Global AI, Malaysia Angle / AI news for Malaysia
NVIDIA's synthetic video detector moves into newsrooms and live streams
Three media vendors are building NVIDIA's AI-video detector into tools newsrooms and streamers already use. It answers one narrow question, and the clips Malaysians argue about usually fail a different test.
In brief
- NVIDIA said on 9 September 2026 that it had expanded its AI for Media toolkit ahead of the IBC conference in Amsterdam, with the Synthetic Video Detector as the authenticity component.[1]
- NVIDIA reports the detector now reaches 99.3% accuracy on text-to-video content and 97.7% on image-to-video, with the largest gains on difficult image-to-video cases. It publishes no test set or methodology.[1]
- Dalet, TwelveLabs and Wowza are each shipping it into working products, moving detection from demonstration into newsroom verification, compliance screening and live streaming.[1]
Detection stops being a demo and becomes a product feature
The question of whether a video is real has moved from research demo to purchasing decision. NVIDIA said ahead of the IBC conference in Amsterdam that three media technology companies are building its Synthetic Video Detector into products newsrooms, compliance teams and streaming providers already use.[1]
The detector is a NIM microservice, first shown at SIGGRAPH earlier this year. NVIDIA is careful about what it claims. It assesses the probability that footage is AI-generated, and calls it another point of analysis in a review process rather than a verdict.[1][2]
What NVIDIA announced at IBC
IBC 2026 ran from 11 to 14 September in Amsterdam, and NVIDIA says the show drew more than 44,000 attendees from over 170 countries. The company announced what it calls a major expansion of NVIDIA AI for Media, a set of GPU-accelerated software kits, microservices and blueprints.[1]
Most of that expansion makes pictures better or cheaper: frame generation for slow-motion replays, super resolution, high dynamic range conversion, lip-synced dubbing. The authenticity work sits apart, because it is the only part aimed at telling a viewer what a picture is rather than improving how it looks.[1]
The accuracy numbers, and what they do not cover
NVIDIA reports that since release the detector has reached 99.3% accuracy on text-to-video content and 97.7% on image-to-video, with especially large gains on the harder image-to-video cases. Those are the two ways synthetic clips are made: from a written prompt, or by animating a still photograph.[1]
What NVIDIA does not publish matters just as much. No test set, methodology, false-positive rate or independent benchmark is attached to either figure. Nor does it say how the detector handles footage that has been re-encoded, screen-recorded or passed through a messaging app, which is how most people receive a suspicious clip.[1]
Three vendors are putting it in front of working teams
Dalet is building the detector into a cloud-hosted verification workflow for news organisations, so editorial staff submit footage and read the scores and metadata inside a Dalet interface. TwelveLabs has made Compliance by TwelveLabs generally available, screening content against regional standards, with the detector adding frame-level authenticity signals.[1]
Wowza takes it furthest from the edit suite. Its Streaming Engine powers more than 35,000 video deployments in over 170 countries, and it will distribute the detector through its Video Intelligence Framework to analyse live feeds in real time, on premises, at the edge, in the cloud or fully air-gapped.[1]
The Malaysian case this week fails a different test
MCMC said on 19 September that it had received complaints about a stage performance video circulating on social media. Preliminary checks found the recording came from a 2018 performance now being spread again. The commission has identified accounts and individuals and is working with police investigators on digital evidence.[3]
Run that clip through a synthetic video detector and it would come back authentic, because it is. The dispute is about when it was filmed and what it is claimed to show. Detection answers whether a machine made the pictures. It does not answer whether the caption is true, and that second question causes most harm here.[3][1]
Why Malaysia should care
Malaysia argues about viral video constantly, and MCMC is investigating one now. That case shows the limit of this technology precisely: the recording is real, so a detector built to spot AI generation would pass it. The harm came from stripping a 2018 performance of its context.
Malaysian broadcasters and news desks
Authenticity scoring is arriving inside the media asset tools you already license, rather than as a separate purchase.[1]
Practical move: Ask your vendor whether the detector is included, what it costs, and who is authorised to overrule its score.
Compliance and platform trust teams
Frame-level authenticity signals can now sit beside existing regional compliance checks, but a probability is evidence, not a finding.[1]
Practical move: Write down the score threshold that triggers a human review before you deploy anything.
Ordinary Malaysian viewers and forwarders
Most clips that cause trouble here are real recordings shown out of context, which no AI detector is designed to catch.[3]
Practical move: Check when and where a video was filmed before forwarding it, and wait for the official statement.
What Malaysians can do now
- Treat an authenticity score as one input, never as a decision, and record who may override it.
- Before forwarding a viral clip, establish its original date and setting rather than its realism.
- Ask any vendor selling detection for its false-positive rate on re-encoded and cropped footage.
What we still do not know
NVIDIA has published results without publishing the test
- No test set, methodology, false-positive rate or independent benchmark supports the 99.3% and 97.7% figures.
- NVIDIA names no price, no availability date for Dalet's workflow and no Malaysian broadcaster or partner.
- Nothing published says how the detector performs on Bahasa Malaysia content or on heavily recompressed clips.
Sources
- 1.NVIDIA Brings Real-Time AI to Broadcast, Sports and Global Streaming at IBC NVIDIA, 9 September 2026
- 2.synthetic-video-detector Model by NVIDIA NVIDIA
- 3.Siasatan Berhubung Video Pementasan Yang Tular Di Media Sosial Malaysian Communications and Multimedia Commission, 19 September 2026
- 4.File:NBC News Control Room Show Logos (50245773423).jpg Wikimedia Commons, 17 September 2012
- 5.File:Television Outside Broadcast Unit (8724322710).jpg Wikimedia Commons, 9 May 2013
- 6.File:Angkasapuri, Kuala Lumpur.jpg Wikimedia Commons, 17 March 2007
- 7.File:KL subway commuters - Flickr - Franck Michel.jpg Wikimedia Commons, 8 May 2024
