Pangram released two new detection models on November 4, 2026: Pangram Image 1.0 for AI-generated images and Pangram Image Flash 1.0 for AI-generated video. The company calls both "state-of-the-art." It reports 99.96% accuracy on its own image test set and 98.7% on the academic DF26 video benchmark.
Those numbers come from Pangram's launch post. They are vendor-reported. This post separates what Pangram claims from what you can check yourself, and explains how to test the tool before you rely on it.
TL;DR: claims, evidence and open questions
| Question | Short answer |
|---|---|
| What is new? | Image 1.0 (images, heatmaps), Image Flash 1.0 (video), native video upload, Chrome extension support, API for both |
| Who made the accuracy claims? | Pangram, in its own launch post |
| What is the image claim? | 99.96% accuracy on 1,200 synthetic images from 12 generators, mirrored from 100 real photos |
| What is the video claim? | 98.7% on DF26, "a 29-point leap over the next best detector" |
| Is it independently reproduced? | Not that we could find on launch day |
| What changed from the research preview? | Minimum image size drops from 512x512 to 256x256; heatmaps; video |
| Does a high score prove an image is fake? | No. It is an estimate, not proof |
What did Pangram actually release?
Pangram is best known for text detection. Earlier this year it released an image detector as a research preview. The November 4 launch turns that into a product line with two models.
Pangram Image 1.0 is the flagship image model. Pangram lists four selling points: performance on frontier image generators, a heatmap that shows where in a picture the AI content sits, robustness to edits, compression and low resolution, and a lower minimum input size. The launch post shows a real photograph that contains an AI-made region, and the heatmap marks that region.
Pangram Image Flash 1.0 is a smaller, faster model aimed at video. Pangram says it analyzes uploaded video frame by frame and reports which parts come back as AI. Native video upload is now part of the Pangram app.
Two product surfaces round out the release. The first is a Chrome extension that flags AI content as you scroll social media, shopping pages and rental listings. The second is an updated API that accepts images and video.
How does Pangram test its image model?
The headline image figure rests on what Pangram calls a synthetic mirror design. The company took 100 real photos and used 12 frontier image generators to create a matching fake of each photo, with prompts written to match the original closely. That gives 1,200 synthetic images. Pangram reports 99.96% accuracy on this set.
The design has a real strength. Because each fake shares a scene with a real original, the detector cannot win by learning that a certain topic or style tends to be fake. It has to find signals in the pixels.
It also has limits you should keep in mind:
- Small base set. One hundred real photos is a small sample. The 1,200 images are many views of the same 100 scenes.
- Pangram chose the generators and prompts. The post does not list every generator in the text we could read, so we cannot tell which tools are in the mix.
- Accuracy needs a threshold. A single accuracy figure hides the trade-off between false positives (real photos called fake) and false negatives. The post does not give a false positive rate for images in the text we read. For comparison, Pangram quoted a one-in-ten-thousand false positive rate for its text labeler, covered in our Pangram Gmail labeler post.
Pangram also reports results on public academic image sets and links them: X-AIGD, Mirage-Test, AIGenImages2026, a Synthbuster plus RAISE comparison, and Community Forensics. The launch post shows the results as charts. We did not re-run any of them.
What is the DF26 video claim?
DF26 is an academic benchmark described in the arXiv paper DF26: We Cannot Tell Fake From Real Anymore. Per the paper's abstract, human performance at spotting AI-generated videos of people talking is close to chance, and state-of-the-art deepfake detectors are not much better. A summary of the paper reports human accuracy on the fake clips of 52.6%, and the strongest detector tested at 69.7% AUROC. The dataset holds 271 real and 2,420 synthetic single-person speaking clips from seven video models. The Hugging Face listing says access is controlled and needs a request.
Against that backdrop, Pangram says Image Flash 1.0 "scores 98.7%" on DF26, "a 29-point leap over the next best detector." Two cautions apply:
- The metric may differ. The paper reports AUROC for detectors and accuracy for humans. Pangram's post says "scores 98.7%" without naming the metric in the text we read. An accuracy figure and an AUROC figure are not the same thing, and the gap between 98.7% and a 69.7% AUROC may not be a like-for-like comparison.
- Pangram ran its own model. The paper's authors did not test Image Flash 1.0. Whether Pangram used the paper's exact split and protocol is not stated in the launch post.
This is the kind of claim that is plausible and also unverified. A jump from near chance to 98.7% on a benchmark built to defeat detectors would be a major result. It deserves a third-party replication before anyone treats it as settled.
The second video test: 129 real clips, 1,529 fakes
Pangram ran an industry comparison on video too. It took 129 real social media videos and generated AI mirrors with 14 frontier video generators, 100 videos each, sampled from the 129 originals. Each generator got the first real frame plus a text description of what happens in the clip. The total is 1,529 videos. Pangram then compared its model against "multiple major AI video detection services."
