Editor’s note: Mark Coddington, Seth Lewis, Tamar Wilner, and Nick Mathews write a regular newsletter on recent academic research around journalism. We’re happy to bring each issue to you here at Nieman Lab.
As we’ve reported previously, AI-generated journalism provokes distrust from readers, with audiences favoring a human touch. That might seem obvious, but a lot of questions lie buried in that broad statement. First, with so many ways to use AI, what does “AI-generated” mean, and how do audiences respond to different types and levels of AI use by journalists? Second, if readers want news outlets to be transparent about their AI use (as previous research has found), but if knowing about AI actually use lowers their trust and perceived accuracy, how should publishers proceed?
A pair of studies from the journal Digital Journalism addresses these questions. The first, “The effects of generative AI in news on media credibility and selectivity: Evidence from a conjoint experiment in Chile,” took an experimental approach. Participants were asked to compare media outlet AI policies, two at a time. The policies varied on seven different dimensions related to AI use and disclosure, and participants had to say which policy they found more credible and which outlet they would choose for news.
The authors — Sebastián Valenzuela, Ingrid Bachmann, Porismita Borah, and Natalia Solís Valdés — found that human oversight was the most influential factor considered. Media outlets that require human review of all AI content were seen as more credible, and were chosen as news sources more often, compared to outlets without such oversight. Participants also thought disclosure of generative AI use was important for credibility and news use. And they were less likely to use an outlet or find it credible when AI automated both objective news stories as well as content requiring nuance and interpretation, compared with outlets that prohibit all AI-automated news writing.
But people didn’t show any preferences for whether news outlets should use AI for menial tasks — this, apparently, doesn’t influence credibility. Neither did they show preferences about AI for use personalizing news formats, or for creating visual content.
The second study, “Beyond the byline: Audience expectations for AI disclosure in news media” by Jessica Zier and Nicholas Diakopoulos, digs more specifically into the conundrum of AI labeling, using interviews. Participants outlined a number of reasons why they want AI labeling, including keeping journalists accountable, avoiding fraud, increasing trust, and acting as a cue that readers may want to verify the information. Interviewees said there’s a vital difference between an article written entirely by AI (which is the origin they assume when labels say “generated” or “made by,”) and one just aided by AI (indicated by labels like “assisted” or “in conjunction”). They were concerned about AI’s capacity for hallucination and bias, so they saw human review as essential, and wanted labeling to address that need. Unlike in the qualitative study, participants had particular concerns about visual content, saying labels were especially needed here.
But participants also explained how labels can backfire. One said, when seeing an AI label, they are likely to think, “I probably need to fact-check this and try and find another article.” Audiences also see journalism as a profession that requires specialized training and ethical integrity, making AI seem to some like a “cop out.” One said about the use of AI: “You can do that as an 11-year-old. You don’t need the training for that if you’re going to use AI to generate your entire article.”
Some of the practical upshots the study’s authors spell out, in a table of recommendations (a nice touch we’d love to see more often):
- Labels can’t be overly technical, but they must be precise.
- An interactive icon that provides information when hovered on could help avoid overwhelming readers.
- Labels should be at the top, not end of the article, so as not to be deceptive.
- At the same time, the industry needs to move towards label standardization to avoid confusing readers.
Together, these studies help build a clearer picture of how and when news outlets need to preserve the “human touch,” and how they should talk about AI. It’s becoming clearer that human involvement is considered a sign of professional responsibility and accountability — essential currency in an age of media distrust. Publishers would also do well to step carefully around AI where subjective or value-based judgments are concerned. And they should think hard about how labels are presented, especially as those labels talk about the human touch and oversight that is so important to audiences.
Research roundup
“Navigating credibility on TikTok: How young adults evaluate and verify information on the platform.” By Luise Anter and Anna Sophie Kümpel, in International Journal of Communication.
It’s easy to dismiss a news diet that’s light on news and heavy on visual social media as simply “believing whatever you happen to see on TikTok.” But of course, TikTok users are making constant credibility judgments about the information they find there, video by video, just as other news consumers are doing it with traditional news articles or broadcast news packages.
The speed of TikTok might make many of the criteria for those judgments minimal, as people use heuristics, rather than systematically processing the information they find there. But what are those quick-and-dirty strategies for assessing credibility on TikTok, and how are they shaped by TikTok’s distinct characteristics — no links, no public sharing, and often extremely short videos?
