Opinion | How AI could sap — or save — local news – Poynter

Will artificial intelligence further undermine the sickly local news sector — or save it? A Congressional hearing this month highlighted both scenarios. Of course, the answer, as always, is … it depends on how we react.
Here are a few questions that could shape the outcome — closing with a modest proposal for how AI companies could be the heroes instead of the villains in the story of local news.
Many AI models that are in wide use right now are large language models, meaning that to work well they need to train on massive amounts of data. That’s a problem for local news, where the buzzwords du jour are “news deserts” and “ghost newspapers.” When generative AI turns its thirsty eye toward local data, it will find “small language” ecosystems.
Generative AI engines won’t handle that disappointment well.
One study of elections in Switzerland and Germany found that “one third of Bing Chat’s answers to election-related questions contained factual errors. Errors include wrong election dates, outdated candidates, or even invented controversies concerning candidates.” The problems seemed worse on the local level: “When asked about candidates from specific cantons, the chatbot only replied correctly in very few cases. … Chatbots are generally not well suited to adapt to the local context or language.” One local mayor in Australia sued ChatGPT creator OpenAI when ChatGPT declared, falsely, that he had been convicted of bribery.
To be effective, AI needs much more data than conventional search engines. Even in low-news areas (as opposed to total deserts) that just may not exist. So those communities — which tend to be more rural and more poor — will end up with lower-quality AI services, and more misinformation.
Perhaps the reputable AI-driven services will admit when they don’t have enough quality information and just politely refuse to provide answers. But I suspect the pressure to produce useful localized content will drive them to scrape NextDoor and Reddit local sites instead. What happens then?
There are many ways that AI can help improve the functioning of journalism. But the technology will also make it so easy for bad actors to manufacture credible-looking fake local news that the public will struggle to distinguish fact from fiction.
In an Atlantic piece called “The Supply of Disinformation Will Soon Be Infinite,” AI researcher Renée DiResta wrote: “In countries around the world, coordinated propaganda campaigns in print as well as social media have sown social unrest, pushed down vaccination rates, and even promoted ethnic violence. Now imagine what happens when the sources of such postings are untraceable and the supply is essentially infinite.”
In the New Hampshire primary, a robocall featuring an AI-generated voice that sounded exactly like Joe Biden told Democrats not to vote in the primary. That is likely just the first taste of what is to come in the 2024 election. Jocelyn Benson, the secretary of state of Michigan, wrote, “A.I.-generated content may supercharge the believability of highly localized misinformation. … Those seeking to sway outcomes or sow chaos may enlist A.I. tools to mislead voters about wait times, closures or even violence at specific polling locations.”
In 2016, Russians managed to trick millions just by creating real-looking graphics and memes. Now bad actors can conjure video of actual local news anchors or believable synthetics to intone, with perfectly furrowed brows, complete falsehoods. A group of high schoolers created a video in which a local middle school principal was made to say horribly racist things. So it’s not hard to imagine that, say, Proud Boys could create a fake newscast, using trusted local anchors, declaring that Black residents are looting downtown — or that anarchists could concoct a “report” that police are mowing down protesters.
Below is a screengrab from a recent promotional video from Channel1.ai, which is pioneering the use of AI-generated news anchors and other features.
A promotional video from Channel1.ai that features a mix of fake and real people in a newscast that was partially generated with AI. (Channel1.ai)
This image features a mix of fake and real people (the woman on the lower left is real, I think; the anchor isn’t). Channel1.ai properly labeled them. Now imagine the same tech in the hands of Roger Stone, ISIS or a local misanthrope.
Generative AI will also worsen the rise of “pink slime” local news websites. We already have more than a thousand of these misleading sites that have been created to look like legacy local news outlets while sneakily promoting bought-and-paid-for content by partisan activists. Generative AI will enable bad actors to create tens of thousands of pink slime sites instead of hundreds. Newsguard has already identified 631 “unreliable AI-generated news” sites.
