On June 17, 2026, Hugging Face switched on a free AI video detector that anyone can use in a browser. The tool scores clips frame by frame and flags the ones that look machine generated. There is no API key, no credit card, and no waitlist. The model behind it shipped the same morning as an open-weight release under an Apache 2.0 license, which means a local newsroom can download the weights and run checks on its own hardware. Hugging Face published the model card and the hosted demo together, and the free endpoint went live at 9:00 a.m. Eastern.
That timing matters. Video detection has been the expensive half of the deepfake problem. Text detection is cheap, image detection is cheap, and video detection has mostly lived behind enterprise contracts. Edit teams at broadcasters paid per minute. Trust and safety teams paid per hour of uploaded content. Everyone else guessed. Now the cheap version is free, and it is not a stripped-down demo. The model handles MP4, MOV, and WebM files up to 200 MB and 90 seconds long, and it returns a per-frame probability score plus a clip-level verdict. Compare that with the open-weight video generators that made this problem urgent, such as the LTX 2.3 release, and the asymmetry is obvious.
The audience is wider than journalism. Teachers checking student submissions. Insurance adjusters reviewing crash footage. Court clerks logging evidence. Dating app moderators. Local election officials in countries with votes this fall. All of them needed a cheap way to ask one question. Was this clip generated? Before this week, the honest answer was often that you should ask a vendor and pay. The hosted free tier is the first credible zero-cost option at scale, though it arrives in a year when free tiers across the industry have been shrinking, a pattern we covered in our look at free AI pricing changes in June 2026.
Competitive context explains the odd generosity. Google moved its own SynthID video detection tool behind paid tiers on June 10, 2026, at $0.004 per analyzed frame. OpenAI still has no public detector for Sora output and leans on C2PA metadata instead. Hugging Face does not sell ads on a detector demo, so the free endpoint is partly a statement. If detection stays proprietary and metered, verification becomes a luxury good. Publishing the weights undercuts that. The company gets something in return, which is a large stream of labeled video that improves the next version.
| Tool | Cost | Free Limit | Open Weights | Max Clip Length |
|---|---|---|---|---|
| Hugging Face video detector | $12/mo Pro, $0.0009 per frame enterprise | 200 clips/day, 2,000 signed in | Yes, Apache 2.0 | 90 seconds |
| Google SynthID video check | $0.004 per frame | None for video | No | Not published |
| OpenAI Sora provenance | Not sold as a detector | C2PA metadata only | No | Not applicable |
| Self-hosted open-weight check | Hardware cost only | Unlimited by license | Yes | Depends on hardware |
What the free AI video detector actually does
The detector is a 2.4 billion parameter model built on a video encoder with a temporal attention head. In plain terms, it watches a clip the way a person does, tracking how light, shadow, and edges behave from frame to frame. Generated video tends to hold still in ways real footage does not. Skin pores vanish between frames. Background text warps when the camera moves. The head catches those inconsistencies and turns each one into a number between 0 and 1. A clip-level score is the weighted average, and the demo overlays a heatmap on the frames that pushed the score up.
Output format is the useful part. You get a JSON payload with a per-frame array, a clip score, and a short list of the frames the model considered most suspicious. A clip score above 0.80 reads as likely synthetic. Between 0.45 and 0.80 the model is unsure, which is the honest place to sit on compressed footage. Below 0.45 it reads as likely real. The demo caps input at 90 seconds and 200 MB, accepts MP4, MOV, and WebM, and processes roughly 40 seconds of 1080p video per GPU second on the hosted endpoint.
Because the weights are open, the hosted demo is not the only path. The quantized checkpoint is about 1.9 GB and runs on a single 16 GB GPU, which puts it in reach of a laptop with a modern discrete card or a small cloud instance. Nonprofits and newsrooms with privacy constraints can run clips locally and never upload them. That is a real difference from every commercial detector on the market, all of which require you to hand over the footage. Our open source model roundup tracks the licenses that make this kind of self-hosting legal for commercial work.
- Clip score above 0.80: likely synthetic
- Clip score 0.45 to 0.80: uncertain, human review recommended
- Clip score below 0.45: likely real
- Input limits: 90 seconds, 200 MB, MP4, MOV, or WebM
The free tier limits, in plain numbers
Free has a ceiling, and Hugging Face published the numbers rather than burying them. Anonymous visitors get 30 checks per hour and 200 clips per day. Signed-in accounts get 300 checks per hour and 2,000 clips per day. The hourly counter runs on a rolling 60-minute window, and the daily counter resets at midnight UTC. Exceeding either limit returns a 429 error with a retry-after header. There is no queue system and no priority lane for free users, so peak-hour latency can stretch from 1 second to 12.
