AI journalism trust: what readers think of AI-made news

Readers are uneasy about AI-made news. In the Reuters Institute's 2025 survey of six countries, 12 percent were comfortable with news made entirely by AI, against 62 percent for news made entirely by humans. A US experiment by Toff and Simon found that AI labels lowered trust slightly, and that listing the sources used largely offset the drop.

EasWrk is a hosting platform, not a research institute. This essay summarises published sources, and the sources are listed at the end. More essays are on the EasWrk Research index.

Do readers trust AI journalism? What the surveys show

Trust in news is low, and AI arrives into that. The Reuters Institute's Digital News Report 2026, published on 16 June 2026, put overall trust in news at 37 percent across markets, the lowest since measurement began in 2015. The US figure was 25 percent. The report found that 42 percent of respondents sometimes or often avoid the news.

The same report found that weekly use of AI chatbots for news was 10 percent globally, up from 7 percent a year earlier. Trust in answers from AI chatbots was 20 percent globally.

The Reuters Institute's Generative AI and News Report 2025 (Simon, Nielsen and Fletcher, 7 October 2025) asked about 2,000 people in each of Argentina, Denmark, France, Japan, the UK and the US. YouGov fielded the survey from 5 June to 15 July 2025. The share comfortable with each kind of news was:

• Entirely human-made news: 62 percent.
• Human-led news with some AI help: 43 percent.
• AI-made news with a human in the loop: 21 percent.
• Entirely AI-made news: 12 percent.

Respondents expected AI to make news cheaper to produce (net +39 points) and less trustworthy (net -19 points). Only 33 percent thought journalists always or often check AI output before publishing. Among people who strongly trust news, 57 percent thought such checks happen. Among people who strongly distrust news, 19 percent did. The numbers describe attitudes, not accuracy. They still show that people accept AI help more readily than AI authorship.

Does disclosing AI use lower trust in news?

Toff and Simon surveyed nearly 1,500 US adults in September 2023, first reporting the study as a preprint in 2023 and publishing it in 2025. More than 80 percent said news organisations should tell readers when AI was used. When articles were labelled as AI-generated, respondents rated a mock news organisation about half a point lower on an 11-point trust scale. They did not rate the content as less accurate or more biased. When the labelled articles listed the news reports used, the trust penalty was largely offset (Nieman Lab, 2023; Toff and Simon, 2025).

The sample skewed slightly more educated and more liberal than the US public, and the study used a mock outlet. It is one experiment. It suggests that a bare label costs a little trust and that visible sourcing recovers it.

AI content farms and the limits of detection

NewsGuard's AI Tracking Center listed 3,749 AI content farm news and information websites in 16 languages as of 23 June 2026. A site qualifies only if it meets four tests: AI produces a large share of its content, the content is published with little or no human editing, its layout or name could lead a reader to think human journalists wrote it, and it gives no clear disclosure that AI produced the content.

Detection tools are not a reliable answer. Liang and colleagues (2023) found that GPT detectors consistently misclassified writing by non-native English speakers as AI-generated, while accurately identifying native writing. The education essay covers the same detector problem in schools.

Language matters too. Wardle and Derakhshan (2017) separate misinformation (false content shared without intent to harm), disinformation (false content shared to harm) and malinformation (accurate material released to do damage). Tandoc, Lim and Ling (2018) reviewed 34 academic articles that used the term "fake news" between 2003 and 2017 and derived six types: satire, parody, fabrication, manipulation, advertising and propaganda. Nielsen and Graves (2017) found that audiences use "fake news" to mean more than fabricated stories, including frustration with news media, politicians and sponsored content.

AI journalism examples: from AP earnings recaps to CNET

In 2014 the Associated Press began using Automated Insights technology to write earnings recaps. TechCrunch reported on 1 July 2014 that coverage would grow from about 300 companies a quarter to about 4,400, in stories of 150 to 300 words. Earnings recaps are formulaic, which suits a template.

