Pangram, the company behind one of the most widely used AI-text detectors, has just moved into video and images. Its new tools flag AI-generated and AI-edited clips, show a heatmap of the suspected regions, and scan photos as you browse. What began as a writing-integrity tool is becoming something closer to universal media forensics.
Why did a text detector get into the video business now, and who actually benefits? Here is what Pangram shipped, what its numbers do and do not prove, and where AI detection goes from here.
Why Pangram Moved Beyond Text
Pangram built its reputation on a single question: was this paragraph written by a human or a model? On November 5, 2026, the company answered a much larger question — was this video, this photo, any of it made by a machine — and answered it with tools aimed at the media people actually fight about now.
Two things shipped at once. First, video detection: subscribers upload a clip and receive an assessment of whether AI was used to create or edit it, with the model distinguishing between partially AI clips and fully AI-generated ones. Second, the company’s Image 1.0 model escaped its research preview and became a general feature, including inside the company’s Chrome extension, which now scans images automatically as you browse and flags ones likely to be synthetic.
Text detection was always a proxy war. The internet’s real arguments about authenticity happen in clips and photos — deepfake ads, synthetic travel vlogs, fabricated product photography, AI news footage. The text business proved the demand; the media landscape proved that text was the smallest slice of the problem. Frontier video models made that obvious to everyone who ships moderation tooling, and the company decided to be first with a consumer-accessible answer.
Video and image detection are available now to existing Pangram subscribers, with no separate product to buy — detection breadth becomes part of the subscription rather than an upsell. Free accounts can check a handful of images without paying, which is a deliberate trust play: let skeptics run their own phone photos through the tool and discover the false-positive edges before an enterprise deal depends on them.
That free tier is also a data strategy. Every skeptic who uploads a real photo sharpens the company’s understanding of where its confidence should drop — a feedback loop that pure enterprise vendors never see. You can try the detector yourself at pangram.com before believing any benchmark slide, including the ones below.
How The Video Detection Actually Works
Pangram’s pitch for video detection rests on three design choices: give a verdict per clip, admit that most clips are mixed, and show your work with a heatmap. Each one answers a specific failure of the text-detector era, where confident one-word labels did more damage than no label at all.
The basic flow is deliberately simple: upload a video, get an assessment of whether AI was used to create or edit the clip. No configuration, no confidence dials, no report writing. The company claims the underlying model “particularly excels at frontier AI video detection” — its language for clips rendered by the current generation of high-end video models rather than the older, sloppier ones. In one internal test the company detailed, the video model flagged AI content with more than 98 percent accuracy.
The more interesting capability is the middle category. Pangram says it can distinguish clips that are partially AI — real footage with synthetic insertions, background replacements, or object removals — from clips that are entirely generated. That is the category most moderation systems refuse to touch, because most of the internet’s edited media is legitimately touched by software. A tool that cannot separate a sky replacement from a full generation is useless to anyone policing real platforms.
The heatmap flags which regions of the video are likely AI, turning a binary verdict into an inspectable one. This matters more for trust than the accuracy number does: a reviewer can disagree with a score, but a highlighted region either contains the suspicious artifact or it does not. Heatmap output is how detection earns the right to be used in consequential decisions — a label alone gets appealed, a marked region gets checked.
| Modality | Output | Claimed accuracy | Main caveat |
|---|---|---|---|
| Text | Probability score | Long-standing core product | Struggles with short, edited, or translated text |
| Images | Flag plus browser scan | 99%+ at 0.27% false positives | Reduced confidence below 512×512 |
| Video | Verdict plus heatmap | 98%+ in one internal test | Internal benchmark, not third-party audited |
Inside The Image 1.0 Model
The image side of the launch is where Pangram made its most aggressive numeric claim — and where the honest fine print is most revealing about how this class of tool actually behaves in the wild.
Pangram Image 1.0, previously a research preview, is now generally available. The company says it flags more than 99 percent of AI images with a false positive rate of 0.27 percent in its testing. Those are extraordinary numbers, and they are exactly the kind of numbers that get quoted without their footnotes: they describe the company’s own test set, built from the generators its team chose to test against, measured at the resolutions that team chose.
The most consequential distribution move is not the model but the browser. The extension now scans images automatically as you browse, flagging ones likely to be synthetic. That shifts detection from a deliberate act — upload, wait, read verdict — to ambient background awareness. CEO Max Spero demonstrated it on Pinterest’s website, and the demo’s imperfections became the most instructive part of the launch.
