Known for its AI text detection, Pangram has officially launched its multimodal detection capabilities. The Pangram Image 1.0 tool, which was previously only available as a research preview, is now fully operational on the web, company API, and the namesake Chrome extension.
According to Pangram, the image detector achieves an accuracy of 99.96% in identifying AI-generated images on its evaluation set. The company also stated that the model's accuracy has improved over the past few months, particularly in handling edited images, low-resolution images, and images generated by the current leading AI models. In comparison with similar tools, TruthScan has an accuracy of 96.33%, AI or Not at 91.38%, and Sightengine at 88.33%, highlighting a significant difference.
However, it should be noted that these figures come from each company's self-constructed evaluation sets, and the criteria are not uniform, so horizontal comparisons should be taken as reference only.
Advantages of Pangram Image 1.0 in Video Detection
For video detection, a different approach has been adopted. Pangram has prepared a lighter model, Pangram Image 1.0 Flash, which processes videos through frame-by-frame detection. The company claims that its accuracy in identifying AI-generated videos on the DF26 benchmark is 98.7%, leading the second place by 29 percentage points; on the same benchmark, TruthScan is at 84.2%, while Sightengine drops to 81.8%. To validate the effectiveness, Pangram selected 129 real videos from social platforms and generated corresponding versions using 14 of the latest AI video models to create a test set.
Pangram also stated that its model has been compared with detectors developed by academic researchers, consistently maintaining a lead in benchmark tests.
The key technical aspect lies in "resilience to manipulation." The model has been specifically trained to see through common disruptive artifacts such as compression, blurriness, and artificial noise—issues that have historically caused detectors to fail. Additionally, it can highlight specific areas in images generated by AI using heat maps, rather than simply providing a yes or no conclusion. The detection threshold has also been lowered: the minimum image size has decreased from 512×512 pixels at launch to 256×256 pixels, although results below 512 pixels will come with a warning of reduced confidence.
In terms of usability, Pangram has integrated image detection capabilities into its Chrome extension, which marks AI content as users scroll through social media, shopping platforms, and rental listings. Users can also directly upload videos on the website for the model to determine if they are AI-generated. It is worth noting that figures like 99.96% come from the vendor's own evaluation set, and real-world performance depends on how many times images have been saved, compressed, and cropped. The tug-of-war between detection tools and generative models currently shows no signs of ending.