Does algorithmic news shape what people think matters?
A randomized experiment with 798 US and UK adults found no uniform agenda-cueing effect from a simulated personalized feed. Human-editor cues shifted US ratings modestly, while UK results were null and the personalization was deliberately crude.
The short version
3 key points
- 1.The experiment randomly assigned 396 US and 402 UK adults to a journalist-curated feed, a simply personalized algorithmic feed or no feed.
- 2.In the US, the journalist-labelled feed modestly raised importance ratings for the lead budget issue and, in a planned contrast, the set of cued issues. No comparable overall effects appeared in the UK.
- 3.The algorithmic feed did not shift issue ratings uniformly in either country. Subgroup findings in the US depended on political interest and ideology and should be treated cautiously because the tailored groups were small.
The answer: not uniformly, and the label and audience mattered
A single encounter with a simulated personalized news feed did not consistently change what adults in the United States or United Kingdom said were the most important national problems. The experiment instead found a narrower result: US participants shown a feed labelled as selected by newsroom editors rated the lead budget issue more important than people who saw no feed, and a planned comparison also found a modest increase across five cued problems. The same journalist cue did not move the UK sample, while the algorithm-labelled feed produced no overall effect in either country.
That is useful evidence against a simple claim that an algorithmic ranking automatically transfers its priorities to an audience. It is not evidence that real recommendation systems are harmless. Participants saw one static feed assembled for the study, not a live service learning from years of behaviour. The researchers found different patterns among small US subgroups defined by ideology and political interest, suggesting that personalization may interact with predispositions rather than act as one common signal. Replication is needed before those subgroup results guide platform policy.[1]
How the 798-person comparison worked
In October 2025, the researchers recruited adults through the international panel provider Dyanta. The final samples were 396 people in the US and 402 in the UK. Their average ages were 44 and 42 respectively, and about 43% in each sample had completed university. Participants first reported political interest and ideology, then were randomly assigned to a journalist-prioritized feed, an algorithmically personalized feed or a no-exposure control. Checks found no baseline difference between conditions in ideology or political interest.
Both treatment groups viewed a static page modelled on Google News. The journalist condition was described as stories selected by newsroom editors and used mainstream and partisan headlines. The algorithm condition was described as tailored to the user. People with low political interest saw entertainment-oriented sources, while politically interested people saw sources aligned with their self-reported left or right position. The journalist groups contained 143 US and 138 UK participants; algorithm groups 149 and 133; controls 104 and 131.[1]
What changed—and what did not
US participants in the journalist condition gave the lead budget topic an average importance score of 4.41 on a five-point scale, compared with 4.13 in the control group. The overall three-condition test was just at the conventional threshold, F(2,387)=2.96 and p=.05. The omnibus test across all five cued issues was not significant, although the prespecified journalist-versus-control contrast was, t(387)=1.98 and p<.05. Non-cued issues did not rise, which weakens the explanation that seeing any news simply made every problem seem more pressing.
The UK sample produced no significant differences for the lead budget issue, the combined cued issues or the comparison issues. Across both countries, the personalized feed did not raise cued-issue importance overall. A deliberately included low-teen-pregnancy headline also did not make participants rate that declining phenomenon as a larger national problem, suggesting they were not merely repeating every visible topic. General attitudes toward AI and algorithmic news did not significantly alter the personalization effect in the fitted regression models.[1]
The subgroup signal is interesting but fragile
Among politically interested, left-leaning US participants, the personalized feed was associated with higher ratings of the cued problems. Among US participants with low political interest, the fitted contrast went in the opposite direction. The paper interprets this as evidence that ideology and interest can condition algorithmic agenda cues. Yet the tailored cells were small: the low-interest algorithm group contained 28 US and 22 UK participants, and the left-leaning groups 57 and 40. A very large reported t statistic for the low-interest contrast is another reason readers should seek replication and underlying data before treating its magnitude as stable.
The simulated personalization was coarse. It selected one of three source mixes from a pre-test response and showed all participants the same five national topics. Real systems draw on far richer histories, update continuously and optimize for objectives that may include engagement, retention or commercial value. Those differences could strengthen, weaken or reverse a cueing effect. The study therefore tests whether a visible curation cue can matter under controlled conditions; it does not estimate the effect of an operating platform on a population.[1]
What this means for editors, platforms and readers
For editors, the US result suggests that explicit human selection can still signal importance even when people distrust institutions. The UK null warns against assuming the same presentation travels across political contexts. The researchers note that their UK topics mirrored the US agenda, including a budget story before a UK budget announcement, so the comparison may have made the British treatment less immediate. Newsrooms should test prominence and explanatory labels with their own audience rather than import an effect from another market.