Above the free ceiling, a hosted Pro plan runs $12 per month for 20,000 clips and access to priority GPUs. Enterprise pricing starts at $0.0009 per frame with volume discounts and includes an on-premise deployment option. That is roughly one quarter of Google’s rate for video watermark checks, which sits at $0.004 per frame as of June 10, 2026.
The structure follows the pattern set across the industry this spring, where free allowances got thinner while paid tiers got more granular. We documented the shift in free tier limits getting tougher in June 2026, and this launch is a partial exception. The cap is real, but the floor is also real. You can run 200 video checks a day at zero cost, indefinitely, without talking to a sales team.
Two hundred checks is not nothing. A metro newsroom publishes maybe 40 original videos a week and receives a few hundred tip submissions. The free tier covers the obvious cases. It does not cover a platform scanning user uploads at scale.
Why video detection became a pricing fight
Detection is a cost center, and vendors treat it like one. Running an inference pass over 90 seconds of video burns real GPU time. Multiply that by every upload on a social platform and the bill gets ugly fast. Google’s answer, published in its developer documentation, was to move SynthID video detection into paid tiers on June 10, 2026. Frame-level watermark checks now bill at $0.004 per frame, which works out to $8.64 for a single 90-second clip at 24 frames per second.
OpenAI took a different route and largely skipped the problem. Its Sora output carries C2PA provenance metadata, and the company points users to that metadata rather than shipping a public detector. Metadata helps when it survives. It usually does not. Every major social platform re-encodes uploads, and re-encoding strips most provenance tags. A screen recording of a video strips them completely. That leaves platforms, newsrooms, and courts holding a file with no chain of custody.
The broader price war is the backdrop here. When one vendor gives away detection, the metered ones start to look extractive. Free detection also creates pressure on the reverse side of the market, the generation side, where API prices have been falling for a year. Our analysis of the AI price war and what it means for developers makes the same point from the other direction. Cheap verification makes cheap generation less useful as a differentiator.
How accurate is it, and where does it break
Model card numbers are decent but not clean. The detector hits 91.4% accuracy on the FaceForensics++ benchmark, 84.2% on lightly compressed video, and 63.1% on clips that were re-encoded twice at low bitrate. That last figure is the one that matters. Most viral video arrives after two or three trips through a platform transcoder, which is exactly where detection gets weakest. Raw benchmark performance and real-world performance sit miles apart.
False positives are the other problem. High-motion sports footage, heavy film grain, and old VHS transfers all look statistically odd to a model trained on clean sources. The model card flags this and recommends a human review step for any score between 0.45 and 0.80. It also notes that training data skewed toward clips under 30 seconds, so confidence drops on longer inputs.
Independent research says the same thing. Stanford HAI has tracked a widening gap between detection accuracy on benchmark data and accuracy on compressed, in-the-wild media, and the 2026 AI Index data shows that gap has not closed. Detection is a signal, not a verdict. Anyone treating a 0.91 score as proof will get burned badly in court.
What newsrooms and moderators are doing with it
Verification desks started testing the endpoint within hours. Reuters Institute survey data has shown for several years that verifying user-generated video is among the hardest and least resourced tasks in digital news. The 2026 numbers are no kinder. Most desk editors have no dedicated detection budget at all. A free tool that returns a documented score in seconds changes the workflow more than it changes the accuracy, because it turns a judgment call into a logged data point.
The practical pattern emerging is two-person verification. One editor runs the clip and records the timestamp, the score, and the model version. A second editor reviews the heatmap and looks for tells a human catches, such as mismatched audio, impossible reflections, or background text that shifts between frames. If the two disagree, the clip gets held. That process costs almost nothing now, which is the whole point.
Moderation teams face a harder trade-off. Automated takedowns based on a detector score are legally risky in the EU and increasingly in the US, and the false positive rate on compressed footage is too high to act on alone. The realistic use is triage. Score everything, route the top decile to humans, and leave the rest. Our AI in journalism statistics page tracks how quickly that kind of tiered review has spread through newsrooms.
