Pangram Claims 99.96 Percent Accuracy Spotting AI Images and a 29-Point Lead on Deepfake Video Detection

The arms race between AI image generators and the tools built to catch them moved a step on November 4, as detection startup Pangram released Pangram Image 1.0 and a smaller companion model, Pangram Image Flash 1.0, which the company describes as state of the art in both AI image and AI video detection. The release upgrades a research preview from a few months ago into a production offering with a Chrome extension and an API.

The flagship model's headline number comes from the company's own evaluation: Pangram built synthetic mirrors of 100 real photographs using 12 frontier AI image generators, for a total of 1,200 synthetic images prompted to closely match the originals, and reports 99.96 percent accuracy on that set. The company also published a heatmap feature that localizes exactly which region of an otherwise real photograph contains AI-generated content, a capability aimed at the increasingly common case of composite images rather than fully synthetic ones.

Robustness is the other selling point. Social platforms compress and downscale images, and bad actors add blur, noise and other artifacts specifically to defeat detectors. Pangram says the new model is trained against such transformations, and it lowered the minimum input resolution from 512 by 512 to 256 by 256 pixels, with results below 512 carrying a reduced-confidence disclaimer. The company also says the model holds up on a set of academic benchmarks including X-AIGD, Mirage-Test, AIGenImages2026, Synthbuster paired with RAISE, and Community Forensics.

On video, the numbers are starker. A recent academic study, DF26, found that humans are barely better than random at telling real videos of people talking from AI-generated ones, and that existing detectors were not much better. Pangram says Image Flash 1.0 scores 98.7 percent on DF26, which it calls a 29-point leap over the next best detector. The company also ran an industry comparison, taking 129 real social media videos and generating AI mirrors with 14 frontier video generators, for 1,529 total videos, and says it outperformed multiple commercial detection services on that set. Frame-by-frame analysis is now native, so a partially AI-edited video gets flagged segment by segment.

All of these figures are self-reported, on test sets the company itself constructed, and Pangram's comparisons against commercial rivals appear in charts the company published without releasing the underlying evaluation harness. Independent researchers have repeatedly shown that detector accuracy degrades on new generators, on images passed through social platforms, and on adversarial perturbations, so the claims deserve the usual caution until third parties replicate them. The references to published academic benchmarks are a step toward verifiability, but the benchmark scores themselves come from the vendor.

The distribution matters as much as the benchmarks. A Chrome extension that flags AI content while users scroll social media, shopping sites and rental listings turns detection from an enterprise tool into a consumer habit, and the API opens the same models to platforms that need to screen uploads at scale. The timing is not accidental: the release lands amid mounting platform battles over AI slop, with game marketplaces, mod sites and stock services all scrambling to quantify how much of their catalogs are machine-generated.

Detection has always been a cat-and-mouse game in which each generator improvement erodes last quarter's accuracy numbers, which is why Pangram's framing emphasizes breadth, robustness to compression and localization rather than a single accuracy figure. If the numbers hold up under independent testing, the practical takeaway is that fully synthetic images and talking-head deepfakes are becoming detectable at very high confidence, while the harder problem, a real photo with a small AI-edited region, now has a tool that at least claims to point at the pixels.