Here at Ars, we’ve long covered the interesting potential and significant peril (and occasional silliness) of AI-generated video featuring increasingly realistic human avatars. Heck, we even went to the trouble of making our own “deepfake” Mark Zuckerberg back in 2019, when the underlying technology wasn’t nearly as robust as it is today.
But even with all that background, startup Channel 1’s vision of a near-future where AI-generated avatars read you the news was a bit of a shock to the system. The company’s recent proof-of-concept “showcase” newscast reveals just how far AI-generated videos of humans have come in a short time and how those realistic avatars could shake up a lot more than just the job market for talking heads.
“…the newscasters have been changed to protect the innocent”
See the highest quality AI footage in the world.
🤯 – Our generated anchors deliver stories that are informative, heartfelt and entertaining.
Watch the showcase episode of our upcoming news network now. pic.twitter.com/61TaG6Kix3
— Channel 1 (@channel1_ai) December 12, 2023
To be clear, Channel 1 isn’t trying to fool people with “deepfakes” of existing news anchors or anything like that. In the first few seconds of its sample newscast, it identifies its talking heads as a “team of AI-generated reporters.” A few seconds later, one of those talking heads explains further: “You can hear us and see our lips moving, but no one was recorded saying what we’re all saying. I’m powered by sophisticated systems behind the scenes.”
Even with those kinds of warnings, I found I had to constantly remind myself that the “people” I was watching deliver the news here were only “based on real people who have been compensated for use of their likeness,” as Deadline reports (how much they were compensated will probably be of great concern to actors who recently went on strike in part over the issue of AI likenesses). Everything from the lip-syncing to the intonations to subtle gestures and body movements of these Channel 1 anchors gives an eerily convincing presentation of a real newscaster talking into the camera.
Sure, if you look closely, there are a few telltale anomalies that expose these reporters as computer creations—slight video distortions around the mouth, say, or overly repetitive hand gestures, or a nonsensical word emphasis choice. But those signs are so small that they would be easy to miss at a casual glance or on a small screen like that on a phone.
In other words, human-looking AI avatars now seem well on their way to climbing out of the uncanny valley, at least when it comes to news anchors who sit at a desk or stand still in front of a green screen. Channel 1 investor Adam Mosam told Deadline it “has gotten to a place where it’s comfortable to watch,” and I have to say I agree.
The same technology can be applied to on-the-scene news videos as well. About eight minutes into the sample newscast, Channel 1 shows a video of a European tropical storm victim describing the wreckage in French. Then it shows an AI-generated version of the same footage with the source speaking perfect English, using a facsimile of his original voice and artificial lipsync placed over his mouth.
Without the on-screen warning that this was “AI generated Language: Translated from French,” it would be easy to believe that the video was of an American expatriate rather than a native French speaker. And the effect is much more dramatic than the usual TV news practice of having an unseen interpreter speak over the footage.
Aside from avatars based on “real people,” Channel 1 also says its newscasters “can be completely generated to have their own personality, appearance, and voice.” An example of one such whole-cloth creation delivers a report on futuristic cars in the sample newscast, but the effect is much less convincing; it looks like a character from a PS5 game delivering the news.
But Channel 1 founder and veteran producer/director Scott Zabielski told the Hollywood Reporter that the company is already looking ahead to a day in the near future when the current small inconsistencies in its AI newscasters will be completely eliminated. “We’re looking down the road at 12 months from now, 18 months from now, three years from now. It is going to get to a point where you absolutely will not be able to tell the difference between watching AI and watching a human being, but we also understand that there’s going to be a pathway from here to there.”
A matter of trust
Convincing video of anchors is only a small part of a newscast, of course. When it comes to the actual news being reported, Channel 1 seems aware that it needs to overcome the inherent skepticism of AI chatbots as the ultimate BS machines.
While the Hollywood Reporter says that Channel 1 “will use large language models (LLMs) to write its scripts,” an AI newscaster in the sample episode clarifies that this is “not fake news. There isn’t a computer somewhere writing its own news stories about things that haven’t happened.” Everything presented is pulled from “trusted sources” and run by “human editors and producers” who check for “accuracy and clarity” to generate news that is “fast, trustworthy, and accurate,” Channel 1 says.