The launch post states that for images below 512 by 512 pixels, “results will come with a reduced confidence disclaimer” — and the Pinterest demo showed why. The extension returned an “unsure” result on two images that are obviously not genuine photos. Low-resolution, heavily compressed, already-synthetic imagery is precisely where detectors hedge, which is the right behavior but a poor fit for platforms whose slop is exactly that: small, blurry, recycled JPEGs passed through five apps.
Small images are not an edge case in slop economics — they are the medium’s natural habitat. Thumbnails, reposts, screenshots of screenshots. If the company wants to police the actual slop ecosystem rather than the clean test set, the sub-512 population has to be its home turf, and “unsure” needs to be a graded output rather than a shrug. Our breakdown of how today’s detectors handle messy real-world input applies almost unchanged to pixels.
The Smartphone Camera Problem
The hardest problem in image detection is not detecting images from generators. It is refusing to flag images from cameras — and cameras themselves have become generative.
“Most phone cameras, including iPhones, do some level of AI postprocessing to add detail to real photos,” Spero explained when describing the build of the image model. Multi-frame stacking, computational HDR, neural sharpening, magic eraser edits — the average smartphone photo is a collaboration between a lens and a small generative pipeline. Any detector trained naively on “real versus AI” learns that this pipeline leaves synthetic fingerprints, and starts flagging pocket cameras as image generators.
Pangram reported that it delayed the image model’s release specifically to test against several iPhone 18 Pros, because early versions of the model flagged the phones’ camera output as AI. That anecdote should be filed under encouraging: a detector vendor that delays a launch to fix false positives is behaving like a firm that knows false positives are what kill enterprise adoption. It is also a warning about the pace of the problem — every new phone generation rewrites what “real” looks like at the pixel level.
Pangram’s stated distinction matters as much as its accuracy: understanding the difference between AI edits and fully AI-generated visuals. Removing a stranger from a family photo and generating the family from scratch are different acts with different ethics and different policies. A detector that collapses both into “AI: yes” creates the same blunt-instrument verdicts that got text detectors laughed out of courtrooms and newsrooms in 2023.
The deeper issue is that phone makers and generative model makers are converging on the same pixels from opposite directions. Detectors must stand in the middle and grade intent. We have covered this from the creator side in our piece on why detection is harder than “real or fake”, where Pangram’s own CEO argued the binary question was the wrong question — words he has now bet the company on answering anyway.
The AI Slop Economy Behind The Launch
Detection tools do not exist in a vacuum. They exist because synthetic media became an economy — production costs collapsed, distribution stayed free, and the resulting flood now has a name: AI slop.
The economics are simple to state. A generation of video that once required a crew, a location, and a week now costs a prompt and a minute of GPU. Content that could not justify its own production cost used to simply not exist; now it exists in the millions, monetized by play count on platforms that pay for engagement regardless of authenticity. The slop is not just low quality — it is low quality with a business model, which is far more durable than low quality alone.
The expansion is best understood as infrastructure building. A subscription that covers text, images, and video becomes a single procurement line for platforms, marketplaces, and publishers — one vendor, one API budget, one trust story. That bundling is the actual product announcement hiding inside the feature announcement: Pangram is no longer selling verdicts on paragraphs. It is selling an authenticity layer — and those get bought at platform scale or not at all.
The buyer fears floods, not individual fakes. No marketplace can halt its catalog over one suspicious upload, but every marketplace can measure what share of this week’s submissions tripped the detector and act on the curve. Detection sold as flood control changes the conversation from forensic disputes about single images to risk management about populations — which is exactly the register enterprise procurement teams think in, and exactly the register this video and image launch was pitched in.
Platforms that learned this the hard way include the biggest labeler of them all: Instagram has flailed at AI-content labeling repeatedly, proving that an inconsistent label is worse than none. Our earlier coverage of Instagram’s AI label mess is required context for anyone assuming labels and detection are the same product.
Who Actually Benefits From Detection Tools
Who actually pays for the new video and image detection, and who merely benefits? The honest answer is that paying customers and beneficiaries are nearly opposite groups — and that gap explains both the launch’s packaging and its limits.
