Realtime AI News
Doubao denies Hengdian extras data fed ByteDance models
Doubao public relations lead Liu Xing said on October 28 that claims Hengdian and ByteDance fed the portraits and voices of 150,000 film extras into a ByteDance model are seriously inaccurate. He said the Hengdian film AI platform exists to build high-quality datasets for productions, that its data belongs to Hengdian or its licensors, and that it will not train ByteDance models.
Doubao public relations lead Liu Xing pushed back publicly on October 28 against a widely circulated claim that Hengdian and ByteDance had fed the portraits and voices of 150,000 film extras into a ByteDance model, calling the story seriously inaccurate. The response was picked up by multiple Chinese outlets the same day, and the real question at its center is not whether film production should use AI but who owns the data in the Hengdian film AI platform, how it is licensed, and what it is used for.
According to Liu, the Hengdian film AI service platform is intended as public digital infrastructure: it turns assets such as the studio's physical locations into high-quality datasets so that production crews can scout, prepare and shoot more efficiently, while lowering the barrier for smaller film companies to use AI and giving more creators a chance to make their work well. That positioning differs fundamentally from the rumor, which described a bulk transfer of extras' likenesses and voices into training data.
The division of roles was spelled out as well. Liu said Volcano Engine is the technical support provider on the project, that all data assets belong to Hengdian or the parties that authorized them, and that any use of a person's likeness or voice requires that person's consent. He also stated plainly that the project's data will not be used to train ByteDance models.
Those statements mark out the three boundaries that matter most in AI data disputes: who holds ownership, how consent is obtained before use, and what the data is ultimately used for. A film studio and its extras are close to a worst-case setting for this, since the material is full of identifiable faces and voices, and any ambiguity about purpose can turn an industry partnership into a public controversy.
In a broader sense, this kind of public clarification has become routine for Chinese model developers. As scrutiny of data provenance and licensing chains intensifies, vendors increasingly have to separate their role as technology provider from the data holder's role, and to say explicitly which data stays out of the training set. The value of the response is less in refuting one rumor than in putting the project's compliance framework on the record.
What to watch now is whether that framework survives contact with verifiable detail: whether the consent mechanism is transparent to the people involved, whether the dataset's permitted uses are set out in writing, and how quickly the platform actually ships in scouting, preparation and shooting. For the film industry, a working public dataset really could lower the cost of using AI, but only if every likeness and voice can be traced back to clear consent.
Why it matters
The denial matters less than the three boundaries it sets out on ownership, consent and training use, giving studio-based AI dataset projects a compliance template to be measured against.
Nearby Updates
All10/28, 16:55
Physical Intelligence in talks to raise $1B at a reported $20B valuation, with no revenue
Physical Intelligence is negotiating a new round of at least $1 billion that would value the robot foundation-model company at about $20 billion, Bloomberg reported. The San Francisco startup, founded in 2024, has no commercial revenue and builds no robots of its own, betting instead on a general robot brain that runs on third-party hardware.
10/28, 16:59
MoleculeMind reports 18 of 19 design projects hit, with all six GPCR targets bound
MoleculeMind published a technical report on MMDesign, its AI protein design platform, covering 19 design projects across 18 targets, 18 of which produced experimentally confirmed binders for a 94.7% project success rate. The report also claims its in-house MMFold structure model roughly doubles AlphaFold 3's top-1 high-quality success rate on antibody-antigen targets in the FoldBench benchmark.
10/28, 17:04
Rokid adds a full-color AI glasses model starting at 4,999 yuan, with a ring for input
Rokid has introduced a new full-color AI glasses model starting at 4,999 yuan, adding a ring for gesture input, according to a retail roundup published by Ebrun on October 28. The report frames the launch as part of an escalating race in AI glasses but gives no specifications, ship date or battery detail.
10/28, 17:30
Legora builds an agentic operating system for legal work on OpenAI models
OpenAI's October 28 customer story details how Stockholm-based Legora uses its models to power contract review, due diligence and agentic legal workflows. The case study says NDA review dropped from about 20% of Legora's legal-team capacity to under 1%, while one Fortune 100 customer cut contract turnaround by 90%.