Pangram releases Image 1.0 to detect AI-generated photos and videos

Public launch of multimedia tools

Pangram, a company known for its AI text detection software, has officially released its multimedia detection tools to the general public. While the company previously kept these tools in a research preview phase, they are now accessible to wide audiences. Users can interact with the technology through a web interface, a company API, or a specialized Chrome extension.

The new release, called Pangram Image 1.0, is designed to identify pictures and videos created by artificial intelligence. The company aims to provide a standard for verifying digital content as synthetic media becomes more common. This move follows Pangram's established reputation as a trusted name in the field of AI text analysis.

The tool is intended to work across various online environments, including social media and shopping platforms. By integrating into a browser, the technology can flag content in real-time as people scroll through their feeds. This accessibility makes it easier for non-technical users to verify the authenticity of the media they encounter daily.

Performance and accuracy benchmarks

According to reports from Yahoo Tech, Pangram claims its image detector reaches 99.96% accuracy on its own evaluation set. The company has spent several months refining the model to handle difficult scenarios. This includes identifying pictures that have been edited, low-resolution files, and content from the most advanced AI models currently available.

The company also released a specific version for video content called Pangram Image 1.0 Flash. This is a lighter-weight model designed to perform frame-by-frame detection to maintain speed. It reportedly achieves 98.57% accuracy on AI-generated videos, which Pangram notes is a significant lead over its closest competitors.

In testing against academic research models, Pangram consistently outperformed other benchmarks. For instance, the company noted that its video detection holds a 29-point lead over the next best detector on the DF26 benchmark. These figures suggest that the tool is currently among the most precise options for identifying synthetic video content.

Technical features and capabilities

Pangram Image 1.0 is built to recognize subtle artifacts that often trick standard AI detectors. The model is trained to see through common issues like artificial noise, blurring, and image compression. By identifying these hidden markers, the tool can maintain high accuracy even when the media quality is degraded.

One of the more advanced features of the image detector is its ability to generate a heatmap. This visual aid highlights the specific parts of a picture that the AI identifies as synthetic. This allows users to see exactly where the manipulation occurred rather than just receiving a binary pass or fail result.

The tool also supports smaller files than it did at launch. It can now check images as small as 256 x 256 pixels, whereas the previous minimum was 512 x 512 pixels. However, the company notes that results for images smaller than 512 pixels will include a warning regarding reduced confidence in the final score.

Testing and real-world use

To verify the effectiveness of the video detector, Pangram conducted a series of tests using social media content. They took 129 real videos and used 14 different AI video models to create synthetic versions of them. This testing process helped the company measure how well the tool handles various types of AI generation.

The Chrome extension is designed to provide a seamless experience for people browsing the internet. It can flag AI content as users visit websites for online shopping, rental listings, or social media. This proactive approach helps users stay informed about the nature of the images they see while browsing different platforms.

For users who want to perform more manual checks, the company provides a website where they can upload videos directly. The model then analyzes the uploaded files to determine if they were generated by AI. This multi-channel approach ensures that different types of users can access the detection technology based on their specific needs.