Anter and Kümpel studied German university students’ strategies by having the students message them links to health- or politics-related TikTok videos, then interviewing them about how they evaluated some of those videos. For these students, evaluating the credibility of the account posting the video was the central strategy for determining a video’s credibility. This evaluation, Anter and Kümpel said, served as “an efficient shortcut”: TikTok as a platform was generally seen as unreliable, and with little room for nuanced argument or documenting supporting evidence, it was difficult to evaluate the credibility based on the message of the video itself.
Still, the criteria by which the students considered sources credible could be heartening for news organizations. Legacy media outlets, professional experts, companies, and political actors were all seen as more credible than “random creators.” Even absent message characteristics that could signal credibility, like a more formal aesthetic style or documented evidence, these source evaluations tended to signal credibility by themselves.
Because they view TikTok’s overall credibility as low, students were loath to verify videos by searching for other videos on the same subject. But verification by leaving the app required more intentional effort than they were inclined to make. So one of the students’ primary verification methods was through the comments, which seemed to enjoy a higher perception of credibility than the actual videos on the platform.
There were elements of the videos themselves that helped support a credibility judgment — whether the videos are largely fact-based, rather than opinion-based, whether they have more formal aesthetics, and whether they’re on topics that are being covered widely all played a role in a video being perceived as credible. As the researchers noted, this presents a conundrum for journalists on TikTok: The algorithm rewards videos with “platform-typical features, such as trending music, fast cuts, or participation in challenges,” but those are the same characteristics that raise skepticism from young audiences. The solution, they conclude, may lie in a “sweet spot” between the two.
“Expressive news preferences: Identity-signaling in news selection.” By Seonhye Noh, in Journalism Studies. One of the trickier problems plaguing political pollsters in recent years has been that of expressive responding — when people say they believe something they don’t in order to express affinity for one’s tribe or hostility toward someone else’s. (For example, Trump voters deliberately giving the wrong answer to a pollster’s question about whether he or Barack Obama had a larger inauguration crowd size.) It makes survey responses more an expression of identity than actual beliefs, and makes political researchers’ job of determining what the public actually thinks even more difficult.
There’s some evidence that this applies to studies of news consumption as well: That people may say they consume news that more closely matches their partisan beliefs than they actually do, as a way to express their political identity and their support for organizations that align with it. That makes it more difficult to determine people’s actual news habits from surveys, but it may also mean that partisan news consumption might be less prevalent than we think.
Noh, a scholar at UCLA, provided the most direct test of this idea yet, creating an experiment in which people selected headlines coming from pro-gun control and anti-gun control perspectives. Some were told they would only be selecting the articles, not reading them, while others were told they would be reading all the articles they selected. They were also given hypothetical scenarios in which their choices would be public or private.
The expressive responding effect (what Noh called “expressive news preferences”) was present: Participants who didn’t have to read the news they selected chose 10% more stories consistent with their partisan beliefs than those who did have to read them. In other words, they were more likely to pick news that matched their views when they were only signaling to a researcher what they would pick, not when they actually had to consume the news. The public/private scenarios didn’t make a difference in the results, perhaps because they were only hypothetical.
But Noh wasn’t expecting whom the effect would be strongest for. She predicted that the strongest partisans would show the most expressive news preferences, matching political science research on expressive responding. But it was the moderate partisans who showed the biggest difference between selections when they did and didn’t have to read the articles. In retrospect, it made sense: Strong partisans actually do primarily want to read news that matches their partisan perspective, not just to signal to a researcher. But weaker partisans’ expressive motives may be more influential in their responses when “political identities are present but not firmly anchored,” Noh wrote. She concluded that we need to be mindful of the strength with which identity influences not just survey responses, but news engagement of various kinds.
“The growing non-commercial basis of U.S. journalism employment: Evidence from one city, 2015–2025.” By Matthew Powers, in Critical Studies in Media Communication. We’re closing in on two decades since calls to reorient American journalism around a more nonprofit and publicly based funding model started to become widespread. And a shift has certainly taken place since then, as a glance at the hundreds of nonprofit news organizations that have sprung up would indicate. But has that funding shift been able to do much to offset the hemorrhage of jobs from commercial news organizations during that time? And have they actually produced a meaningful change in the historic commercial orientation of journalism in the U.S.?
