Pangram built its reputation on prose. On Wednesday the New York startup pushed into pixels: Pangram Image 1.0 for still photographs and Pangram Image Flash 1.0 for video, the latter analysing clips frame by frame.
The pitch is narrow and specific. Image 1.0 does not only answer "AI" or "not AI"; it returns a heatmap that localises the synthetic region inside an otherwise real photograph — a generated mural on a real restaurant wall, a swapped object on a real table. Flash 1.0 takes video natively: upload a clip and it reports which stretches of it look generated. Both are available in the web app and the API, and Image 1.0 ships in the company's Chrome extension, which now flags content in social feeds, shopping pages and rental listings.
The accuracy claims are strong and, for now, self-graded. Pangram says Image 1.0 catches 99.2% of AI images and falsely flags 0.27% of real ones. The headline figure — 99.96% accuracy — comes from a test it designed itself: 100 real photographs, synthetic "mirrors" of each produced with 12 frontier image generators using prompts matched to the original, 1,200 images in total.
The video number rests on outside work. DF26, an academic benchmark published in September, contains 2,691 clips of people speaking to camera; its authors found humans and existing detectors barely beat chance. Pangram says Flash 1.0 scores 98.7% there, 29 points ahead of the next-best detector. The company also built its own comparison across 1,529 videos, mirroring 129 real social clips with 14 generators.
Why video matters more than stills: talking-head fakes are where money actually moves — executive impersonation, fake endorsements, romance and investment fraud. Still-image detection has been commoditised for two years. Frame-by-frame video analysis sold by the API call is new, and it is expensive in exactly the way that keeps competitors out.
The caveats are structural rather than incidental. In August 2026 Pangram's text detector flagged a Berkeley professor's op-ed as partly AI-written; she disputed it. Reuters testing found Meta's own image detector missed cropped fakes from its own generator more than half the time. A YouTube network that paid real actors to read AI-written scripts defeated platform detectors. Each is a reminder that a detector is a probabilistic classifier, and that the grading is ultimately done by whoever pays for it.
One technical change cuts both ways: the minimum input resolution drops from 512 to 256 pixels, with a lower-confidence warning below 512. That is a practical decision, because social images arrive compressed and small — and it widens the surface on which a false positive can happen.
The larger argument is that detection and provenance are complements, not rivals. Signed capture in the C2PA mould works only for cooperative cameras and editors; it says nothing about the uncooperative majority uploading from a phone gallery. Classification is the only tool for that majority, which is why it will remain an arms race rather than a solved problem.
Commercially, Pangram raised $9 million led by Menlo Ventures in July, when image detection was still a research preview, and its text detector already runs on Substack. Platforms are moving too: Snapchat has stopped recommending AI videos on Spotlight.
What to do with a verdict: treat it as one signal in a stack, log which model version produced it, because thresholds move between releases, and never let a score alone trigger an enforcement decision about a person. In a dispute, the heatmap is the more defensible artefact — it points at a region a human can inspect.