The catch: watermark removal and the arms race
Detection has a structural problem that no model release fixes. Detectors are trained on the generators that existed when the training set was built. Generators keep shipping. An open-weight video model released in March 2026 produces footage the June detector has never seen, and the accuracy drop on unseen generators usually runs 15 to 25 points. The model card admits this in a limitations section that is longer than the results section.
Then there is adversarial stripping. Adding imperceptible noise to a generated clip before upload drops detection accuracy sharply, and the code to do it is public. Metadata is even easier to remove. Anyone who re-exports a clip through a free editor wipes C2PA tags by accident. The result is a verification arms race where the defense has to be right every time and the offense only has to work once.
Free detection does not win that race. It changes who can play. A student with a laptop can now run the same class of check that used to cost $8.64 per clip. That is the real shift, and it lines up with a broader move where flagship capabilities get pushed to paid tiers while the free floor gets a thinner but genuinely useful version, a trend we tracked in major AI model tier changes.
What to watch next
Hugging Face staff have said a full-length version is in training and will handle clips up to 10 minutes and 4K resolution. An API with the same weights, plus batch endpoints, is on the roadmap for the third quarter of 2026. The open weights will stay Apache 2.0 according to the model card, but the hosted endpoint is the piece most likely to change. Free inference tiers rarely stay free once GPU costs show up on a balance sheet.
Regulation is the other pressure point. The EU AI Act transparency obligations for synthetic media take effect on August 2, 2026, and they put the burden on generators to label output rather than on detectors to catch it. That is the right architecture, though it does nothing for the archive of unlabeled video already online. Expect vendors to pitch detection tools as compliance products in the months before that deadline.
For anyone using the tool today, three habits matter. Record the model version with every score, because results shift between releases. Keep the original file, because re-encoding destroys evidence. And treat every score as one input among several. A detector that says 0.91 is telling you where to look, not what to conclude.
Free AI News may earn a commission if you sign up for a paid plan through links on this site. This does not affect our reporting.
Frequently Asked Questions
What is an AI video detector?
An AI video detector is a model that analyzes a video file frame by frame and estimates the probability that the footage was generated or altered by AI. It looks for temporal inconsistencies, such as flickering textures, warped background text, and lighting that does not behave the way a real camera records it. Modern detectors return a per-frame score plus a single clip-level verdict.
Is the Hugging Face AI video detector really free?
Yes. The hosted demo requires no API key, no credit card, and no waitlist. Anonymous visitors get 200 clips per day, and accounts get 2,000. There is a paid Pro plan at $12 per month for 20,000 clips with priority GPUs, but the free tier is permanent rather than a trial.
How accurate is AI video detection in 2026?
The new model card reports 91.4% accuracy on the FaceForensics++ benchmark and 84.2% on lightly compressed video. Accuracy falls to 63.1% on clips that were re-encoded twice at low bitrate. That last figure matters most, because most viral video travels through several platform transcoders before anyone sees it.
Can AI video detectors be fooled?
Yes, and easily. Adding imperceptible adversarial noise before upload drops detection accuracy sharply, and the code to do it is public. Detectors are also trained on older generators, so clips produced by a video model released weeks earlier can slip through. Provenance metadata is even weaker, since re-exporting a clip through any editor usually strips it.
Do I need a GPU to run the detector myself?
You need a GPU to run it at useful speed. The quantized checkpoint is about 1.9 GB and fits on a single 16 GB card, which covers a modern desktop GPU or a small cloud instance. CPU inference works but is slow enough that a 90-second clip can take several minutes.
Will the free tier change later?
The open weights will not, because Apache 2.0 is irrevocable once published. The hosted endpoint is the fragile part. Free inference tiers across the industry have been shrinking through 2026, and a detector that costs GPU time per request is a likely candidate for tighter caps.
What Should You Remember?
- Free access: Hugging Face opened a hosted AI video detector on June 17, 2026, with no API key or payment method required.
- Daily caps: Anonymous users get 200 clips per day and 30 per hour; signed-in users get 2,000 per day and 300 per hour.
- Open weights: The 2.4 billion parameter model ships under Apache 2.0 and runs on a single 16 GB GPU.
- Accuracy caveat: Detection falls to 63.1% on twice re-encoded clips, the form most viral video takes.
- Price contrast: Google bills SynthID video checks at $0.004 per frame, roughly $8.64 for a 90-second clip.
- Workflow shift: Newsrooms can now log a detection score for free, turning a judgment call into documented evidence.
- Watch the tier: Hosted free limits are the piece most likely to tighten once inference costs land on the balance sheet.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