Large language models can write on any topic, which removes that limit. In January 2023, CNET issued corrections on 41 of the 77 stories it had written with an internal AI engine, according to Engadget's report of 25 January 2023. The errors included incomplete company names and unclear wording, and some stories contained phrasing that was not original. CNET's editor-in-chief said the outlet had paused the tool and would restart it when confident.

Coddington (2015) separates computer-assisted reporting, data journalism and computational journalism as related but distinct forms. Diakopoulos (2019) gives a book-length account of how algorithms entered newsrooms.

Newsroom jobs and money before AI arrived

Pew Research Center (2021) used US Bureau of Labor Statistics data to show that newsroom employment fell 26 percent from 2008 to 2020, from about 114,000 to about 85,000. Newspaper newsrooms fell 57 percent, from roughly 71,000 to about 31,000. Digital-native newsrooms grew 144 percent, from 7,400 to about 18,000.

Pew's newspaper fact sheet estimates US newspaper advertising revenue at $49.4 billion in 2005 and $9.6 billion in 2020. Both sets of figures predate generative AI. They show the pressure AI arrived into. In this essay's reading, fewer staff means fewer people to check machine output. That is an inference, not a finding of either source.

Local news deserts in 2026

Medill's State of Local News 2026 report, released on 13 October 2026, counted 238 news desert counties. That is 26 more than a year earlier, the largest one-year rise since Medill began keeping records five years ago. Newspapers closed at nearly three a week this year. The reports I read do not show AI sites filling these gaps. They count different things, and this essay does not claim a link.

The New York Times Company shows the other end of the market. It reported $1.55 billion in subscription revenue for 2022, including $978.6 million from digital-only subscriptions, and 9.55 million subscribers at year end (New York Times Company, 2023). Large outlets can sell journalism directly. Many local papers cannot.

Algorithmic distribution and what it amplifies

Pew's 2026 survey of 5,166 US adults, fielded 24 to 30 August 2026, found that 59 percent get news from social media at least sometimes. The figure was 53 percent in 2020 and in 2025. In 2026, 40 percent got news regularly from Facebook, 34 percent from YouTube, 23 percent from Instagram and 20 percent from TikTok.

Vosoughi, Roy and Aral (2018) studied about 126,000 news cascades on Twitter from 2006 to 2017, shared by about 3 million people. False stories were 70 percent more likely to be retweeted than true ones, and true stories took about six times as long to reach 1,500 people. The authors attributed the faster spread mainly to humans, not bots.

That study predates generative AI and does not show that AI changed the pattern. If AI makes emotional content cheaper to produce, the incentive that study documents becomes cheaper to serve. That is an inference. The social media essay covers engagement algorithms in more detail.

Platform accountability: Section 230, the DSA and content provenance

In the United States, Section 230 of the Communications Decency Act says that no provider or user of an interactive computer service shall be treated as the publisher or speaker of information provided by another information content provider (47 U.S.C. 230).

The EU's Digital Services Act (Regulation (EU) 2022/2065) applies from 17 February 2024. It covers very large online platforms, which have 45 million or more average monthly active recipients in the Union. Their risk assessments must cover actual or foreseeable negative effects on civic discourse and electoral processes, and public security.

The Coalition for Content Provenance and Authenticity (C2PA) publishes the Content Credentials technical specification, which it describes as a "nutrition label for digital content". Its steering committee includes the BBC, Google, Meta, Microsoft and OpenAI. A provenance label only helps if publishers and platforms both adopt it.

What publishers can do about trust

The surveys above point to two practical steps. Show the human checks, because readers doubt that journalists check AI output. List the sources behind a story, because that largely offset the trust penalty in the one experiment I found. Both are inferences from the studies, not tested results from newsrooms.

The deeper limit is the one the Reuters Institute data shows. Overall trust in news was 37 percent before AI-made news became common. AI did not cause that figure, and better labelling alone will not fix it.