Pangram, the company best known for detecting AI-written text, has released a pair of models that flag AI-generated images and video, and it is putting numbers on the claim that they work. The pitch lands at a bad moment for anyone who relies on a photo or a clip as proof of something. Pangram published the details on its blog.
The image model is Pangram Image 1.0, a follow-up to a research preview the company put out a few months earlier. Its headline feature is a heatmap: when a real photograph has AI-made content mixed in, the tool marks where in the frame the synthetic part sits. That matters for edited pictures, where a yes-or-no verdict on the whole file would miss the point.
For its accuracy figure, Pangram took 100 real photos and had 12 current image generators produce lookalikes from closely matched prompts, giving 1,200 synthetic images. The company says the model reached 99.96 percent accuracy on that set. Pairing each fake with a real original keeps the test honest about subject matter, though Pangram built the test itself, and the blog post does not name the commercial detectors it was compared against.
Robustness is the other claim. Social platforms compress and shrink pictures, and people trying to dodge detection add blur or noise. Pangram says it trains on those distortions and that the model holds up. It also lowered the minimum image size from 512 by 512 pixels to 256 by 256. Below 512, results carry a reduced-confidence warning.
Video goes to a separate, smaller model called Pangram Image Flash 1.0, which analyzes uploads frame by frame and shows which stretches come back as AI. Pangram cites an academic study called DF26, which found people only slightly better than chance at telling real talking-head videos from AI ones, with other detectors not much better. On that benchmark, Pangram reports 98.7 percent, 29 points above the next best detector.
The company also ran its own industry comparison. Its team gathered 129 real clips from social media and made synthetic copies with 14 video generators, each seeded with the real clip’s first frame and a written description of what happens next. That produced 1,529 videos in total. The setup is deliberately hard, since the fake and the original share the same scene.
Access comes three ways. Video upload is now built into Pangram’s site, a Chrome extension flags AI content while you scroll social feeds, shopping pages, and rental listings, and the API now accepts both images and video. The rental-listing example is the telling one: the buyers for this kind of tool are not only newsrooms but anyone who pays money based on what a picture shows.
Detection is an arms race with a structural disadvantage, because every new generator can be tested against public detectors before it ships. Pangram’s results cover generators that exist today, and the blog offers no figures for ones that have not been released. Independent testing will say more than a vendor’s benchmark.
For any team that moderates user uploads or verifies listings, the practical step is cheap: run a few hundred of your own known-real and known-fake files through the API and see whether 99.96 percent survives contact with your data.
Reported by the Pangram blog (Isaac Kasahara), in its post “Introducing Pangram Image 1.0”; the source file carries no publication date.
