Nieman Foundation at Harvard
An overwhelming majority of readers would like news publishers to tell them when AI has shaped the news coverage they’re seeing. But, new research finds, news outlets pay a price when they disclose using generative AI.
That’s the conundrum at the heart of new research from University of Minnesota’s Benjamin Toff and Oxford Internet Institute’s Felix M. Simon. Their working paper “‘Or they could just not use it?’: The paradox of AI disclosure for audience trust in news” is one of the first experiments to examine audience perceptions of AI-generated news.
More than three-quarters of U.S. adults think news articles written by AI would be “a bad thing.” But, from Sports Illustrated to Gannett, it’s clear that particular ship has sailed. Asking Google for information and getting AI-generated content back isn’t the future, it’s our present-day reality.
Much of the existing research on perceptions of AI in newsmaking has focused on algorithmic news recommendation, i.e. questions like how readers feel about robots choosing their headlines. Some have suggested news consumers may perceive AI-generated news as more fair and neutral owing to the “machine heuristic” in which people credit technology as operating without pesky things like human emotions or ulterior motives.
For this experiment, conducted in September 2023, participants read news articles of varying political content — ranging from a piece on the release of the “Barbie” film to coverage of an investigation into Hunter Biden. For some stories, the work was clearly labeled as AI-generated. Some of the AI-labeled articles were accompanied by a list of news reports used as sources.
A couple of limitations to note. The news articles shown to participants, though sourced from tech startup HeyWire AI that sells “actual AI-generated journalistic content,” ran under a mock news org name, and the lack of real-world implications and associations may affect results. The sample of nearly 1,500 people also skewed slightly more educated and more liberal than the U.S. public at large. (There’s a wide — and widening — partisan divide when it comes to trust in news media.) This is a working paper or pre-print, meaning the findings have not yet been peer-reviewed.
Co-author Toff has said the idea for this research came after he was asked about trust toward AI-generated news — and he didn’t know the answer. A few takeaways from the resulting experiment and a conversation with the co-authors:
On an 11-point trust scale, survey respondents who saw the news stories labeled as AI-generated rated the mock news organization roughly half a point lower than those shown the article without the label — a statistically significant difference.
The respondents, interestingly, did not evaluate the content of the news article labeled as AI-generated as less accurate or more biased.
The researchers found the largest difference in trust among those who were familiar with “what legitimate news production and reporting entails.” People with lower levels of “procedural news knowledge,” as the researchers put it, generally did not dock the news orgs trust points for labeling content as AI-generated.
There’s some hope that generative AI could increase trust among those with the lowest confidence in media. Given historically low trust in media among Republicans in the U.S., perhaps some audiences would see generative AI as an improvement over professional journalists? An earlier experiment found that presenting news as sourced from AI reduced the perception of bias among people holding the most hostile partisan attitudes toward media. More recently, an editor from a German digital news site that experimented with AI-assisted content said an audience survey suggested some readers seem to favor “the mechanical accuracy of technology” over the “the error-prone or ideologically shaped person.”
But co-authors Toff and Simon found no improvement in this experiment. Their research showed no changes from AI disclosures among the least trusting segments of the public.
Future research could still explore whether different labels could build trust with certain segments of the public, Toff said in an email.
“I wonder if there are ways of describing how AI is used that actually offer audiences more assurances in the underlying information being reported perhaps by highlighting where there is broad agreement across a wide range of sources reporting the same information,” Toff said.
“I don’t think all audiences will inevitably see all uses of these technologies in newsrooms as a net negative,” he added, “and I am especially interested in whether there are ways of describing these applications that may actually be greeted positively as a reason to be more trusting rather than less.”
Increasing transparency has been a hallmark of many efforts to improve trust in journalism, from a “show your work” ethos to enhanced bylines. With AI tools still regularly spitting out misinformation and hallucinating sources, being given the opportunity to double check original source material is highly encouraged. To that end, some respondents were shown a list of real-life news sources that the AI used to generate the article. Links for the Barbie (left) and Hunter Biden (right) examples:
Researchers found that when a list of sources was provided alongside the news article, labels disclosing the use of AI did not reduce trust. In other words, the “negative effects associated with perceived trustworthiness are largely counteracted when articles disclose the list of sources used to generate the content.”
Confirming previous studies, Toff and Simon found an overwhelming majority believed news organizations should “alert readers or viewers that AI was used” — more than 80% across all respondents. Among those who said they wanted to see a disclosure, 78% said news organizations “should provide an explanatory note describing how AI was used.”
The researchers also accepted open-ended responses from study participants, which resulted in practical suggestions to label AI-generated content (“a universally accepted symbol” or “industry-wide labels” similar to the “standard way nutrition information is displayed on food products”) and some statements of blanket disapproval (“or they could just not do this,” one wrote).
“While people often say they want transparency and disclosure about all kinds of editorial practices and policies, the likelihood that people will actually click through and read and engage with detailed explanations about the use of these tools and technologies is probably quite low,” Toff said.
The nutritional labels mentioned by one respondent might be instructive for thinking about what news consumers want. “People want companies to disclose what’s in their food even if 99% of the time they aren’t going to actually read through the ingredient list,” Toff noted.
It’s easy to forget it’s only been one year since ChatGPT was released and helped kickstart a seismic shift in the tech industry. Many in journalism — and in our audiences — are still getting to know the technology and perceptions, for better or worse, may evolve. (The researchers found, for example, that respondents who said they heard or read “a lot” about news organizations using generative AI were more likely to say they thought AI did a better job than humans in writing news articles.)
“Audiences are already often deeply skeptical if not downright cynical about what human journalists do (and a lot of news that people encounter in their social media feeds doesn’t give them much reason to feel otherwise). Inevitably as these tools become more widely used, newsrooms will need to grapple with how to effectively communicate what these technologies are and are not being used for, and we know so little about how to do that,” Toff said. “We already expect quite a lot from the public in terms of media literacy to be able to navigate the contemporary information environment; the use of these technologies in news adds a whole other layer to that, and neither newsrooms nor the public have a very well developed vocabulary to navigate that on either side.”
Simon stressed these early findings should not deter news organizations from setting up rules around the responsible use and disclosure of AI and noted that comparative work around disclosures is “well underway.” News organizations should consider where disclosure makes sense (when an article was largely written by AI, for example) and where it may not (when journalists used an AI-transcription tool to transcribe interviews to inform the story).
The full working paper is available online or you can skim a thread from one of the co-authors.