Ad platforms police synthetic endorsements. Stock marketplaces protect licensed libraries from AI floods. Newsrooms verify user-submitted footage before airing it. Creator platforms paying by engagement want to know whether the engagement was farmed. None of them are buying truth; they are buying defensible process.
| Buyer | Worst failure mode | What they need from the detector |
|---|---|---|
| Ad platforms | Synthetic endorsements at scale | Video verdicts plus heatmaps |
| Stock marketplaces | AI floods licensed libraries | Image detection at upload time |
| Newsrooms | Fabricated footage aired as real | Explainable region flags |
| Creator platforms | Slop farming payout systems | Bulk partial-AI grading |
The Chrome extension makes ordinary browsers the largest unintended beneficiary of this launch. Nobody who installs it is buying an authenticity layer; they are getting a personal lie detector for product photos, news thumbnails, and influencer vacations. That ambient audience matters commercially — consumer tools create public expectations that enterprise contracts eventually have to honor.
The least discussed beneficiary is the seller who runs the detector and ships the product anyway. Detection tools create a documented check: a marketplace that scanned with Pangram Image 1.0 and approved a fake can claim process, while one that never scanned has nothing. That is why detector marketing leans on coverage percentages — the tool’s value to the buyer includes the audit trail it produces.
If you have followed the earlier debates about whether Pangram deserved its gold-standard reputation in text — our long take is progressiverobot.com — the same tension now applies to pixels: reputation is an asset until the first viral false positive, and every buyer’s audit trail becomes the detector’s discovery rule.
Limits, False Positives, And The Arms Race
No detection tool survives contact with its own limitations, and Pangram’s are now public. The company’s launch post is unusually honest about them — but honesty about limits does not neutralize them.
A 0.27 percent false positive rate sounds negligible until you multiply it by platform scale: on ten million uploads a day, that is twenty-seven thousand real photos accused of being fake every morning, before breakfast. Every one of them is a support ticket, a wrongly delisted product, a journalist blocked from publishing. Enterprise buyers will demand their own false-positive measurements on their own distributions, and internal numbers will not survive that audit unchallenged in every vertical.
The launch post’s own concession that sub-512-pixel images come “with a reduced confidence disclaimer” marks the detector’s floor — and the slop ecosystem lives on the floor of that distribution. Compression strips the artifacts detectors rely on; reposts destroy them further. A qualitative map of how confidence degrades explains most future product decisions:
Detectors are adversarial instruments: every public detector invites generators trained to fool it, and this one’s behavior is now public knowledge. Frontier video models improve quarterly, and each generation erodes the artifact gap that detectors like Pangram’s exploit. The strategic question is not who wins today’s snapshot but whether detection can stay a product rather than becoming a patch cycle — and whether buyers keep paying for verdicts that carry a disclaimer in half the input space.
That is the honest frame for the whole category, tools like this one included: detection is worth having and cannot be trusted alone. More on the deeper pattern in our AI models and tools hub.
What To Watch Next
Pangram’s expansion will be judged less by launch-week numbers than by four developments over the next year. The company has repositioned itself from a niche integrity tool to infrastructure for the authenticity economy; infrastructure gets graded on adoption, durability, and damage control.
The first large platform to publish third-party measurements of the video detector — rather than quoting the company’s internal 98 percent — will define the category’s credibility standard. If the detector survives that audit well, verification becomes a procurement checkbox across the industry. If it stumbles, expect every buyer to quietly treat detector scores as one more risk signal rather than a verdict.
Watch how quickly frontier video models produce output the detector flags as real. The arms race moves in quarters, not years, and a single public demonstration that top generators defeat Image 1.0 or the video model resets enterprise timelines. The companies whose models this tool hunts are the same companies with the compute to study it closely.
The browser extension is this launch’s most interesting experiment: detection as ambient consumer utility rather than forensic service. If millions of users start seeing AI flags while browsing, public expectation shifts — and platforms that have resisted labeling will face a constituency that already sees the labels privately.
The company spent three years proving that text could be graded, and three months proving that the market wanted the whole media stack graded at once. The video and image launch is credible, well-paced, and honestly caveated — which makes it the most serious attempt yet to make authenticity a purchasable commodity. Whether the authenticity economy pays for it is next year’s story.
One quiet thing to watch: whether the company begins publishing its adversarial test sets. A detector that releases its bait is either confident or cornered, and the difference tells you everything about where this category lands.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