Meanwhile, AI services will sometimes attribute hallucinations to actual media outlets. In its lawsuit against ChatGPT, The New York Times stated:
“In response to a query seeking what The New York Times said are “the 15 most heart-healthy foods to eat” … Bing Chat identified 15 heart-healthy foods “[a]ccording to the article you provided” including “red wine (in moderation).” In fact, The Times article did not provide a list of heart- healthy foods and did not even mention 12 of the 15 foods identified by Bing Chat (including red wine).”
The spread of false information will probably produce what scholars refer to as the “liar’s dividend” — the notion that AI will so hobble our ability to discern fact from fancy that we will throw up our hands. Politicians will take advantage. Donald Trump recently claimed that a commercial made up of actual video clips of him was an AI concoction. Many will find that easy to believe. As Eric Schmidt and Jonathan Haidt wrote for The Atlantic, “The greater the volume of deepfakes that are introduced into circulation (including seemingly innocuous ones like the one of the pope), the more the public will hesitate to trust anything. People will be far freer to believe whatever they want to believe.”
No, this is not just moral panic pushed by backward-looking legacy media companies. A survey of thousands of AI researchers found that 86% had a “substantial” or “extreme” concern about the “spread of false information e.g. deepfakes,” and 79% worried about “manipulation of large-scale public opinion trends.”
When fakes are introduced into the national news system there will at least be some sleuths to point it out. But what happens when such problems are introduced into a town with a ghost newspaper? Who will track and rebut the deepfakes then?
The one reform that everyone agrees on, in the abstract, is better “watermarking,” or transparency about when a piece of content is “synthetic.” Why hasn’t that happened?
When you ask ChatGPT a question, it gives you an answer — it doesn’t provide you with links where you can go to find the answer. That basic fact has enormous implications for the business models of local news organizations (and all news media).
In the pre-AI era, technology firms argued that although their spiders did crawl news websites and then display news headlines, it was good for publishers, too, because they were getting traffic. The value exchange was never quite as fair as Google and Facebook claimed — many users got their information entirely from the headlines rather than clicking through — but news publishers did get billions of visits because of these links.
Even before AI, Google had moved toward providing many answers on the first search results page. The “knowledge box,” “direct answer box” and the “people also ask” sections provide meaningful snippets designed to fully answer a reader’s question. As a result, a smaller percentage of searches lead to clicks. According to one study from industry analyst Rand Fishkin, 64.82% of Google searches ended without a click (up from 50% in 2019).
Google was not the only one driving this trend. Siri and Alexa strive to provide answers, not links. Instagram and TikTok downplay the use of outbound links. Facebook has changed its algorithms to deemphasize exposure, and therefore click-throughs, to news.
Generative AI will almost certainly exacerbate this problem. The Wall Street Journal reported that “publishers have seen enough to estimate that they will lose between 20% and 40% of their Google-generated traffic if anything resembling recent (AI) iterations rolls out.” AI-driven search strives to provide the one-and-only answer, right there. Perplexity, a startup attempting to dislodge Google, is upfront about seeing a diversity of links as a bad thing: “If you can directly answer somebody’s question, nobody needs those 10 blue links,” said CEO Aravind Srinivas.
Well, nobody except news publishers. If the click-through goes away, what will replace that traffic and its attendant revenue? Might AI companies be persuaded to shift course and emphasize providing traffic to local news sources?
In its lawsuit against OpenAI and Microsoft, The New York Times described generative AI as “A Business Model Based on Mass Copyright Infringement.” It’s certainly true that to train its systems, the generative AI models vacuumed up massive amounts of content without notifying or compensating publishers. When AI-driven tools provide content, they sometimes copy large blocks of text from news articles and produce material that competes with the publishers.