Platforms should make personalization legible and preserve routes to a shared agenda. That can mean explaining why a story is shown, offering a non-personalized or editor-led view and allowing comparison with a cold-start feed. For public-service communicators and teachers, the result supports teaching readers to distinguish a story's intrinsic public importance from the signals created by placement, repetition and presumed personalization. It does not support replacing professional news judgment with a claim that audiences will correct every distorted ranking themselves.[1]
Evidence limits and what would change the assessment
The study measured issue-importance ratings immediately after one exposure. It did not measure article reading, memory, factual knowledge, later media choice, civic action or voting. Panel samples can approximate demographic distributions without being probability samples. The stimuli used credible headlines, but their source, tone and political meaning could differ between countries. The authors did not receive funding, declared no competing interests and report institutional ethics approval; participant-level data are not public because of privacy and ethics restrictions.
Confidence would rise with preregistered replications that use locally salient agendas, larger ideological subgroups and behavioural outcomes. Field experiments could compare an editor-led feed, a transparent personalized feed and an opaque personalized feed over weeks while measuring exposure diversity, knowledge and whether people revise issue priorities. Independent audits should also test actual platform outputs. Until then, the strongest conclusion is modest: curation signals can shape salience under some conditions, but an algorithm label alone did not move these two samples as one bloc.[1]
What this means for people
- Readers should treat ranking and repetition as editorial signals, not neutral measures of a topic's public importance.
- Newsrooms may preserve common ground by keeping an identifiable editor-led view alongside personalized feeds.
- Teachers and public-service communicators can make feed provenance and personalization part of media-literacy practice.
Global context
The experiment directly compared US and UK adults and found different country-level results despite similar average attitudes toward AI. It cannot establish that regulation caused the difference, and its UK agenda may have been less locally appropriate. Replication across languages, political systems and actual platforms is needed.
What the evidence does not yet show
- One immediate exposure to a static simulated feed, not longitudinal use of a live platform
- Panel samples rather than national probability samples
- Small personalization subgroups and a coarse ideology-and-interest treatment
- UK topics were chosen to mirror the US agenda and may not have carried the same meaning
- No measures of later knowledge, behaviour, civic action or real-world news consumption
What to watch next
- Preregistered replications with country-specific agendas
- Field tests on live feeds with independent access to ranking data
- Effects on knowledge and exposure diversity rather than immediate ratings alone
- Larger samples for political-interest and ideology interactions
- Whether transparent editor-led and cold-start options preserve common ground
Living evidence record
Impact record IAI-0UQPWNE
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
15 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Communication and Change published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 15 October 2026
Research question, topics and regions
Whether journalist-labelled or algorithmically personalized headline feeds alter adults' ratings of which national problems are important
Topics: News personalization · Agenda setting · Algorithms · Political communication · Media literacy
Regions: United States · United Kingdom · Europe · North America
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Watch the background, then check the evidence
The Royal Society
AI-generated illustration · not a video frame
YouTube loads only when you choose. Press play in the player to start. If playback is unavailable, use “Watch on YouTube” below.
Healthcare, culture & human impact · Background explainer
Published 2026-07-01 · 44 min 28 sec
Is this the AI future we want? | The Royal Society
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Source checked 14 October 2026. Check available captions on the publisher’s video page.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Society & Media
Can people spot AI-generated images?
A preregistered experiment with 2,091 adults in Denmark found that people were slightly more accurate than random probability guesses, yet on average would have scored better by assigning every image a 50% chance of being AI-generated. The test used one 2024 image tool and does not measure belief, sharing or harm.
10 min · 1 source
Society & Media
Do AI agents polarise each other?
A peer-reviewed simulation found polarization and homophilic clustering when 1,000 language-model agents repeatedly interacted. The result is a warning about synthetic agent societies, not evidence that real voters or social platforms will behave the same way.
8 min · 2 sources
Society & Media
How did fake experts reach real newsrooms?
OpenAI says two covert influence operations used false journalist identities and a front research centre to place material in real outlets. AI mostly helped with drafting, translation and internal reporting; the harder failure was identity and source verification, and claimed reach remains only partly corroborated.
10 min · 2 sources
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.