Pangram launches AI image and video detection models

Pangram, a New York startup known for spotting AI-written text, launched Pangram Image 1.0 for still images and the smaller, faster Pangram Image Flash 1.0 for video, which checks video frame by frame, it said on Wednesday. Both run in its web app and API, and Image 1.0 also ships in its Chrome extension, which scans images as a user browses.

The image model can draw a heatmap showing where AI content sits inside a picture, such as a fake object added to a real photo. On video, Engadget reported, the tool separates clips that are partially AI from those entirely generated and flags which parts of the frame are likely synthetic. The models also work on smaller files than before, down to 256 by 256 pixels from 512 by 512, and give a lower-confidence warning below 512. The API now handles both images and video. The company had released a research preview of Image 1.0; the model is now available through the Chrome extension.

Accuracy claims and how they were tested

Pangram says Image 1.0 catches 99.2% of AI images and wrongly flags 0.27% of real ones. Engadget put the detection claim at more than 99%, with the same false positive rate.

The 99.96% figure comes from a test Pangram built. It made AI copies of 100 real photos with 12 leading image generators, using prompts that closely matched each original. Pangram designed the test, not outside reviewers.

For video, the company used DF26, an academic benchmark of 2,691 clips of people speaking to camera, published in September. Pangram says Flash 1.0 scores 98.7% on DF26, 29 points ahead of the next-best detector. Engadget reported more than 98% accuracy for the video model, citing one internal test the company detailed.

Why detection remains difficult

The DF26 benchmark is unkind to everyone else. Its authors found that both people and leading detectors did little better than chance at telling real clips from fakes. That gap is the context for the 98.7%.

Detectors still make mistakes. 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. Pangram has misfired too: in August its text tool flagged a Berkeley professor's op-ed as partly AI-written, and she disputed the finding.

CEO Max Spero said most phone cameras, including iPhones, do some AI postprocessing. He told Pirate Wires the company delayed the image model to test it on several iPhone 18 Pros because it had initially flagged their images as AI. Spero shared an image of the Chrome extension running on Pinterest; it returned an "unsure" result on two images that are obviously not genuine photos.

Availability and funding

Pangram said the tools are available now to subscribers, and free accounts can check a handful of images without a paid plan. 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.

Pangram says it trains the models to see through compression, blur and noise, which platforms and people trying to fool detectors can add.