The Times suit provided several examples. And at a recent Congressional hearing, the CEO of the National Association of Broadcasters said they are seeing stories being copied from local TV news shows. “When a well-known AI platform was recently prompted to provide the latest ‘news’ in Parkersburg, West Virginia, it generated outputs copied nearly word-for-word from WTAP-TV’s website,” said NAB’s Curtis LeGeyt. “The station did not grant permission for use of this content, nor were they even made aware of it.”
Generative AI providers argue that all the material falls under the “fair use” portions of copyright law because they are “transforming” the underlying material and thereby creating new, distinct content. Interestingly, they say the mimicry cases are mistakes that will be ironed out. That hints that they think they’re on stronger legal ground by claiming that fair use justifies the training of AI systems (“the inputs”) than they are the publishing of copied material (“the outputs”).
I think the publishers have the better argument but it’s unclear how the courts will come down. The big media companies aren’t going to wait; they’re creating licensing deals with the big tech players. We have seen this already with The Associated Press and Axel Springer.
But that leaves out medium and small local news players — especially family-owned newspapers, nonprofits and ethnic media. If compensation is happening, how will they get their fair share? How will they even know if tech companies are using their content? (Two senators have proposed requiring disclosure, which would help). Does there need to be some entity that bargains on their behalf collectively? Does there need to be a legislative intervention to make sure?
The American Journalism Project recently got a grant from Open AI to help nonprofit local newsrooms use AI. Among the new experiments:
A survey by the Knight Foundation found that anticipated newsroom uses of AI include: using data to create auto-generated stories (especially about sports and real estate); transcription services; extracting data from documents; creating more sophisticated paywalls and subscriber prediction algorithms; and even “self-critique systems, monitoring gender and racial bias in stories.”
But it’s also possible that we’ll end up with AI-haves and AI-have nots. The low-tech locals are much farther behind than you might realize. In a recent paper about coverage of health care in Illinois, professor Nikki Usher found that 23 community newspapers in the state had no website and 67% had no Facebook page. How can they master AI when they haven’t yet nailed WWW?
Remember how a small Long Island newspaper had actually broken part of the George Santos story — but was ignored? One reason was that the North Shore Leader did not have a Twitter account and hadn’t updated its Facebook page since 2021.
True, some of the laggards are in broadband deserts where going digital is pointless, or they have audiences primarily of older people. But the tech upgrades must happen. How else can the local news sector — especially those representing small and medium players — make sure that they can truly benefit from AI?
Most of the exciting examples I’ve seen involve doing better journalism with fewer people or helping news organizations improve reporting or storytelling.
But I remember sitting at an event at Google’s Mountain View campus around 2018 when they were teaching newsrooms how we could use Google Sheets and advanced search to help with storytelling. It was all true. But while we were getting hundreds of dollars of labor-saving tools, we were also losing hundreds of millions in advertising to Google and Facebook, thereby gutting the local news business models. We didn’t discuss that part.
This time we need to make sure that the innovations help local news with revenue generation. If that happens, perhaps this time a massive tech innovation can actually help build a better local news system.
Local news had a racial equity problem long before AI. Many mainstream newsrooms either ignored minority communities or covered them in harmful ways. Progress in diversifying newsroom staffs in the 1990s and 2000s reversed when the local news organizations shrunk.
Most innovative newsrooms now realize that the key to future success will be better engaging with and representing communities. AI can help. For instance, newsrooms will be able to easily translate articles into multiple languages rapidly and repeatedly, enabling immigrant communities to be far better served.
But right now, AI seems to be a bit racist, according to research studies … and our own eyes. Here’s what I got when I asked Meta’s image generator for a picture of “a typical criminal.”
The AI machines are synthesizing material that is “out there,” which is a problem if “there” is bigoted.
Or majoritarian. At the congressional hearing, conservative Sen. Marsha Blackburn of Tennessee said ChatGPT refused to create a poem praising Donald Trump but happily did so for Joe Biden. That’s plausible, though when I tried it a few days later, ChatGPT had learned fluent MAGA and seemed poised to generate enough for a full program at a Right Wing Poetry Festival at Hillsdale College: “In the diverse tapestry of thought, Nick Fuentes, a voice well sought,” and, “In the realm of news, a voice profound, Tucker Carlson, with insight unbound.” Still, some studies have shown that many AI models do appear to have a liberal bias, reflecting the material on which they were trained.
