Detecting AI text was already a shaky business. Detecting AI video is harder, and the people building the generators are not standing still. So when Pangram announced that it is extending its detector to video and images, the right reaction is interest plus a raised eyebrow.
Pangram is best known for its text detector, which has built a reputation as one of the better ones in a field full of false accusations and weak results. Now subscribers can upload a video and get an assessment of whether AI was used to create or edit it. The company says the tool can separate clips that are partly AI from those that are fully generated, and it adds a heatmap that points to the parts of a video most likely to be synthetic. That partial-versus-full distinction matters. Most real-world misinformation is not a fully generated clip. It’s a real video with a swapped face, a cloned voice, or an edited scene.
On performance, Pangram says its video model “particularly excels at frontier AI video detection” and beat other detectors in testing. One internal test showed more than 98 percent accuracy. The image model, Pangram Image 1.0, flagged more than 99 percent of AI images with a false positive rate of 0.27 percent. Both figures are self-reported, and neither has outside validation yet. Detector accuracy tends to drop when a new generator ships, so a benchmark today says little about next quarter.
The image model is now available through Pangram’s Chrome extension, which scans images as you browse and flags likely AI ones. There are limits. Below 512 x 512 pixels, results carry a reduced confidence disclaimer. That likely explains why CEO Max Spero’s demo on Pinterest returned “unsure” on two images that were plainly not real photos. Pinterest is full of small thumbnails, so that is not a minor edge case.
The more interesting detail is the iPhone problem. Spero said most phone cameras, iPhones included, apply AI postprocessing to add detail to real photos. Early versions of the model flagged those shots as AI. He told Pirate Wires the company delayed release to test on several iPhone 18 Pros. This is the core difficulty for every detector. The line between a computationally enhanced photo and a generated one keeps blurring, and false positives against ordinary camera output would destroy trust fast.
Pangram is entering a crowded field. Hive and Reality Defender sell detection to platforms and enterprises, and provenance efforts like C2PA content credentials try to solve the problem from the other direction by labeling content at creation. Pangram’s angle is a consumer-friendly product with a browser extension, plus a text-detection brand people already know. Free accounts can check a handful of images, and the full video and image tools require a paid subscription.
Does it matter? Somewhat. A single vendor covering text, image, and video is useful for newsrooms, moderators, and teams vetting user submissions. But detection is an arms race, and the defender has to be right every time. Until independent testing backs up the claims, treat these tools as one signal, not a verdict.
