Generative artificial intelligence is moving from a novelty to an information gateway, but the growth in news use remains modest and uneven. A 2026 Reuters Institute report says the share of respondents in six countries who used generative AI to get news rose from 3% in 2024 to 6% in 2025 and 9% in 2026. That is a fast increase from a small base. It is not evidence that AI has replaced news websites or that users trust every answer they receive.
AI use is rising, but news use remains limited
The report’s distinction between general AI use and AI used specifically for news is important. People may use a chatbot to draft a message, translate a paragraph or plan a trip without asking it to interpret current events. The six-country comparison offers a consistent time series, but it should not be treated as a universal global estimate. Survey methods, country selection, internet access and how respondents understand “using AI for news” all shape the result.
What the survey found
In the United States, 12% said they had used generative AI for news in the previous week, according to the Reuters Institute report. The authors note that this level rivals the online reach of some individual news brands in their comparison. That is a useful signal for publishers: an AI interface may become a meaningful point of contact. But comparing weekly self-reported use with a brand’s reach does not show equivalent attention, trust, time spent or willingness to pay.
For audiences, the appeal is straightforward. A conversational system can answer follow-up questions, summarize a long document or explain unfamiliar terminology in a way that feels tailored. It can make information easier to approach, especially for readers who do not know which publication or source to consult. Yet convenience shifts responsibility. If a summary compresses conflicting evidence into a confident paragraph, the user may not see what was omitted or who produced the original reporting.
The risk to publisher visibility
Newsrooms also face a distribution question. Search engines historically sent readers to publisher pages, where they could see context, corrections, author information and related reporting. An AI answer may instead keep the reader inside a platform. If publishers’ work informs answers but traffic does not return to publishers, the economic relationship may change. The scale of this effect is not established by the Reuters Institute survey; it requires traffic data, licensing information and direct reporting with publishers and platforms.
The democratic concern is broader than revenue. Journalism depends on editorial processes: assigning reporters, checking documents, contacting sources, distinguishing allegation from fact and correcting errors. A language model can help users navigate information, but it does not automatically reproduce those responsibilities. If a chatbot gives an answer without clear source links, readers may have difficulty checking whether it reflects a published report, an official record or an inference assembled from several pages.
Useful newsroom applications need safeguards
There are reasonable reasons to use AI in news work. It can assist with transcription, translation, document search, data extraction and accessibility, when trained staff verify the output. But efficiency is not a substitute for reporting. A model can make a task faster while still missing context, mishandling names or inventing a link between facts. The key newsroom question is not simply whether AI is used, but which tasks it performs, who reviews the result and how readers are told about material assistance.
Build trust through traceable sourcing
For publishers, practical safeguards begin with visible sourcing. AI-generated summaries should link to original reporting and primary records, identify uncertainty and avoid presenting interpretation as established fact. Newsrooms should maintain an accessible corrections policy and preserve human responsibility for editorial decisions. When a model is used in production, disclosure should be specific enough to explain the role it played rather than relying on a vague label such as “AI-assisted.”
For platforms, the central test is whether they can provide traceable answers and fair visibility to the reporting they draw on. A source list is useful only if the sources actually support the claims and the links work. Users should be able to compare information, reach the original article and understand when evidence is incomplete. These design choices influence whether AI helps people find journalism or makes its origins harder to see.
The 9% figure is therefore best read as an early warning and an opportunity, not as a settled verdict. The audience is expanding, while the habits, trust and economics around AI-delivered news are still taking shape. Researchers will need repeated surveys, and publishers will need transparent analytics, to identify who uses these tools, for which subjects and with what outcomes. A strong information ecosystem should make new interfaces answerable to the journalism they present.
[Interview needed: a reader, a newsroom editor and a researcher on whether AI summaries change how people verify news.]
What better evidence would look like
To understand whether AI news tools help or harm public understanding, researchers need more than a single question about weekly use. They can compare how readers evaluate identical claims when delivered by a chatbot, a news article or a search page; record whether readers open linked sources; and measure whether they can identify uncertainty or correct errors afterward. Such studies should include different languages, ages and levels of digital access. A survey response tells us what people remember doing, but observed tasks and interviews can reveal why they chose a tool and what they did with its answer.
Publishers can also track referral traffic and the share of their work appearing in AI-generated answers, while protecting user privacy and explaining the limits of their analytics. Platforms should publish clear documentation about source selection and give publishers ways to report attribution problems. None of this requires rejecting new interfaces. It makes adoption accountable to outcomes: whether people find reliable reporting, whether original journalism remains visible and whether readers can verify what they are told. The Reuters Institute’s trend line gives a reason to investigate these questions now, before use becomes commonplace enough that design choices are harder to change.
Audiences are not one group with one set of needs. A student may ask for background on a conflict, a voter may seek current local procedures, and a person with low vision may rely on speech output. Each use case raises a different standard for freshness, detail and source attribution. A system that summarizes stable background material can still fail when asked about a court order issued hours ago. News products should make dates visible, route users to authoritative updates and avoid answering time-sensitive questions from stale material. Researchers can test these features with real users and publish results that others can replicate.