Skip to content

The Detector Now Points at the Part of Your Photo You Edited

Pangram Image 1.0 adds heatmaps, a 256-pixel floor and frame-by-frame video checks. The verdict that matters is no longer "AI or not". It is "partly AI", and that describes almost every retouched photograph.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A magnifying glass held over a stack of dark photographic prints, with one small bright patch catching the light inside the lens
Editorial illustration; not a detector output or a benchmark result.

On November 4, Pangram Labs took its image detector out of research preview. Text detection had been the company's reputation for three years; the image tool had been available in a limited form since July. Now there are two models: Pangram Image 1.0 for stills, and a smaller, faster Image Flash 1.0 for video, which checks clips frame by frame. Both run through an API, and the image model also ships in the Chrome extension.

The feature list is short. Heatmaps that mark where in a picture the AI content sits. A minimum input size cut from 512 to 256 pixels. Robustness training against compression, blur and noise. The rest is numbers, and the numbers are the part worth arguing with.

A detector that outputs a percentage invites one argument: is it right. A detector that outputs a map invites a different one, and that second argument is the one creators will actually have to survive.

The number, and who built the test

Pangram evaluated Image 1.0 on 100 real photographs. It fed descriptions of each photo to 12 frontier image generators, which produced 1,200 synthetic near-copies. The model scored 99.96% accuracy on that set. Against commercial rivals on the same set, Digital Trends reports TruthScan at 96.33%, AI or Not at 91.38% and Sightengine at 88.33%. The Next Web adds a second pair of figures from the company: 99.2% of AI images caught, and 0.27% of real images wrongly flagged.

The mirror-set design is the right instinct. By asking generators to reproduce real photos as closely as possible, the test makes content recognition useless as a shortcut, so the model has to key on generation artifacts instead of "this looks like a stock photo of a beach". That is a real methodological improvement over the "here are 500 AI images and 500 photos" style of test that used to dominate.

It is still a test built by the company that sells the detector. The 12 generators are almost certainly the ones the model was trained against, the photos are chosen by the same people, and none of the headline accuracy figures have been independently reproduced. The heise writeup says so plainly. Read the 99.96% as a ceiling for the benchmark, not as the accuracy you will see on your own files.

A percentage of a photograph

The 0.27% false positive rate sounds like nothing. It is one wrongly accused image per 370 genuine ones. If you deliver a 40-photo product set, that is roughly one image in ten shoots that comes back labelled as machine-made even though you shot it. At 2,000 images a week for a catalogue, the expected count is about 5.4 false flags, though any given week could be zero or higher. That is an estimate, and it assumes the reported rate even applies to your catalogue images.

Then the heatmap changes the shape of the accusation. The old detector answered "is this AI". The new one answers "how much of this, and where", and for a large share of professional photography the honest answer is a small region. Generative fill to remove a light stand. A sky replaced because the client wanted blue. A logo inpainted off a T-shirt. A dust patch cleaned on a museum scan. Every one of those is a real photograph that now contains AI-generated pixels, and locating those regions is what the heatmap is built to do.

That is not a flaw in the detector. It is a description of what image work became. The useful consequence is that "partly AI" needs to stop being an accusation and start being a category with a disclosure attached. A retoucher who says "yes, the sky is generative, here is the plate" is in a completely different position from someone claiming a fully synthetic portrait was photographed. The map tells you where to ask the question. It does not tell you what the answer means.

Frame by frame is the right call for video

For video, Pangram points at a paper rather than a claim. DF26 (arXiv:2609.07369) tested people on real and generated talking-head clips and found they were barely better than chance at telling them apart. Existing detectors were only slightly better. Flash 1.0 scores 98.7% on that benchmark, which the company reports as 29 percentage points ahead of the next best detector.

The design choice that matters is per-frame analysis. A clip is not one object. It is a sequence in which 12 frames may be regenerated and 200 may be untouched, and a single whole-clip score throws that structure away. Frame-level output localizes the synthetic section, which is the only form of the answer that is useful in a dispute: "these two seconds are generated, the rest is camera".

Pangram built its own video comparison set for this, and the construction is worth knowing because it shows the same mirror logic: 129 real videos from social platforms, a vision language model describing what happens in each, then 14 frontier video generators producing 100 videos each, sampled from those 129 originals with a description plus the real first frame as input. That is 1,529 videos including the originals. Again, company-built, company-scored.

Independence cuts both ways here. A detector that could be beaten by naming a generator is a bad detector, and Pangram is clearly training against the newest ones. But the published figures and the production figures are not the same measurement, and the gap between them is where every provenance dispute will land.

Three things it still cannot do

It cannot tell you who made the image. Detecting synthesis is a property of pixels. Authorship is a property of a person, and no classifier has access to that.

It weakens on the pipeline that social platforms apply. Pangram reports training and testing against compression and downscaling, which is why the 256-pixel floor exists at all: a thumbnail crop is often all a platform keeps. But below 512 pixels the result carries an explicit reduced-confidence warning, and most of the images people argue about online have been re-encoded, cropped and screenshotted at least twice. The verdict does not become useless there. It becomes a lead, and a verdict on a 300-pixel screenshot should not be treated as evidence.

It cannot separate "AI-assisted" from "AI-generated and tidied up by hand". Those two workflows can end with pixel-level similarity and completely different amounts of human craft. The map is a possible indication of where machines touched the file — it can flag genuine regions and it can miss edits. It says nothing about the hours on either side of that touch.

What to keep, starting now

If any part of your output involves generated pixels, the file you deliver is no longer sufficient evidence of how it was made. The map is a claim about regions, so the counter-evidence has to be regional too.

Keep the plate. The pre-edit frame or the original RAW, with its own capture metadata, is the artifact that answers a heatmap most directly, because it lets you check each flagged region against what was there before. Keep the working file, not just the export: layer stacks and history are the record of what you changed and in what order, and they cost nothing to store next to the deliverable. Keep the native resolution, because a 1024-pixel JPEG that has been through three apps is a weaker artifact than the 6000-pixel original, and detectors agree. And when a client asks, answer with the region rather than the mood: "the sky is generated, everything else is the shoot" is a defensible sentence, and the plate plus the layer stack are what make it checkable. The heatmap is the thing that prompted the question, not the proof of the answer.

There is a compliance reason this is arriving now. The EU AI Act's transparency obligations have been phasing in since August 2026, and heise frames the detector market as a response to them. But do not mistake a detector for a labeling system. C2PA provenance metadata travels inside the file and is routinely stripped by editing, re-saving and platform upload. Invisible watermarks such as SynthID are built to survive more than that, including common edits and compression, which is why visible watermarks end up optional while the invisible ones stay in the argument, and why detecting text watermarks is its own arms race. Detection is what you reach for after the metadata is gone.

The practical rule I would take from this release is unglamorous. Assume every delivered image will be scored, and assume a small region of it may be flagged as generated even when you did almost nothing. Then keep the plate, keep the layers, keep the resolution, and write down what you changed. That is a ten-minute habit per job, and it is the difference between a heatmap that supports your account and one that contradicts it.

We wrote about the weaker version of this problem a week ago, when a Spanish literary prize went to a novel a text detector scored as 99% AI. The lesson there was that a detector score is not a verdict. The image version sharpens it: the score has become a map, the map is more informative, and it is still not a verdict. It is an opening question with coordinates.

Sources

Share this guide
X LinkedIn