Powers’ new study suggests that the answers to those questions may actually be yes. Powers, a professor at the University of Washington, examined the makeup of newsroom employment in one U.S. city, Seattle. He built a comprehensive database of all full-time local journalists in the city in 2015, then updated it 10 years later. He classified those journalists in two related ways. The first was by ownership — public, nonprofit, private (owned by individuals, families, or small groups of investors), and market (publicly traded, or owned by private equity or hedge funds). The second was by funding — market, philanthropic, or government.
Overall employment declined by surprisingly little — 431 journalists in 2015 to 411 in 2025, or 4.6%. Of the 85 positions lost, more than 80% of them were in privately or market-owned organizations. And every one of the 65 positions gained were in public media. (Without those 65 new jobs, the employment decline would have been 20%.)
That shift meant that the share of Seattle journalists employed via public ownership more than doubled, from 10.2% to 26.5%, passing up market-based ownership as the second-ranked form of ownership in the city. The broadest share by far is private ownership, though that percentage dropped from 58.7% to 52.3%. Several privately owned news organizations ( The Post-Intelligencer, Seattle Weekly, The Stranger) have been gutted (or further gutted) since 2015, but the reason private ownership’s share of employment remained relatively steady was the Seattle Times’ use of philanthropic funding to pay for 30 staff positions.
In fact, the share of philanthropically funded journalism positions across the city doubled from 17.2% in 2015 to 35.3% in 2025. Taken together, Powers concluded that the expansion of public media and the increased philanthropic funding of newsroom positions “have sharply reduced the number of newsroom jobs that would have otherwise been lost during the past decade. They have also made segments of the journalism job market increasingly reliant on non-commercial support.” Powers acknowledged that these changes aren’t a panacea, and the local job market for journalists is still in structural decline. “These actions do, however, suggest that the rate and nature of this change is not inevitable,” he noted. “If politicians, philanthropists, and citizens wish to act to address the situation further, options exist for doing so.”
“Going with the mainstream: Exploring GPT representation of journalistic culture.” By Taewoo Kang, Tim Vos, Thomas Hanitzsch, Neil Thurman, Imke Henkel, Sina Thäsler-Kordonouri, and Wiebke Loosen, in The International Journal of Press/Politics. What kind of a journalist is ChatGPT? Journalists around the world embody a range of values and conceive themselves as taking on a variety of roles that are not only divergent from one another, but sometimes directly at odds. And we know that large language models absorb the biases of what’s been fed into them and how they’ve been engineered, which would include some of these wide array of values. So if journalists are asking ChatGPT to play journalistic roles for them, what kind of values and roles are they getting?
Kang and colleagues answered that question in a pretty elegantly simple way: They took 39 survey questions that were asked of journalists in the widely used Worlds of Journalism study and asked them of ChatGPT. They adapted them for LLM prompts and asked the questions repeatedly, slightly altering the wording of the prompts and including specific national contexts of the U.S., U.K., and Germany to ensure ChatGPT answered them reliably and that results weren’t being influenced by details in wording. They then compared the results to survey results of journalists from those three countries.
The type of journalist whose answers most closely matched ChatGPT was a centrist, full-time contract worker with a TV journalism background and a bachelor’s degree. ChatGPT least resembled right-leaning journalists, part-time journalists, journalists working for news agencies, and those without a formal degree. It didn’t align any differently across men or women.
ChatGPT’s alignment shifted a bit as questions drilled down into particularly areas. In its epistemological beliefs (about the nature of reality and knowledge), ChatGPT aligned with journalists identifying as left-leaning and female, and those working at online-native employers. But on ethical questions, ChatGPT became more centrist.
There were some variances by country as well: In the German sample, ChatGPT was the most centrist, and in the U.K. it was the most left-leaning and had the strongest resemblance to TV journalists. In the U.S. it was the most closely aligned with male journalists. Overall, ChatGPT more closely resembled German journalists’ values than those in the U.K. or U.S.
All together, the findings provide some empirical evidence pinning down what precisely are the journalistic values and biases of ChatGPT as a system. Specifically, the authors note that the model’s alignment with majority groups is cause for concern, as it “points to the need for continued vigilance among human journalists regarding the potential systematic omission of minority perspectives in LLM outputs.”
