Pangram Launches AI Detection Tools for Images and Videos With 99% Accuracy Claims

Pangram Expands AI Detection Capabilities Beyond Text

Pangram, widely recognized for its AI text detection tools, has launched two new models designed to identify AI-generated content in images and videos1

1

2

2

. The New York-based startup released Pangram Image 1.0 for still images and Pangram Image Flash 1.0 for video analysis, marking a significant expansion of its detection capabilities into multimedia content2

2

. These AI detection tools are now available through the company's web app, API, and Chrome extension, allowing users to scan content as they browse social media, shopping platforms, and rental listings3

3

.

Impressive Accuracy Rates Claim Market Leadership

Pangram claims its image detection model achieves an astounding 99.96% accuracy rate on its own evaluation set, with a false positive rate of just 0.27%1

1

3

3

. In the company's internal testing, Pangram created AI copies of 100 real photos using 12 leading image generators with prompts closely matching each original2

2

. These self-reported figures position Pangram well ahead of competitors, with TruthScan achieving 96.33% accuracy, AI or Not reaching 91.38%, and Sightengine at 88.33%3

3

. For AI-generated videos, Pangram Image Flash 1.0 demonstrates 98.7% accuracy on the DF26 academic benchmark, a 29-point lead over the next-best detector2

2

3

3

.

Advanced Features for Detecting AI Fakes

The Pangram Image 1.0 model introduces a heatmap feature that highlights which specific aspects of an image are likely AI-generated, enabling users to distinguish between partially AI-edited content and entirely AI-generated visuals1

1

2

2

. The video detection model performs frame-by-frame analysis and can differentiate between clips that are partially AI-modified versus completely AI-generated1

1

. Pangram CEO Max Spero explained that the model has been trained to see through compression, blur, and artificial noise that platforms and individuals trying to fool detectors might add2

2

. The company has also improved its handling of low-resolution images, now supporting images as small as 256 x 256 pixels, down from the initial 512 x 512-pixel minimum, though results below 512 pixels carry a reduced-confidence warning2

2

3

3

.

Technical Challenges and Development Hurdles

Developing robust AI detection tools presented unique challenges for Pangram. Spero noted that most phone cameras, including iPhones, perform AI postprocessing to add detail to real photos1

1

. "We had to build a model that's robust to this postprocessing while understanding the difference between AI edits and fully AI-generated visuals," he explained1

1

. The company delayed its image model release to test it on several iPhone 18 Pros after discovering that initial versions incorrectly flagged images from the phones' cameras as AI-generated content1

1

. For video testing, Pangram used the DF26 benchmark, an academic dataset of 2,691 clips of people speaking to camera published in September, where both humans and leading detectors previously performed little better than chance at identifying fakes2

2

.

Market Position and Future Implications

Pangram raised $9M in a funding round led by Menlo Ventures in July when it first previewed image detection capabilities2

2

. The company's text detector already runs on Substack, and its expansion into multimedia detection positions it as a comprehensive solution for identifying AI-generated content across formats2

2

. However, deepfake detection remains an imperfect science. AI detectors still make mistakes, and even Meta's own image detector missed cropped fakes from its generator more than half the time in Reuters tests2

2

. Pangram's accuracy figures are self-reported rather than independently verified, and the company previously faced controversy when its text tool flagged a Berkeley professor's op-ed as partly AI-written, which she disputed2

2

. As platforms like Snapchat crack down on AI videos and the volume of AI-generated content continues to grow, reliable detection tools will become increasingly critical for maintaining trust in digital media.