Doubao denies claim that 150,000 extras' likenesses and voices trained its model

Doubao publicly denied a claim circulating on October 28 that the likenesses and voices of 150,000 film extras had been fed into a large model, marking the company's first direct response to that specific allegation.

The rumor itself was concrete: that a large pool of background performers had their faces and voices used as training data. If true, the allegation would touch not only data-sourcing compliance but each performer's control over their own likeness and voice, and their right to be paid for it.

The response was a denial, and the public statement did not go into further detail. That leaves one confirmed fact — the company says the claim is not accurate — while the question of what actually happened along the data chain still needs a fuller account.

The sensitivity is easy to understand. Likeness and voice are among the most contested categories of training data right now. Digital humans, voice cloning and video generation all depend on large collections of faces and voices, and the authorization chains behind such material are harder to trace than text corpora. Individuals rarely know whether they have been "trained on" at all.

For Doubao, the denial is also damage control. Once an impression forms that a model was fed performer data, partnerships in film and content production become harder to close. But a denial is not an explanation: what the public wants to know is not only whether it happened, but where the data came from, who authorized it, and how someone opts out.

Three things are worth watching. Whether Doubao follows up with a fuller account of its data practices. Whether performer groups push for clearer claims over how their likeness and voice may be used. And whether, as scrutiny of AI training data intensifies, rumor-and-denial cycles of this kind become routine.