Pangram’s new models spot AI fakes in images and video, it says

Pangram launched Pangram Image 1.0 for AI images and the faster Image Flash 1.0 for video. It claims near-perfect scores on its own tests and 98.7% on the DF26 academic benchmark. Both are in its web app and API, and Image 1.0 is in its Chrome extension.

Pangram, a New York startup known for spotting AI-written text, has launched two models for detecting AI images and video, the company said on Wednesday. Pangram Image 1.0 checks still images, while the smaller, faster Pangram Image Flash 1.0 checks video frame by frame.

The image model can also draw a heatmap showing where AI content sits in a picture, such as a fake object added to a real photo. Pangram says it catches 99.2% of AI images and wrongly flags 0.27% of real ones.

In its own test, Pangram made AI copies of 100 real photos with 12 leading image generators, using prompts that closely matched each original. Image 1.0 scored 99.96% accuracy on that set. The test was designed by Pangram, not by outside reviewers.

For video, Pangram used DF26, an academic benchmark of 2,691 clips of people speaking to camera, published in September. Its authors found that both people and leading detectors did little better than chance at telling real clips from fakes.

Pangram says Flash 1.0 scores 98.7% on DF26, 29 points ahead of the next-best detector.

The new models also work on smaller images, down to 256 by 256 pixels from 512 by 512. Results below 512 pixels carry a lower-confidence warning. Pangram says it trains the models to see through compression, blur and noise, which platforms and people trying to fool detectors can add.

Image 1.0 now runs in Pangram’s Chrome extension, which flags AI content on social media, shopping and rental sites. Users can also upload videos in the web app, and the API now handles both images and video. Platforms are cracking down too: Snapchat has stopped recommending AI videos on Spotlight.

Pangram raised $9M in a round led by Menlo Ventures in July, when it first previewed image detection. Its text detector already runs on Substack.

AI detectors still make mistakes. In August, Pangram’s text tool flagged a Berkeley professor’s op-ed as partly AI-written, which she disputed.

Even Meta’s own image detector missed cropped fakes from its generator more than half the time in Reuters tests. On YouTube, a network that paid real actors to read AI scripts beat its detectors.