Pangram has unveiled two detectors for AI-generated media: Image 1.0 for still images and Image Flash 1.0, a faster model that analyzes video frame by frame. Image 1.0 also produces a heatmap highlighting which regions of a photo appear synthetic. The company claims near-perfect accuracy, but a significant share of those figures stems from tests it designed itself.
- New detectors: Image 1.0 and Image Flash 1.0 identify AI-generated images and videos, available via web app, API, and browser extensions.
- Impressive but self-reported numbers: Pangram claims 99.96% accuracy on its own image test set and 98.7% on the DF26 video benchmark, a 29-point lead over the next best tool.
- Lower resolution support: The minimum image size drops from 512×512 to 256×256 pixels, though results below 512 pixels carry a reduced-confidence warning.
What the new detectors do and where they work
The models analyze images and videos and estimate whether they were produced by AI, with tools aimed at both everyday users and enterprises. Both are available in the web app and the API. Image 1.0 also powers a Chrome extension that flags AI content on social networks, shopping sites, and rental listings, while users can upload videos to the web app for Image Flash to review.
A concrete technical improvement: the system now handles images as small as 256×256 pixels, down from a previous 512×512 minimum. Results below 512 pixels include a lower-confidence disclaimer, an honest acknowledgment that detection is more fragile at low resolution. The heatmap in Image 1.0 is the most useful feature for users: it doesn’t just say “this is AI” but points to where.
Pangram says the models are trained to resist common evasion techniques such as blur, noise, and compression artifacts. Browser extensions for Chrome and Firefox operate in real time, analyzing images without manual uploads, and the API now supports both image and video detection for third-party developers.
Reading the numbers with caution
Pangram states that Image 1.0 detects 99.2% of AI images and falsely flags only 0.27% of real ones, while Image Flash achieves 98.7% on the academic DF26 benchmark. These are impressive figures, but with caveats worth highlighting.
First, in a test with AI copies of 100 real photos generated by 12 different tools, Pangram reports 99.96% accuracy. That test was designed by Pangram, not external reviewers, which doesn’t invalidate it but means it should be treated as an in-house figure. The second caveat weighs more: DF26, the only independent benchmark, contains 2,691 clips of people speaking to camera and was published in September; its authors concluded that both humans and leading detectors “barely beat chance.” That Image Flash scores 98.7% there, 29 points above the second best according to Pangram, is notable but clashes with that broader picture of difficulty.
Reliability is the crux of the matter, because a false positive has consequences. Pangram’s own text tool flagged an op-ed by a Berkeley professor as “partially written by AI” in August, which she denied. And according to Reuters tests, Meta’s image detector failed more than half the time with cropped images. The problem isn’t just catching what’s fake but avoiding accusations against what’s authentic—crucial when discussing AI-edited images that, according to an MIT study, double memory errors.
Why there’s a market for this
Synthetic content is already used to scam, manipulate, and fake reviews, and something is needed to flag it. Pangram is a New York startup known for detecting AI-written text; its text detector already runs on Substack, and in July it raised $9 million (about €8 million) in a round led by Menlo Ventures when it first introduced image detection.
The use case is evident in phenomena such as ads featuring celebrities who “recommend” investments that are actually deepfakes, or fake reviews that the law already pursues and that consumers should know how to spot before buying.
The sensible reading is that detectors like these are necessary and, at the same time, dangerous if trusted blindly. Necessary because the alternative—an internet where no one distinguishes real from generated—is worse. Dangerous because a 99% figure invites treating its verdict as proof, and no detector is that: Pangram’s own history with that professor and Meta’s with cropped photos demonstrate it. The prudent approach is to use them as what they are, one signal among several, never as a sole judge. And to demand that the tests supporting those percentages be conducted by independent third parties, not the company selling the product. As long as the only external benchmark says detecting AI “barely beats chance,” any in-house 99% deserves at least a raised eyebrow.