatmap highlighting likely AI content, rather than returning only a judgment about the whole clip. Pangram says it can separate partial AI use from fully generated footage.

Image 1.0 likewise produces a heatmap intended to locate AI-generated material within an image. Pangram’s example shows a real photograph containing an AI image, illustrating why a result may need to identify a region rather than classify every pixel alike. Developers can access both image and video detection through the company’s updated API, its interface for connecting other software.

What the accuracy figures measure

Pangram reports 99.96% accuracy for Image 1.0 on an evaluation built around synthetic versions of real photographs. It started with 100 real photos and used 12 image generators to create 1,200 synthetic images, with prompts closely matching the originals. That figure describes performance on the company’s evaluation set, not a measured success rate for every image someone might encounter while browsing.

Its industry video comparison uses a similar design. Pangram collected 129 real social-media videos and generated 100 synthetic clips with each of 14 video generators. Each generated clip began with a real video’s first frame and a machine-written description of its action. The resulting comparison covered 1,529 videos across real and synthetic footage.

The purpose of these matched pairs is to keep the scene and subject similar. Pangram says that makes the test harder and reduces the risk of a detector learning to label particular kinds of content as AI. The intended distinction is between a real recording and a generated counterpart, rather than between unrelated subjects.

Separately, Pangram says Image Flash 1.0 scored 98.7% on DF26, beating the next-best detector by 29 percentage points. The DF26 paper describes a narrower setting than video generally: single-person public-speaking clips, including direct-to-camera recordings, official statements and studio interviews. Its dataset contains 271 real videos and 2,420 synthetic ones.

The researchers found that humans and the deepfake detectors they tested performed close to random chance. That finding explains the challenge the benchmark poses; Pangram’s much higher score is the company’s own reported result, distinct from the researchers’ findings.

Real photos complicate the boundary

Size is not the only complication. Pangram says it specifically trains its image detector to handle compression, downscaling, blur and added noise. Those changes can arise when images move through social platforms or when someone deliberately tries to fool a detector. The company says its model is robust to these transformations.

Camera software presents a different problem. CEO Max Spero said many phone cameras, including iPhones, use AI postprocessing to add detail to real photos. He told Pirate Wires that Pangram delayed the image-model release to test several iPhone 18 Pro devices after the detector initially flagged their camera images as AI-generated.