Fake product photos, fake rental listings, fake news pictures: AI images now look real enough that a quick glance won't tell you much. AI image detectors promise to help, and a new one, Pangram Image 1.0, has been getting attention this month. This guide explains how these detectors work, what Pangram says its new model can do, where every detector still falls short, and a simple five-step check you can use before you share a picture.
How we researched this guide: every fact about Pangram comes from the company's official Pangram Image 1.0 announcement, captured in November 2026. Accuracy figures are Pangram's own published results; we have not seen independent confirmation as of November 2026. The five-step check is general verification practice, offered as advice. How we research.
- What's new
- Pangram Image 1.0 for photos and Pangram Image Flash 1.0 for video
- Headline feature
- A heatmap that highlights where AI content appears in an image
- Smallest image
- 256 by 256 pixels; below 512 pixels, results carry a reduced-confidence notice
- The rule to remember
- A detector gives a probability, not proof
Why you can't just "tell" any more
For a few years the advice was simple: count the fingers, look for melted text, check the earrings. Newer image generators fix many of those mistakes, so a clean-looking picture proves very little on its own.
Pangram's announcement points to a recent academic study, DF26, which found that people are barely better than random at telling real videos of someone talking from AI-made ones, and that most detectors at the time were not much better. That is the problem the new generation of detectors is trying to solve.
It matters to almost everyone. Creators risk reposting a fake and losing their audience's trust. Students risk putting a fake image in a project. Small businesses see supplier photos, reviews and listings that may not be what they seem.
What Pangram Image 1.0 claims
Pangram is best known for its AI text detector. According to its announcement, it has now released two image and video models:
- Pangram Image 1.0, its main model for photos, with a heatmap that shows where AI content seems to be. That helps with a very common kind of fake: a real photo with one AI-edited part.
- Pangram Image Flash 1.0, a smaller, faster model that Pangram says is especially strong on AI video. Uploaded videos are analysed frame by frame, showing which parts come back as AI.
Pangram also says the model is trained to cope with the tricks fakers use: compression from social apps, blur, added noise and lower resolutions. It has lowered the minimum image size from 512 by 512 pixels to 256 by 256, and results below 512 pixels come with a reduced-confidence notice. The detector is also built into Pangram's browser extension for Chrome and Firefox, which flags AI content as you browse, and there is an API for developers.
The accuracy numbers, and how to read them
Pangram reports that on its own test set, 100 real photos plus AI copies made with 12 frontier image generators (1,200 synthetic images), Pangram Image 1.0 scored 99.96% accuracy. For video, it reports that Flash scored 98.7% on the DF26 benchmark, 29 points ahead of the next best detector it compared.
Those are strong figures, but keep two things in mind. First, they are the company's own tests, published by the company. That doesn't make them wrong; it means nobody independent has confirmed them yet. Second, a test set is clean and labelled. The images in your feed have often been screenshotted, cropped, filtered and re-uploaded several times, which is harder for any detector.
Pangram's site invites you to try it for free. Check its pricing page for the current limits before you rely on it for anything important.
How AI image detectors work, and their honest limits
An AI image detector is itself an AI model. It is trained on very large collections of real and AI-made images and learns tiny patterns that separate them, patterns your eyes usually can't see. Pangram describes one smart training approach: take a real photo, ask AI generators to make a close copy of the same scene, and teach the detector to tell the two apart. That way it learns the subtle differences, not just obvious cartoon-style fakes.
Every detector, however good, shares the same limits:
- It gives a probability, not proof. "Likely AI" is not the same as "definitely fake".
- It can be wrong both ways. It can miss a fake, and it can flag a real photo. The second mistake is the dangerous one: publicly calling a real person's photo fake can hurt someone who did nothing wrong.
- It is in a race. New generators appear every few months, and a detector may miss a brand-new one until it is updated.
- There is a grey zone. Many real photos are touched up with AI tools: a stranger removed, the sky extended, the light fixed. A detector may call that AI; a person might call it basically real.
The best way to think of a detector is as a smoke alarm. When it goes off, you check. You don't accuse anyone just because it beeped.
A 5-step check before you share any picture
This routine takes about two minutes and works with or without a detector.
- Pause. Strong emotion is a faker's best tool. If a picture makes you angry, scared or amazed in a second, slow down.
- Find who posted it first. Tap through to the original account. Is it a known outlet, the brand itself, a real person with a history, or an account that appeared last week?
- Run a reverse image search, for example with Google Lens. You will often find the original photo, an older date, or an existing fact check.
- Look for a second source. Real events usually have other photos from other angles and other people. A fake often exists in only one picture. This is the step most people skip.
- Use a detector as a final signal. If it agrees with everything else you found, good. If it disagrees, keep digging before you decide.
If you still can't tell, the safe choice is simple: don't share it, or share it clearly labelled as unconfirmed.
What this means for creators, students and small businesses
Creators: the most important images are your own. If you use AI images or AI clips in your videos and posts, label them. It keeps your audience's trust, and it is the honest thing to do.
Students: run the five steps on any picture that goes into a project or presentation. A fake image in your slides can cost you marks, and the check takes two minutes.
Small businesses: be careful with supplier product photos, reviews, and property or rental photos when you book space. Pangram itself names shopping and rental listings as places where knowing what's real matters.
The bottom line
AI image detectors are getting better, and a heatmap that points to the edited part of a picture is a genuinely useful step. But no detector is a judge. Your best protection is still a routine: pause, find the source, search, look for a second angle, and only then use a detector. Our video on this topic, out on 19 November, walks through the same check in under ten minutes.
Source: Pangram: Introducing Pangram Image 1.0 and Pangram Image Flash 1.0


