Pangram says its new AI video detector flags 98 percent of generated clips

Pangram, the company behind a widely used tool for detecting AI-generated text, has released new tools for flagging AI-generated images and video — the segment of synthetic media flooding social platforms fastest right now. The launch matters because detection has lagged badly behind generation: video models from major AI labs produce increasingly convincing clips, and platforms, publishers, and fact-checkers have had few reliable ways to sort real footage from generated slop. Pangram is now staking a claim to that job across all three media types: text, image, and video.

By Sophie Lindqvist

What can the new tools actually do?

With the update, Pangram subscribers can upload a video and get an assessment of whether AI was used to create or edit the clip. The company says the tool can distinguish between clips that are "partially AI" and those that are entirely AI-generated. That distinction is the hard part of detection work — a clip that mixes real footage with a generated background is a different problem than a fully synthetic one, and conflating the two has tripped up other detectors.

The video tool also ships with a heatmap feature. It flags which aspects of the video are likely to be AI, rather than delivering a single verdict on the whole file. In practice, that means an analyst can see where in a clip the manipulation likely sits. Takeaway: the tool doesn't just say fake or real — it says where and how much.

How accurate does Pangram claim the detection is?

The numbers come from the company's own testing, which is worth keeping in mind when evaluating them. Pangram says the underlying model behind its video detection "particularly excels at frontier AI video detection" — meaning clips produced by the newest, most capable video generators. In one internal test detailed by the company, the video model flagged AI content with more than 98 percent accuracy. Pangram also claims it was significantly better at picking out AI videos than other detection tools in its testing.

On the image side, the company reports even stronger figures. Its image model flagged more than 99 percent of AI images, with a false positive rate of 0.27 percent. A false positive here means a real photo wrongly labeled as AI — the error type that can damage reputations and get authentic evidence dismissed, which is why that 0.27 percent figure carries as much weight as the 99 percent one. Takeaway: the claimed accuracy is high, but it's self-reported, and the caveats start immediately.

What are the caveats?

Resolution is the first one. The company says that at resolutions below 512 x 512 pixels, "results will come with a reduced confidence disclaimer." That's a meaningful limit, because low-resolution images are exactly what spreads fastest on social platforms. An image shared by Pangram CEO Max Spero — showing the Chrome extension in action on Pinterest's website — returned an "unsure" result on two images that are obviously not genuine photos. So the tool's weakest performance showed up in a demo the CEO himself posted.

Modern phone cameras pose a subtler problem. "Most phone cameras, including iPhones, do some level of AI postprocessing to add detail to real photos," Spero explained in a post on X. "We had to build a model that's robust to this postprocessing while understanding the difference between AI edits and fully AI-generated visuals." This is the core tension in image detection today: nearly every photo consumers take is already machine-touched, so a detector has to separate computational photography from synthetic imagery.

That problem was concrete enough to delay the launch. Spero told Pirate Wires that Pangram held back the release of its image model so it could test it on several iPhone 18 Pros, because the model initially flagged images from the phones' cameras as AI. Getting that wrong at scale — labeling every iPhone photo as fake — would make the tool unusable for exactly the audience that needs it. Takeaway: the hard engineering problem isn't spotting obvious fakes, it's the gray zone where real photos are machine-processed and fake ones are machine-made.

How do you get the tools?

The image model has a history worth noting. Pangram previously released a research preview called Pangram Image 1.0. That model is now available through the company's Chrome extension, which automatically scans images as you browse and flags the ones likely to be AI. The extension turns detection from an upload-and-check workflow into a passive layer over the web — closer to how most people would actually encounter a detector in daily use.

The full set of new AI image and video detection abilities is available now to anyone with a Pangram subscription. Users with a free account can check a handful of images without signing up for a paid plan, which gives the curious a way to test the claims before paying. Takeaway: paid subscription for the full toolkit, limited free checks for images.

Why this matters beyond one company

Pangram built its reputation on text detection, and this expansion tracks where the synthetic-media problem is moving. Image and video generators have improved faster than the tooling to audit them, and the volume of AI "slop" — low-effort synthetic content designed for clicks — has grown accordingly. Detection vendors that can cover text, images, and video in one product position themselves as infrastructure for platforms, newsrooms, and anyone else who needs to trust what they see.

The stakes cut both ways. A detector that catches 98 or 99 percent of generated content raises the cost of passing off fakes as real. But false positives — real photos labeled fake — carry their own cost, and Pangram's own admitted growing pains, from iPhone 18 Pro misfires to "unsure" results on obvious fakes, show how narrow the reliable zone still is. The company is betting that continued model work keeps widening it; the next round of frontier video generators will test that bet.