Pangram has introduced two AI detection models—Image 1.0 for still images and the faster Image Flash 1.0 for video—that demonstrate near-perfect accuracy based on internal and academic benchmark tests, aiming to improve identification of AI-generated visual content across digital platforms.
- Image 1.0 detects AI images with 99.2% accuracy and flags 0.27% false positives
- Image Flash 1.0 achieves 98.7% accuracy on DF26 video benchmark, outperforming competitors
- Models resilient to reduced image size and common visual distortions like compression
What happened
Pangram, a New York-based startup known primarily for AI-written text detection, has launched two new models for identifying AI-generated visual content. Their Image 1.0 model targets still images, while Image Flash 1.0 processes video frame by frame at high speed. Both leverage Pangram’s advanced machine learning techniques to detect synthetic elements in photos and videos with remarkable precision.
The company deployed these models in several formats: Image 1.0 is integrated into Pangram’s Chrome extension, which flags AI content on social media and various sites, while both models are accessible through the web app and API. This facilitates detection workflows for publishers, platforms, and developers who want to verify image and video authenticity.
Why it matters
The rise of AI-generated media poses growing challenges around misinformation, digital fraud, and trust in online content. While there are existing detection tools, many struggle with accuracy, especially under conditions like compression, blur, or smaller image sizes. Pangram’s models push detection capabilities forward by achieving near-perfect scores on internal tests and a leading 98.7% on the DF26 academic video benchmark, significantly outperforming current alternatives.
Their technology’s ability to produce heatmaps pinpointing AI-generated regions within images provides additional context for users and platforms examining suspicious content. This is crucial as deepfakes and AI-manipulated visuals become more sophisticated and harder to spot by human viewers or less capable tools.
What to watch next
Pangram’s footprint in the AI detection ecosystem is rapidly expanding, particularly with its $9 million funding round led by Menlo Ventures earlier this year and its existing text detection service on platforms like Substack. Adoption of these new image and video models will be a key factor to watch, especially as social networks and marketplaces increase their focus on combating AI fakes.
However, challenges remain in the broader industry as no detector is infallible. Past incidents, such as tools mislabeling genuine content or missing subtle deepfakes, highlight the need for ongoing improvements. Monitoring how Pangram and competitors refine their models to handle adversarial content and integrate into content moderation workflows will be critical for the future of digital authenticity verification.