Nvidia has released a new artificial intelligence tool designed to help newsrooms identify whether videos are synthetically generated or captured from real events, addressing growing concerns about deepfakes undermining trust in video journalism. The Synthetic Video Detector, unveiled at Siggraph (a major graphics and computer technology conference), analyzes video frame-by-frame to produce a likelihood score indicating whether footage contains synthetic content.

Why Does Detecting Synthetic Video Matter for News Organizations?

As AI video generation technology has become increasingly sophisticated, journalists and broadcasters face a critical challenge: distinguishing authentic footage from convincing artificial recreations. Video remains the primary medium through which audiences witness breaking news and major events in real time, making the integrity of video content essential to public trust.

"If we cannot tell the difference between a synthetic video and a real one, it can erode public trust when videos are presented as news, as something that came from the real world," explained Rev Lebaredian, vice president of physical AI simulation technology at Nvidia.

Rev Lebaredian, Vice President of Physical AI Simulation Technology at Nvidia

The tool addresses this tension by leveraging the same AI technologies used to generate synthetic videos and applying them in reverse to detect when footage has been artificially created. Lebaredian noted that this dual capability offers a practical solution: "Fortunately, the same AI technologies we're developing to generate these videos can also be used to detect when a video is real or synthetic".

Lebaredian

How Does the Detection Tool Actually Work?

The Synthetic Video Detector operates by analyzing submitted video on a frame-by-frame basis, using dual vision-transformer models called DINOv2 and DINOv3 to examine individual frames. The system produces frame-level classification scores that are then averaged and combined into an overall video-level synthetic-likelihood score.

Editorial teams can use this score to prioritize which clips require closer review, flag or quarantine questionable footage, or escalate specific videos for deeper investigation. Importantly, Nvidia emphasizes that the tool is intended to supplement, rather than replace, existing journalistic verification practices. The microservice is specifically designed to remain effective even after videos undergo common newsroom and social media processing steps, including compression, resizing, cropping, and re-encoding, all of which can otherwise degrade a detection system's accuracy.

What Are the Performance Metrics and Technical Capabilities?

In Nvidia's internal testing, the detection model achieved impressive accuracy rates depending on video compression levels. The system reached 92% accuracy when analyzing uncompressed video, 87% accuracy at 15% compression, and 82% accuracy at 50% compression, reflecting the added difficulty of identifying synthetic content in more heavily processed footage typical of real-world social media distribution.

The system's processing speed is also noteworthy. It can process 1080p video in as little as 22 milliseconds on Nvidia RTX systems and approximately 30 milliseconds on Nvidia L40 GPUs, meaning the analysis happens nearly instantaneously. The latest revision of the underlying model achieved an area-under-the-curve (AUC) score of 0.9614 and an accuracy rating of 0.9453 on Nvidia's internal test set. The AUC metric measures how effectively a classifier system ranks genuinely synthetic content above authentic footage, independent of any specific detection threshold.

How Can News Organizations Deploy This Technology?

Nvidia designed the microservice to be deployable in a range of environments depending on an organization's specific security and operational needs. These deployment options include:

• On-Premises Systems: Organizations can run the detector on their own internal servers to maintain complete control over video data and processing.
• Edge Computing Setups: The tool can be deployed closer to where video is initially captured and streamed, allowing real-time detection.
• Hybrid Infrastructure: Organizations can combine on-premises and cloud-based deployment for flexibility and scalability.
• Air-Gapped Environments: The system can operate in isolated networks that remain disconnected from external systems, meeting strict regulatory requirements around data residency and security.

This flexibility is particularly significant for organizations most exposed to synthetic media risks, including broadcasters, government agencies, financial institutions, and operators of critical infrastructure, which often face strict regulatory requirements around data protection and operational control.

Who Is Already Using This Technology?

Streaming infrastructure company Wowza has already begun integrating the Synthetic Video Detector through its Wowza Video Intelligence Framework, bringing real-time synthetic video detection capabilities into livestreaming workflows. This integration is particularly significant given that Wowza's platform spans more than 35,000 deployments across over 170 countries.

By pairing the detection tool with video infrastructure that customers already use, Wowza aims to make AI-assisted verification available at the point where video is initially captured and streamed. This allows organizations to flag potentially synthetic content in real time while keeping sensitive footage contained within their own secured environments.

What Are the Limitations of This Detection Tool?

Nvidia was clear in clarifying the specific scope and limitations of the detection service. The system does not evaluate whether the events depicted in a video are factually accurate or true. Instead, it strictly assesses whether the video itself was synthetically generated. Newsrooms should treat the resulting scores as a triage and escalation tool rather than a definitive final judgment, emphasizing that human editorial oversight and established verification practices remain essential components of the overall process.

The model was trained and evaluated using a combination of real and synthetic video content, including footage generated through diffusion-based AI methods. As AI video generation technology continues to advance in realism and accessibility, tools like Nvidia's detector are expected to become an increasingly important, though not standalone, part of how newsrooms work to preserve public trust in the video content they publish and distribute.