The AI anchors of Xinhua TV

by Andrew Cochran
updated October, 2019

In the 1980s, the UK production Max Headroom parodied a computer-generated figure known for its sarcasm and digital stutter. 30 years later, AI-powered personalities are presenting the news-of-state in China. It’s all serious — no stuttering here, at least not on purpose. And if a stutter were to happen, no doubt there’d be a swift software upgrade. The bigger question: can there be an upgrade for sincerity, too?

They are the world’s first AI-enabled news presenters. They became their virtual selves thanks to developments by Sogou, China’s second-largest search engine company. Sogou collaborated with Xinhua News Agency, the state-run press agency and China’s largest media organization. In a prepared statement, Sogou says their technology can ‘revolutionize’ human-computer interactions.

Sogou has deep pockets for deep learning. In 2017 the company raised US$585 million on the New York Stock Exchange to advance its work in AI. Their largest shareholder is one of China’s foremost tech companies, Tencent, which is the world’s biggest games company.

More than the news is being presented on TV. Sogou‘s synthetic anchors show AI technology at work in a country where AI advances are government policy. China has set a goal of being the world leader in AI technologies by 2030.

Television news as a testing ground

Television is an interesting place to stake ground. TV newscasts are high profile. They also are substantially the same around the world, making export possibilities easier. The newcast format follows fixed conventions from beginning to end, as if following silent rules. The most important stories of the day roll out in order of their ‘news value,’ amplified with details from reporters. We’ve already seen AI systems excel in orderly environments – witness chess, poker, and GO. TV news stories differ daily, but their presentation is remarkably routine, nation after nation. Recurring tasks are fertile ground for AI.

A news program is like a freight train. The engine is the news brand. The newsroom assembles a series of train cars each day, with changes of the day as the cargo for each car. They are joined together by a presenter (or two), coupling one piece to the next. They’re even called ‘anchors’ for their stabilizing effect.

The role of anchors is also to connect with viewers, making the aggregated changes of the day easier to absorb. Having the news presented by other-than-humans might be more clinical or disturbing for a human audience. But is it?

The answer may hinge on whether sincerity can be a science. The secret sauce of a news anchor is their trustworthiness. It lives in some and not others, and the intangible “it” factor takes more than a favourable face ratio and the demeanour to go with it. They must be believable. The news they present must not only be true in fact but also appear to be true in perception, too.

The Sogou method

This is where Sogou’s approach is intriguing. Inventing a new life-like form requires many choices. Male or female? Tall, short, or in-between? Complexion? Hair colour and style? Voice? Mannerisms? Which will be the most believable? When humanity is synthesized, how do you decide?

Sogou ‘casts’ their AI figures by replicating real people. Each is a well-known personality. In this case, they are existing anchors on Xinhua TV. Copying from real humans solves identity questions and, maybe, attaches the essential intangible: familiarity for the audience. Might it include built-in believability? As with human anchors, it will be up to the audience over time. Public opinion sampling in China has many variables. It is hard to imagine we’ll know.

Two plus one

The first two AI anchors were unveiled in November 2018. It was a celebrated event, an occasion at the World Internet Conference, a high-profile annual event in China that the Guardian calls ‘China’s Davos for the tech sector‘. Both figures replicated male anchors on Xinhua. One of them is programmed to speak in English.

In a hand-out video, the English-speaking figure said he could ‘work tirelessly’ to present the news. Xinhua said in an accompanying statement they would be using the AI figures for ‘reducing news production costs and improving efficiency’. 

YouTube link to a text-on-video report by South China Morning Post, November 2018
(may be preceded by a commercial)

A third AI figure, modelled on a female anchor at the network, was announced by Xinhua three months later. Shortly after, she reported from another high-profile event, the Two Sessions conference, the annual gathering of Chinese leadership similar in prominence to a sitting of parliament.

YouTube link to video from New China TV

The technology

The science behind the AI figures is unclear, even whether they are computer-generated images or a form of robotics. One report from The Verge suggests they are digital video composites. Another, from China Daily/Asia News Network, quotes an AI professor describing they use ‘micro-electric motors,’ suggesting a more mechanized approach.

Handout materials from the company only say they integrate ‘advanced image detection and prediction capabilities’ with speech synthesis, ‘which allow the virtual anchor to broadcast text inputs in real-time.’ There is no mention of presentation methods.

Either way, the AI figures are fully programable. As anyone who has been on television knows, ‘being yourself’ in front of an audience is not always easy. Humans sometimes use talent coaches, specialists who help news anchors and business executives present themselves more naturally in front of the camera. One of the techniques is using gestures while speaking. It can take several sessions to get it right. With the Sogou anchors, it takes only a tweak in software. The following video shows progress three months after the first version (watch the hands).

YouTube New China TV

Is imitation really the sincerest form of flattery?

There are many implications of AI-replicated anchors. Are their human counterparts entitled to compensation for using their likeness? Does an indefatigable replica provide value or peril to the human anchor? An AI replica can be on-air while its human is otherwise occupied, perhaps on assignment, vacation, or taking a nap. More airtime gives the human TV personality a higher profile. But the AI figure could say things their human counterparts might find unpalatable, damaging a trusted reputation.

AI figure Xin Xiaomeng (left) and Xinhua anchor Qu Meng (right)

Then there are the legal issues. In the West, there is a nascent ‘law of personality,’ where concepts of ownership pivot on the extent there is a distinctive personality to begin with. Is having a digital copy de facto proof personality rights exist? Some jurisdictions for privacy law offer a ‘right to be forgotten.’ Should there be a right not to be replicated? How might the copies be protected from further copying or derivations? In patent law, The Supreme Court of Canada determined in 2002 that a higher life form cannot be patented. That case involved a transgenic mouse. What about a synthetic human? If you had a digital twin, who would own it?

A threat to human anchors?

Much more tangible is what this means for job security. Many human news anchors read from a text that’s written for them. Directions may be from a producer in the control room. An AI figure that makes no mistakes and has no limits on how long it can work could be a competitor for the job. One of the human anchors on Xinhua, who was replicated as an AI anchor, has been thinking about what it means for him. In the following report from South China Morning Post, he says his strengths are presenting ‘richer emotions and a more realistic human voice.’

YouTube link to text-on-video report by South China Morning Post, November 2018
(may be preceded by a commercial)

Sogou says their news figures could be forerunners of more AI-generated characters. They foresee artificial figures in education, medicine, and law.

‘Vocational avatars’

Sogou has since announced two additional AI anchors and three more AI characters. One is an AI anchor for television in Dubai. The second is also an anchor figure, for Russian TV. Both are modelled on known personalities; both stake additional trajectory points in the evolution of this new form.

Joining them is an online judge, part of China’s Internet court system, and two authors of storybooks, again replicating human counterparts. Sogou refers to them all as ‘vocational avatars.’

This is a cohort unlike any we’ve seen. There likely will be more, behaving at their avuncular best. Will we feel more comfortable?

Links in this article (in the order embedded above)