Deadline reports that Channel 1 stories will draw from “a yet-to-be-announced news agency,” original reporting from “independent journalists,” and AI-generated stories built from “trusted primary source[s]” like government documents. This sourcing seems to work well enough in the sample episode, with a wide range of stories that sound like slightly rewritten summaries of the kind of straight reporting you’d get from a wire service like the AP or Reuters.
As presented, the pieces won’t win any awards for in-depth investigative reporting. But they suffice as a kind of simple news digest that summarizes and repackages on-the-ground reporting from actual humans.
The Channel 1 sample newscast also sprinkles in a few examples of AI-powered new analysis, like a graph categorizing the sentiment of 30,000 recent social media reactions to the launch of the Tesla Cybertruck. “We are not gathering primary data sources, but we are processing that data,” Mosam told Deadline.
Then there’s the imagery that goes alongside those talking heads. For its sample episode, Channel 1 relies on B-roll “file video,” fair use content, and licensed video from other sources (a lot seems to come from licensed video clearinghouse Storyful, based on the credits). That’s the kind of visual background noise you’re probably already used to if you watch a lot of TV news reports.
But Channel 1 also wants to use AI-generated images to add flavor to some stories, as in one example about a child who dialed 911 just to give a police officer a hug. Channel 1 compares the (clearly labeled) AI image of that phone call happening to a kind of “courtroom sketch” that can provide “footage of events where cameras were not able to capture the action.” Strong labeling will be key for this kind of “we weren’t there” AI-generated footage, as we’ve already seen that many people on social media easily mistake video game clips for real war reporting.
The future of news?
Following its public sample rollout this week, Channel 1 says its immediate plans for next year include a “daily news program in multiple countries and languages.” The Hollywood Reporter reports that the company wants to quickly expand to “produc[ing] between 500 and 1,000 segments daily,” which could help power a planned app that will provide personalized news and personalized newscasters. The company says it wants to partner with “trusted news brands” and recruit “independent journalists” comfortable with their work being delivered by AI anchors.
Whether there’s a market for this kind of AI adaptation of traditional TV news is an open question. But the wider possibilities for the technology are a bit mind-boggling. Imagine, for instance, if a company decided to loosen the definition of a “trusted source” to include overtly ideological propaganda. Soon, it will be easy (and relatively cheap) to put that kind of AI-generated agitprop into the mouth of a trustworthy-looking AI avatar—one that might soon be almost completely indistinguishable from a real human.
Or what if someone decides to cut costs further by removing the human sourcing and editorial oversight of the news entirely? Why not just tell an LLM to generate a summary of the most important news stories each day (culled from search engines and social media postings)? That summary could easily be delivered by a human-looking AI avatar—with accompanying AI-generated imagery for “footage of events where cameras were not able to capture the action”—and let it loose with minimal additional human intervention. Think “Endless Seinfeld,” but for 24/7 news.
That kind of fully automated news system might lose viewers’ trust if it obviously hallucinates some patently incorrect information. But even large factual issues might not matter if a portion of the audience doesn’t notice (or care) when it comes from an authoritative-sounding AI avatar. Such a news system might eventually sustain itself economically with even a small audience, given the lack of ongoing human labor costs for things like reporters or camera operators.
Zabielski told Deadline that for the anchors who merely read off teleprompters, “it’s hard to imagine the technology isn’t going to come along at some point for those people. But the reality is most of the people in the news business are also journalists and reporters that also appear on camera… We still need that reporting to be done, whether or not an avatar is delivering that or we’re the ones going on camera.”
That might be true for now. But eventually, some AI avatars will start spouting “news” that is not specifically sourced to any one human reporter. We’ve already seen AI-generated text jockeying for search-engine ranking positions—and sometimes crowding out actual reported news or trustworthy information. With this kind of technology, the same spammy problem could soon be coming for the world of video news as well.