Pangram, the AI-detection company best known for flagging machine-generated text, has officially launched Pangram Image 1.0 and Pangram Image Flash 1.0, a pair of models the company says set a new state of the art for detecting AI-generated and AI-edited images and videos. The announcement, published on Pangram's own blog on November 4 and picked up by Engadget, Unite.AI and Yahoo Tech, converts a research preview that the company released a few months ago into a generally available product with API support.
The launch comes at a time when synthetic media has become genuinely difficult to distinguish from authentic content, both for casual scrollers and for the moderation systems platforms rely on. It is one of the most closely watched corners of AI industry coverage right now. Pangram's bet is that dedicated detection models, trained specifically against the outputs of today's frontier generators, can restore some of that lost signal.
Frontier Image Detection With 99.96% Accuracy
The flagship Pangram Image 1.0 model was evaluated against a benchmark the company built itself: 100 real photographs, each mirrored with synthetic versions produced by 12 frontier AI image generators using prompts closely matched to the original photo, for a total of 1,200 synthetic images. On that evaluation set, Pangram reports 99.96% accuracy — a figure that, if it holds up under independent scrutiny, would put the model well ahead of the commercial and academic detectors the company compared it against.
Beyond a simple real-or-fake verdict, the model outputs heatmaps that localize exactly where AI content appears within an image. That matters for the increasingly common middle ground between fully synthetic images and untouched photographs: marketing images with AI-generated backgrounds, photos retouched with generative tools, or composites that mix real subjects with generated elements.
Pangram also claims the new model is robust to the kinds of transformations that reliably defeat naive detectors — social media compression, deliberate blurring and noise, and low upload resolutions. The company says its previous research preview required images of at least 512x512 pixels, while Image 1.0 accepts inputs down to 256x256, with a reduced-confidence disclaimer shown below the old threshold.
Video Detection and a 29-Point Leap on DF26
The companion model, Pangram Image Flash 1.0, is aimed at video, and it is here that the company's claims are most striking. Pangram cites a recent academic study, DF26, which found that humans are barely better than random when distinguishing real videos of people talking from AI-generated ones — and that existing automated detectors were not much better. On that same DF26 benchmark, Pangram reports a score of 98.7%, which it describes as a 29-point leap over the next best detector.
To benchmark against the commercial market rather than academia, Pangram constructed a difficult comparison set from 129 real social media videos, generating synthetic mirrors of each using 14 different frontier video generators — 100 videos per generator, seeded with the first real frame of each source video and driven by generated descriptions of the action. Across the resulting 1,529 videos, the company compared Flash 1.0 against multiple major AI video detection services. The mirror methodology is designed to remove content bias: each real video and its synthetic twin share the same scene and subject, so a detector cannot game the test by learning which topics or aesthetics correlate with AI content.
The Flash model also performs frame-by-frame analysis, which lets it flag AI-edited segments inside otherwise authentic footage — a capability aimed at the growing problem of partially manipulated videos rather than fully synthetic clips.
From Text Detection to a Full Synthetic Media Stack
Pangram built its reputation on AI text detection, where its classifiers are widely used in education and publishing, and Yahoo Tech's coverage of the launch described the company as the gold standard in that category. Extending into images and video is a natural expansion, and it positions Pangram as a one-stop shop for platforms, newsrooms and enterprises trying to police synthetic media across formats.
The new models ship with API support for both image and video detection, and a Chrome extension that applies detection to images and videos as users scroll their feeds — moving detection from a back-office moderation tool to something closer to consumer-grade protection. Pangram says it also benchmarked the models against academic datasets including X-AIGD, Mirage-Test, AIGenImages2026, Synthbuster paired with RAISE, and Community Forensics.
The Detection Arms Race Continues
One caveat deserves emphasis: every number in Pangram's announcement comes from Pangram's own evaluations, conducted on its own mirror-based benchmarks. Independent validation is how detection tools ultimately earn trust, and the history of AI detection is littered with tools whose accuracy collapsed against generators they had never seen. Pangram itself acknowledged as much in its post, thanking the community for trying the research preview and finding its weaknesses.
Still, the launch is a meaningful marker of how far detection technology has progressed in response to generative video's rapid improvement. If the reported DF26 result survives third-party testing, the gap between what humans can perceive and what machines can verify will only widen — which is precisely the market Pangram is betting on.
For platforms and publishers, the practical takeaway is that high-accuracy detection at low resolutions, including frame-level video analysis, is becoming available as a commodity API. The next phase of the synthetic media problem may be less about whether detection is possible and more about who deploys it, and at which layer of the internet's plumbing.
Stay Ahead of AISynthetic media, detection tooling and the platforms that adopt them are moving targets — stay current on all of it.
Read more AI news →
