Pangram Launches AI Image Detection Models That Point at the Fake Pixels

AI image detection gets a practical upgrade

Pangram launched two new AI image detection models on November 4, 2026, and they are aimed at a problem that has moved from research labs into everyday life. Image 1.0 detects AI-generated images, while the smaller Image Flash 1.0 analyzes video frame by frame. Both add a heatmap view that highlights the parts of a picture that look synthetic, so you do not just get a yes or no verdict. You get a map of where the suspicion sits.

That detail matters. A single score for a whole image is hard to act on, especially when only one element in a real photo has been altered. A heatmap lets a journalist, a marketplace moderator, or a regular user look at the flagged region and decide whether it makes sense. If the hands look wrong in a portrait, or a product in a listing photo has impossible reflections, the heatmap points straight at it.

Pangram says Image 1.0 identifies 99.2 percent of AI-generated images with a 0.27 percent false-positive rate, and reports 99.96 percent accuracy on a company-created test set. Those are vendor numbers, so treat them as a claim to be tested by independent reviewers, not as settled fact. The more interesting claim is about coverage: the models now accept images as small as 256 by 256 pixels, down from 512, although smaller images carry a reduced-confidence disclaimer.

How AI image detection works in practice

The release also moves detection out of a standalone tool and into places people already work. Image 1.0 is built into a Chrome extension that flags synthetic material across social media, shopping, and rental sites, and an updated API supports both image and video checks. That distribution matters more than any single benchmark, because detection only helps if it shows up before someone shares, buys, or rents based on a fake picture.

Video is the harder half of this launch. Image Flash 1.0 is built for speed, and Pangram says it scores 98.7 percent on DF26, a benchmark described in a September 2026 preprint, which it says is 29 points above the next-best detector. Again, that is a vendor-reported result on a specific test, and real-world video is messier than any benchmark: compression, re-uploads, and crops all degrade signals. But a fast video model at least makes frame-by-frame checking feasible at scale, where a slow model would only ever be used on a handful of clips.

There are limits worth stating plainly. Pangram's July 2026 research preview could not confidently detect deepfakes and face swaps, and the company has not claimed those cases are fully solved. Low-resolution images still get a weaker verdict. And no detector can settle the question of intent: a synthetic image used as obvious satire, art, or illustration is not the same as one used to deceive, and a tool cannot tell you which one you are looking at.

Why this launch lands now

The timing is not accidental. AI-generated images and video have become cheap to produce and hard to spot, and the usual defenses (watermarks, metadata, provenance labels) only work when the creator cooperates. Detection is the fallback for everything else: the screenshot of a screenshot, the listing photo with no metadata, the clip that arrived through three different apps before it reached you.

That is also why Pangram's heatmap idea deserves attention even if you doubt the headline accuracy numbers. A detector that explains itself, even roughly, is more useful than a black box with a higher score. It gives a human reviewer something to check, and it gives the person who posted the image a specific claim to answer. In moderation work, that is often the difference between a fair decision and a coin flip.

Still, detection is an arms race, and every public detector trains the next generation of generators to evade it. The honest way to use tools like these is as one signal among several: check the source, reverse-search the image, look for corroboration, and treat a clean verdict as "no evidence of fakery found" rather than proof of authenticity. Pangram's models look like a useful step in that direction, provided users keep that distinction in mind. The fake-pixel heatmap is a good aid to judgment, not a replacement for it.

For readers who want to check their own habits, our guide to checking a link before you click it covers the same skeptical routine in 30 seconds, and our fake online store guide shows how synthetic images feed shopping scams. Detection tools help, but the habit of pausing before you trust a picture helps more.