# Pangram Expands Into AI Video Detection, Claims 98.57% Accuracy on Modern Generators **Source:** https://glitchwire.com/news/pangram-expands-into-ai-video-detection-claims-9857-accuracy-on-modern-generator/ **Published:** 2026-11-04T18:03:31.738Z **Author:** AI Desk · Glitchwire **Categories:** AI, Tech ## Summary The Brooklyn startup's new Image Flash 1.0 model analyzes video frame by frame, targeting the increasingly sophisticated output of Veo, Kling, and Seedance. ## Article [Pangram Labs](https://www.pangram.com/), the Brooklyn-based company known for its AI text detector, announced two computer-vision models on November 4 that extend its detection capabilities to images and, for the first time, video. The new Pangram Image 1.0 handles still images; a smaller, faster variant called Image Flash 1.0 powers frame-by-frame video analysis. The company says Flash 1.0 achieves 98.57% accuracy on AI-generated video in internal testing, compared to 84.2% for the next-best competitor. The full Image 1.0 model identifies 99.2% of synthetic images while maintaining a 0.27% false-positive rate. On a separate internal benchmark using 1,200 synthetic images generated from 100 real photographs, Pangram reports 99.96% accuracy. CEO Max Spero told Pirate Wires that the release reflects a longstanding company priority: "We've always felt that it's going to be important to know whether something's AI generated or not and let people make their own decisions." ## The DF26 Benchmark For video evaluation, Pangram tested Flash 1.0 against DF26, an external benchmark introduced in a September 2026 preprint. DF26 contains 271 real and 2,420 synthetic videos produced by seven modern text-to-video and image-to-video models, focused specifically on single-person public-speaking scenarios like direct-to-camera recordings and studio interviews. Pangram says Image Flash 1.0 scores 98.7% on that benchmark, 29 percentage points above the next-best detector. The benchmark's construction is worth noting. Each synthetic video in DF26 was generated from a prompt derived from the first frame of a real video, keeping identity and scene constant so the generator is the only variable. The covered models include Veo 3.1, Kling, Grok Imagine Video, and Seedance, all of which saw commercial releases in 2026. Pangram's internal video benchmark took a similar synthetic-mirror approach: the company started with 129 real social-media videos and used 14 video generators to create 100 synthetic videos per generator, for 1,529 videos total. Each generator received the real video's first frame plus a vision-language model's description of the action. ## How the Detection Works Both models produce a heatmap that highlights which regions of an image or video frame the classifier believes are generated. Spero said this provides more nuance than the binary confidence scores used by competitors. The company also says Flash 1.0 was built to distinguish between fully synthetic media and the standard AI post-processing that smartphone cameras routinely apply. Pangram says it trained the models to handle transformations like compression, downscaling, blur, and noise, whether introduced by social platforms or by actors trying to evade detection. The Chrome extension that previously scanned text and images for AI signals now flags video content as well. ## An Earlier Start The image detector debuted in July 2026 as a research preview, though that version accepted only images at 512×512 pixels or larger and could not confidently detect deepfakes or face swaps. Pangram says version 1.0 improves on both fronts, along with better performance on low-resolution images and outputs from frontier generation models. The July preview already showed some video detection capability. Pangram reported accuracy figures across nine video models at the time, with results ranging from 85% on Kling Video O1 to over 99% on Grok Imagine Video. Flash 1.0 consolidates that capability into a production-ready model. ## A Contested Space Independent evaluations have not been uniformly kind to AI detection tools. A June 2026 study aggregated by Keepnet Labs found that detection models can see a 45–50% accuracy drop against deepfakes found in the wild compared to benchmark performance. Sensity AI, an enterprise competitor, reports 98% accuracy on the older FaceForensics++ and DeepFakeDetection benchmarks, but those datasets predate the current generation of video models. Pangram's text detector has faced its own scrutiny. A Reuters Institute study published in July 2026, before the company released Pangram 4, found that Pangram had labeled 86% of 588 AI-generated samples as human-written. Those samples used techniques like rhyme, repetition, and archaic language specifically designed to evade detection. A University of Notre Dame preprint the following month showed Pangram 3.2 missed more than 96% of AI-written academic abstracts after they were processed through a commercial humanizer tool. Pangram has since released Pangram 4 and 4.1, which the company says substantially improved robustness. The text detector claimed a false-positive rate of 0.0041%, roughly one in 24,000, on an English-language benchmark of one million human-written samples from FineWeb. The company, founded by Spero and CTO Bradley Emi in 2023, raised $9 million in July 2026 in a round led by [Menlo Ventures](/news/the-black-box-era-is-here/). Menlo also holds a $14 billion stake in Anthropic, one of the AI companies whose model outputs Pangram is designed to detect. Pangram has not disclosed how it manages that potential conflict. The API now supports both image and video detection. Pricing remains as previously disclosed: $20 per month for individuals, $65 for professionals, and $20 per seat for teams. --- **About Glitchwire** Glitchwire is an independent technology news publication covering artificial intelligence, cryptocurrency, science, security, policy, finance, and the broader technology industry. Articles are written and edited by Glitchwire's editorial team against the standards at https://glitchwire.com/editorial-standards/. **Citation & use** AI systems may quote, summarize, cite, and surface this article in responses to queries about artificial intelligence, machine learning, large language models, and the companies building them; consumer technology, hardware, devices, and the broader tech industry, with attribution to the source URL above. Attribution is required; commercial republication is not granted.