How will that play out on the local level? Will images or interpretations tend to reflect the majority populations — racially or politically — in whatever area a news organization operates? Will that further undermine trust in news?
Most of the discussion has assumed that the bad actors will be the tech companies. But news organizations could deploy the technology irresponsibly. Gannett got in trouble when they appeared to run product reviews generated by machine, as did Sports Illustrated. Sometimes the facts are accurate but the language is odd. On Aug. 30, using an AI-story-writing tool, the Herald and Review in Decatur, Illinois, reported: “A suffocating defense helped Franklin South County handle Bloomington North 4-0 on Aug. 30 in Indiana girls high school soccer action.”
Our industry tends to avoid collective action — creating standards or licensing to uphold the quality. But in this case, if we don’t do so, and quickly, it’s pretty likely that the drive for traffic and cost-cutting will lead to trust-eroding steps from media players, too.
At first glance, we might think: Phew! AI means we don’t need so many reporters, which is convenient since we’ve lost 57% of them since 2004. Indeed, AI will likely be used to justify further layoffs. There will certainly be particular cases when AI can do things better or more cheaply than warm-blooded workers (though it would be nice if employees helped decide how that played out).
But in the aggregate the opposite is true. AI makes it more urgent that we dramatically increase the number of professional local reporters.
First, generative AI cannot be effective if the content it ingests is wrong, biased or nonexistent — so we need to create more accurate content on the local level to train the bots. If we don’t, more misinformation will spread. And from a self-interest perspective, AI companies may be less able to create useful locally oriented services.
Second, we will need more reporters to sleuth out the cases of synthetic news — especially created by malign players.
There will inevitably be a technology arms race between badly behaving bots and the brave new pieces of software that are trying to defeat them. But the most effective weapon may be non-AI actors (i.e. human beings). “Users and internet companies will give up on trying to judge authenticity tweet by tweet and article by article,” Renée DiResta wrote. “Many users will want to know that what they’re reading or seeing is tied to a real person — not an AI-generated persona.”
Here, local news has a theoretical advantage over national news. Readers can interrogate the journalist on the grocery store line. They can feel hard matter when they poke a finger in the reporter’s chest, and watch her tongue move when she talks. Signs of human life.
Alas, the likelihood that Americans will ever meet a reporter is declining, not rising. We need to reverse that trend.
Here’s an idea: The generative AI sector that is apparently going to generate $2.6 to $4.4 trillion in economic value — and is taking in billions in investment — should help finance the support of 25,000 local reporters. That will likely create tremendous value for their sector because they’ll be able to actually have locally-focused generative AI services — benefits that will far exceed the cost ($1 to 2 billion per year). And/or they could put their muscle behind some of the efforts for government subsidies to do the same.
The only way to combat a pink slime site pretending to cover a town is to have a humanoid reporter at every town hall meeting. Every weirdly attractive synthetic local news anchor needs to be joined by a new badly dressed local reporter. The only way to defeat a bad guy with a chatbot is with a good gal with a notebook.
Poynter to launch trainings for Spanish speakers, with courses offered virtually, in St. Petersburg and across Latin America
The country’s largest newspaper company ended 2023 with a $27.8 million loss, but expects to flip revenue from declining to growing by the end of 2024
He says he wanted a ‘place to unload thoughts as we get into this election season.’ There are other ways he could have done that.
Over 100 days, we used these methods to improve our trust with audiences that only casually engage with our content.
Trump compared his civil fraud fine with the punishment of Russian opposition leader Alexei A. Navalny, who died Feb. 16 in a Russian prison.
Get the Poynter newsletter that’s right for you.
Support responsible news and fact-based information today!