Pangram’s AI detector now reads video too, and it tops its own benchmarks

Pangram has released Pangram Image 1.0 and Image Flash 1.0, new models that detect AI-generated images and video, according to a Techmeme item pointing to the company’s announcement post. Pangram describes the image model as state-of-the-art, and says the smaller Image Flash model powers its new video detection.

One model for stills, a lighter one for video

Pangram Image 1.0 is the full image detector. It follows a research preview the company released a few months ago, which was its first step into detecting synthetic pictures after years of focusing on text. The company says the new version is trained to hold up when images are compressed, shrunk, blurred or noisy, the conditions in which pictures usually circulate on social networks.

The detector also produces heatmaps that show which regions of an image look AI-generated, so a reader can see where the model’s suspicion comes from instead of getting only a verdict. Image Flash 1.0 is described as a smaller, faster model built for frontier AI video. Video can be uploaded directly, and the tool analyzes it frame by frame and reports which parts are flagged.

Pangram’s post compares both models with several academic benchmarks and with other commercial detection services. In its own wording, the results put it ahead of the field. On video, it also says people are “barely better than random” at telling real footage of talking people from AI footage, which is its argument for a dedicated tool.

Where you can use it today

Image detection is available in Pangram’s web app. A browser extension for Chrome and Firefox flags AI content while users browse social media, shopping sites and rental listings. The API now supports both images and video, so other products can call the detectors directly. The post does not state pricing and points readers to Pangram’s pricing page. It also does not publish a false positive rate for the new release in text form, a gap that matters for anyone who might use a detector to judge a real person’s photo.

The preview’s record shows what to check in 1.0

The analysis here rests on the earlier preview, because the new post leans on charts rather than error rates. In its preview post, Pangram published a false positive rate measured on older human-made images, a test that matters more than headline accuracy because a detector that wrongly labels real photos does harm. The same post admitted that deepfakes and face swaps were not reliably detected, and that attempts to strip or defeat watermarks were still being worked on. The 1.0 post does not clearly settle either point.

Outside testing of the preview was mixed. A TechCrunch test, republished by Yahoo Finance, reportedly found that the tool caught several generated images but labeled a photographed AI image as human. We read that only as a search-result summary and did not open the article, so treat the detail as secondhand. It does show the practical weak spot: an AI image that has been re-photographed or heavily edited is harder than a clean file.

The new numbers belong to Pangram. The benchmark sets, the generators used to build its internal test images and the commercial detectors it compared against are chosen by the company, and none of the results has been independently verified. Image generators also change often, so a detector that leads today can fall behind after a major model release.

What the announcement leaves open

The post does not name every generator behind its internal tests, which makes it hard to judge how the models cope with the newest releases. It gives no independent audit and no statement on how often the detectors will be retrained. For newsrooms, platforms and marketplaces, the useful test is to run the tool on their own material, including edited and re-photographed files, and to treat a flag as a lead for human review rather than proof.

The release does add something that was missing from Pangram’s product: one account that covers text, still images and video. Whether that convenience comes with trustworthy error rates is the question the company’s next published evaluation, or an outside lab, would need to answer.