Pangram launches AI image and video detectors, citing high test scores
The startup reports a 98.7% score on an academic video benchmark, while its separate image test was designed internally rather than by outside reviewers.
New York startup Pangram has introduced two detection models: Pangram Image 1.0 for still pictures and Pangram Image Flash 1.0 for video. The company said the Flash model is smaller and faster, examining individual video frames for AI-generated content.
Pangram reported a 98.7% score for Flash 1.0 on DF26, placing it 29 points above the next-best detector. Published in September, the academic benchmark contains 2,691 clips showing people addressing a camera. Its authors found that human viewers and leading detection tools performed only slightly better than chance.
For Image 1.0, Pangram reported 99.96% accuracy in an internally designed evaluation. It used 12 leading image generators to produce AI versions of 100 real photographs, with prompts closely matching the originals. Outside reviewers did not design that test.
The company separately says its image model detects 99.2% of AI images while incorrectly identifying 0.27% of real images as artificial. It can produce a heatmap locating suspected AI elements, including objects inserted into otherwise genuine photographs.
Both models support images measuring 256 by 256 pixels, compared with 512 by 512 previously. Results for images below 512 pixels include a warning about lower confidence. The models are available through Pangram’s web app and API; Image 1.0 also works in its Chrome extension.
Detection errors remain possible. In August, Pangram’s text detector identified part of a Berkeley professor’s opinion essay as AI-written, a finding she disputed.