The post names the test but, in the text we could read, does not name the competing services. The comparison appears as a chart. Without names, versions and settings, readers cannot reproduce the comparison. That is the main gap to close if you plan to buy on the strength of it.
Pangram also lists academic video results on AIGVDBench, DF26 and Ivy-Fake, claiming state of the art on several. Again, these are its own runs.
Robustness: what the post says and what it leaves out
Social platforms compress and downscale images. Bad actors can add blur or noise to push a detector toward "real." Pangram says it trained for these transformations and found the model robust. It also lowered the minimum input from 512x512 to 256x256. Below 512 pixels, results come with a reduced-confidence disclaimer. That is an honest note: small thumbnails remain harder.
What the post does not show in the text we read is a table of accuracy by edit type, such as JPEG quality, crop, screenshot or re-encode. If you work with screenshots of screenshots, test your own samples.
How does this fit the wider detection picture?
Detection is one of three approaches to the "is this real?" question. The others are watermarking and provenance.
| Approach | How it works | Strength | Weakness |
|---|---|---|---|
| Detection (Pangram) | A classifier estimates whether pixels look AI-made | Works on content with no label | Statistical; can be wrong; arms race with generators |
| Watermarking | Generator embeds a hidden signal | Strong when present | Only covers cooperating generators; can be stripped |
| Provenance (C2PA) | Signed metadata records how a file was made | Proves a chain of custody | Missing on most files; lost on re-upload |
For the other two approaches, see our guides to Google's SynthID verification site and LinkedIn's content credentials rollout. A sensible workflow uses all three: check for credentials first, a watermark check second, and a detector as the fallback.
Why detection claims deserve care
Detector scores have already caused real disputes. In one case we covered, a novel's AI score became the center of an argument over a literary prize; see the Goncourt and Pangram claimed-vs-verified post. The lesson there applies to images and video: a score is evidence, not a verdict.
Pangram's text tools are already embedded in products. Substack's AI detection uses Pangram. If the image and video models spread the same way, a false positive will carry weight. A real photographer or a real news clip could be flagged.
Real cases show why the stakes are high. A photo contest winner in a video category was disqualified after judges found the entry was AI-generated, as we reported in the Nikon Small World in Motion post. Platforms are also making policy on AI content, such as itch.io deleting AI slop and adding disclosure. Both cases need reliable tools, and both would be hurt by a detector that is confident and wrong.
How to test Pangram Image 1.0 yourself
You can run your own mirror test in an afternoon. Do not rely on the vendor's samples.
- Collect real items. Use photos or clips you took yourself, so you know they are real. Include phone photos, screenshots and compressed social uploads.
- Make fakes. Generate matching images or clips with two or three current generators. Use similar prompts to your real items.
- Degrade both sets. Re-save at lower JPEG quality, crop, resize to 256x256 and 512x512, and re-upload through a messaging app.
- Record every result. Track false positives on real items, not only hits on fakes. One wrongly flagged real photo matters more to you than one more correct hit.
- Compare to a second tool. Run the same files through at least one other detector and a provenance check.
- Check the heatmap. For edited photos, see whether the heatmap points at the region you changed. A heatmap on the wrong area is a warning sign.
If you build products on top of the API, log the confidence value and never auto-block on a single score. If your team uses AI agents that read web images, remember that attackers use fakes to steer agents; AgentBeam, the agent security platform from the explainx.ai team, stops AI agents before they take dangerous actions.
What readers are asking
Does a heatmap prove where the fake is? No. It is the model's attribution, a guide for human review. Treat it like a highlighter, not a forensic report.
Can generators beat the detector next month? Probably, in time. Detection and generation chase each other. Pangram's own earlier text work shows regular re-training. A model released today covers today's generators; the DF26 paper's central point is that quality gains erase old cues quickly.
Why does low resolution matter? Fewer pixels carry fewer traces of how an image was made. That is why Pangram added a reduced-confidence note below 512 pixels.
Should newsrooms and schools adopt it? Use it as a triage tool, not an arbiter. Pair it with source checks and provenance data.
Bottom line
Pangram Image 1.0 and Image Flash 1.0 are a real product release with a readable launch post, an API and a Chrome extension. The accuracy figures of 99.96% on images and 98.7% on DF26 video are Pangram's own and use test sets Pangram built or ran itself. They are promising. They are not independently confirmed.
What would settle it: a third-party test with named generators, a published false positive rate on real media, and a clear statement of the metric used on DF26. Until then, run your own mirror test and keep a human in the loop.
Details reflect Pangram's launch post and the DF26 paper as of November 4, 2026. Product limits and prices can change.
Related reading
- Pangram's Gmail AI labeler and its false positive rate
- Substack's AI detection built on Pangram
- Goncourt, Pangram and a disputed AI score
- Google SynthID verification site guide
- LinkedIn content credentials for AI images
- Nikon Small World in Motion winner disqualified
- itch.io deletes AI slop and adds disclosure