Pangram Launches Image 1.0, an AI Image Detector That Pinpoints Edited Areas
Pangram, known for its AI text detection tools, launched Pangram Image 1.0 on November 4, an image detector that displays a heatmap of suspected areas, along with Image Flash 1.0 for video. Its near-100% success rates, however, primarily stem from tests designed by Pangram itself.
Pangram, a startup known for its AI text detection tools, launched Pangram Image 1.0 on November 4, 2026, to detect AI-generated or edited images, alongside Pangram Image Flash 1.0, a lighter model highlighted for video. The most tangible improvement is not the headline accuracy score, but a heatmap that reveals exactly where synthetic content appears within an image.
Mapping Suspicious Areas
Traditional detectors typically return a single, binary verdict: “AI” or “not AI.” Yet a significant share of misleading imagery consists of real photos that have been retouched: an object inserted, a face altered, or a background swapped out. Pangram Image 1.0 returns a grid of localized scores that pinpoints synthetic elements surrounded by real-world imagery; the launch example highlights an AI-generated wall mural blended into a real restaurant photo.
According to the API documentation, each scan yields one of six verdicts (ranging from “No AI” to “AI Detected,” plus “Partially AI” for partially generated visuals), a probability score between 0 and 1, a confidence rating, and, when present, C2PA metadata (the “Content Credentials” embedded by select tools and cameras) and EXIF data. If a validated C2PA manifest confirms AI generation, it can settle the assessment without running the model. Otherwise, Pangram claims to inspect the pixels themselves, without depending on watermarks.
Another update: the detector now accepts images starting at 256 x 256 pixels, down from 512 previously, though inputs below 512 pixels trigger a reduced-confidence warning. Pangram also claims to train its models to withstand compression, blur, and noise—artifacts frequently applied by social media platforms or added deliberately by bad actors to evade detection.
Video, Chrome Extension, and API
Image Flash 1.0 scans videos frame by frame to flag generated segments. The web platform now supports video uploads, the API covers both media types, and Image 1.0 is integrated into Pangram’s Chrome extension, which flags generated content directly while users browse social networks, marketplaces, or rental listings.
Figures That Require Scrutiny
Pangram claims a 99.96% accuracy rate on a benchmark of 1,200 images synthesized by 12 recent generators using prompts modeled on 100 authentic photos, alongside a 98.7% score for Image Flash on DF26, an academic dataset of talking-head videos. Its product page boasts a 99.5% score on 1,130 images when compared against five commercial detectors. According to The Next Web, Pangram reports catching 99.2% of AI images while misidentifying real photos only 0.27% of the time. These benchmarks were largely designed by Pangram itself, and no independent third-party evaluations have yet been published for this release.
Our take
The heatmap represents genuine functional progress: for a journalist, a content moderator, or a buyer vetting a listing, knowing where an image was altered is far more actionable than an aggregate percentage. Still, an automated detector remains a clue, not definitive proof. Even a 0.27% false-positive rate, across the billions of images circulating on social networks, translates into substantial volumes of authentic photos erroneously flagged, and every new generator reignites the cat-and-mouse game between creation and detection. Scores nearing 100% stem above all from tests curated by the vendor. Verifying C2PA metadata, when available, remains the most reliable signal; pixel inspection merely fills in the blanks left by missing provenance data.
Sources
- Pangram — Introducing Pangram Image 1.0 (November 4, 2026) Primary source
- Pangram — API Documentation, Image Detection Primary source
- Pangram — AI Image Detector Product Page Primary source
- The Next Web — Pangram's new models spot AI fakes in images and video, it says Secondary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
