Can AI Images Be 99% Detected? A Look at Pangram's New Chrome Extension
機械翻訳 / Machine-translated
A
AI Friends
1日前・14分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever looked at a photo on social media and wondered, "Is this even real?" On November 4, 2026 (US time), Pangram Labs announced a new model and Chrome extension claiming image detection accuracy of "over 99%." This article breaks down what it offers and what to watch out for.
According to ITmedia AI+'s coverage, Pangram Labs is a US-based startup. On November 4, they announced "Pangram Image 1.0," a model for distinguishing AI-generated images.
Alongside it, they also released "Pangram Image Flash 1.0," a faster model designed for video. They also introduced a Chrome extension called "Pangram AI Detection," which alerts users when an image they're viewing may have been AI-generated.
The predecessor, "Pangram Image 0.1," was released as a research preview in July 2026. Version 1.0 reportedly incorporates user feedback and improves handling of deepfakes (convincingly realistic fake imagery), low-resolution images, and images created by the latest generation of AI.
Image 1.0 is said to detect 99.2% of AI-generated images, with a false positive rate of 0.27%. A false positive means mistakenly flagging a genuine photograph taken by a human as "AI-generated."
In a test using 1,200 synthetic images based on real photographs, the model detected 99.96%. It reportedly exceeded 99% on most other benchmarks (standardized performance tests) as well.
Flash 1.0, the video-oriented model, scored 98.69 on "DF26," an AI video detection test—reportedly 29 points ahead of the second-place model.
It's important to keep in mind that these figures come from the company's own announcements. The ITmedia AI+ article also does not report any third-party verification.
For instance, even with a 99% detection rate, a 0.27% false positive rate remains. If you run 10,000 genuine photos through the system, a simple calculation suggests roughly 27 of them would be incorrectly flagged as "AI-generated." Using this tool alone as grounds to accuse someone of posting fake content would still be a stretch.
According to the Chrome Web Store listing, you can right-click any image or text to get a verdict, which appears in the bottom-right corner of your screen.
On X, LinkedIn, Substack, Reddit, and Medium, posts are labeled with badges: "Human," "AI-Assisted," or "AI." Opening the sidebar lets you view the human-to-AI ratio as well.
A free Pangram account is required to use the extension. According to ITmedia AI+, installing the extension itself is free, but automatic scanning of social media is a paid plan feature. The listing page showed English as the only supported language.
Imagine you're browsing product reviews on an e-commerce site and start wondering, "Could this photo be a fake AI-generated review?" With the extension, a single right-click gives you a rough answer.
The API (a mechanism that lets other software call on this functionality) pricing for businesses and developers is as follows:
For example, if a news editorial team checks 1,000 submitted images every day, Image 1.0 would cost $250 per day. For high-volume use cases, it may be practical to mix and match with the cheaper Flash 1.0.
Pangram isn't the only AI image detection tool out there. HIVE AI Detector is another Chrome extension that can scan text and images on a page with a single click. Its listing page does not mention pricing.
What sets Pangram apart is its dedicated video model, the ability to automatically badge social media posts, and its API for integration into enterprise systems. That said, accuracy figures from Pangram and other providers are all self-reported, so there are no results from head-to-head comparisons under identical conditions.
Pangram is primarily known as a service for detecting AI-written text. It announced a $9 million funding round on July 29, 2026 and has since expanded into image detection.
For users in Japan, this extension is relatively easy to try, though the listing page is in English only. Account creation and results are also displayed in English, which may present a bit of a barrier for those less familiar with the language.
Social media managers and PR professionals in Japan could use it to verify images received from business partners or to fact-check photos tied to viral controversies. E-commerce site operators might also find it useful for checking product images submitted by sellers.
Imagine a desk editor at a news outlet who, during breaking disaster coverage, needs to quickly determine whether a posted photo is authentic. In that case, the detection tool's results should be used alongside direct confirmation with the original poster and verification of the shooting location.
Installation is free. However, it has been reported that automatic social media scanning requires a paid plan.
We wouldn't recommend it. The figures are the company's own, and a 0.27% false positive rate still exists. Treat verdicts as a reference point and use them alongside source verification.
There is a video-oriented "Flash 1.0" model that displays suspected AI-generated segments using a red heatmap (a diagram where color intensity indicates likelihood).
The Chrome Web Store listing showed English as the only supported language. Japanese support has not been confirmed.
Start by trying the extension on one image that catches your eye, and compare the result against your own impression.
This article is a cross-post from AI Friends.
