LD-10 — Persuasion, Intermediation & the Information Order
EVIDENCE CUT
25 September 2026
CURRENT SOCIOLENS POSITION
The strongest evidence does not support one simple claim that “AI persuades society.”
It supports four different statements:
* frontier language models can produce meaningful attitude change in controlled, opt-in text conversations;
* newer systems can outperform strong human text persuaders under experimental conditions, and low-cost behavioural effects are now measurable;
* population-scale political influence is still unestablished because exposure, persistence and transmission to consequential behaviour remain weakly measured;
* the clearest population-level transformation already observed is intermediation: AI answer layers embedded in search reallocate attention, referrals and source exposure under platform and jurisdictional rules.
The centre of gravity of this debate is therefore:
INTERMEDIATION UNDER INSTITUTIONAL CONDITIONS.
Persuasion capability matters. But capability is not the same thing as realized social influence.
THE QUESTION
When, and through which mechanisms, does generative AI change persuasion, information intermediation and the public information environment?
A useful answer requires separating:
* attitude change;
* behaviour;
* vote intention;
* validated voting or turnout;
* exposure and adoption;
* source selection;
* referral and attention;
* citation fidelity;
* trust;
* platform defaults;
* provider and jurisdictional rules.
1. PERSUASION CAPABILITY IS REAL — AND VINTAGE-SENSITIVE
System design matters
ESTIMATE.
In large preregistered UK experiments, Hackenburg and colleagues found that post-training and prompting increased measured conversational persuasiveness by as much as 51% and 27% relative to the study baselines.
The same research found a trade-off: interventions that increased persuasion also reduced factual accuracy.
These are relative changes in the study's persuasion metric, not 51-point or 27-point opinion shifts.
Cross-national effects are heterogeneous
ESTIMATE.
A Nature study of AI conversations around elections found measurable changes in stated candidate preferences in the United States, Canada and Poland.
In the preregistered primary analyses:
* the U.S. candidate-preference effect was about 2.9 points on a 0–100 scale;
* the corresponding Canada and Poland effects were around 10 points on the study's scale.
These are experimental attitude outcomes.
They do not establish:
* that 10% of voters switched;
* that election results changed;
* or that the same effect generalizes to ordinary, unsolicited exposure.
Newer frontier systems can exceed strong human comparators
ESTIMATE — PREPRINT.
In four preregistered 2025–26 UK experiments involving 18,978 conversations, frontier AI systems out-persuaded several human comparator groups in text chat.
Against elite debaters, the measured advantage was 4.6 percentage points.
When the AI was constrained to human message length and writing speed, that advantage disappeared.
This suggests that throughput — the ability to produce more argument and factual material during the interaction — is part of the measured advantage.
It does not establish superiority in face-to-face persuasion or population-scale influence.
2. “AI ≈ HUMAN” IS NOT A TIMELESS FACT
A 2025 meta-analysis of seven earlier studies found no average difference between LLMs and humans, with substantial heterogeneity.
That result remains useful.
But newer model vintages now produce different findings.
The correct conclusion is not that one study “wins.”
MODEL VINTAGE IS PART OF THE ESTIMAND.
Older and newer systems should not be averaged as though “LLM persuasion” were a stable treatment.
3. PERSONALIZATION: A CORRECTION MATTERS
A preregistered study found that personalized GPT-4 outperformed human debate opponents.
But an Author Correction published by Nature Human Behaviour on 3 September 2026 changed an important interpretation.
After correction, the direct comparison between personalized and non-personalized GPT-4 did not reach conventional statistical significance.
So two statements must remain separate:
* GPT-4 versus human opponents: supported in that experiment.
* personalization itself as the cause of the advantage: not established by that direct comparison.
Corrections are part of the evidence, not a footnote.
3A. A HIGH-PROFILE RESULT REMAINS UNDER FORMAL CONCERN
A widely cited 2024 Science study reporting durable reductions in conspiracy belief has been under an Editorial Expression of Concern since 11 June 2026.
Science is evaluating corrected analyses after reporting and data-processing issues were identified.
The authors state that corrected results retain the original direction, significance and substantive size.
R07 therefore records the publication-status fact but does not use the original substantive effect magnitude as a settled public claim.
4. SOME EFFECT PERSISTS — BUT “DURABLE” NEEDS A HORIZON
ESTIMATE.
In one UK experiment, about 36%–42% of GPT-4o's immediate attitude effect remained at a one-month recontact.
That is evidence of partial persistence.
It is not evidence that AI-created beliefs remain stable indefinitely.
R07 does not establish general persistence beyond currently studied short horizons.
5. BEHAVIOUR IS NOT THE SAME AS ATTITUDE
ESTIMATE — PREPRINT.
Two preregistered 2026 UK experiments found that brief AI conversations increased real petition signing by:
* 12.8 percentage points in one study;
* 19.7 points in another.
The same research found almost no relationship between how much attitudes changed and how much behaviour changed.
This matters conceptually.
Persuasion can mean changing the view of someone who initially disagrees.
Mobilization can mean moving someone who already agrees to act.
Those are not the same mechanism.
The measured actions were relatively low-cost and occurred under paid, opt-in experimental exposure.
R07 does not establish effects on real-world turnout or validated voting behaviour.
6. REACH IS THE MISSING BRIDGE
A large per-contact effect is not automatically a large social effect.
The transmission chain is:
EXPOSURE → PER-CONTACT EFFECT → PERSISTENCE → BEHAVIOURAL TRANSMISSION.
Every link matters.
A peer-reviewed 2026 study found that, conditional on exposure, LLM conversations could persuade about as well as established campaign advertising under the studied conditions.
Its cost simulation estimated roughly US$48–$75 per persuaded voter for LLM chatbots versus around US$100 for traditional methods.
But established campaign methods were easier to scale.
That is why per-contact capability and population influence must remain separate.
7. CHATBOT NEWS USE IS GROWING — BUT STILL MINORITY
SURVEY ESTIMATE.
The Reuters Institute Digital News Report 2026 found:
* weekly use of standalone AI chatbots for news rose from 7% to 10% across 48 markets;
* only 1% said AI was their main source of news.
Across 27 markets, 4% of all respondents said they always or often clicked through from chatbot news answers to original sources.
Among the smaller subgroup who actually used chatbots for news, the corresponding rate was about 42%.
Those percentages use different denominators.
DENOMINATORS ARE PART OF THE CLAIM.
8. DEFAULT AI SEARCH IS A DIFFERENT KIND OF EXPOSURE
The strongest current population-level mechanism is not a voluntary political chatbot conversation.
It is AI embedded inside an information gateway people already use.
AI Overviews
ESTIMATE — WORKING PAPER.
In a U.S. desktop field experiment, conditional on an AI Overview appearing:
* outbound organic clicks fell by 39.8% relative;
* zero-click searches increased by 34.5%;
* sponsored clicks were unaffected.
The study found no measurable improvement in perceived search quality or ease.
AI Mode
ESTIMATE — PREPRINT.
In a preregistered one-week U.S. field experiment, forced use of Google AI Mode:
* reduced external click-through by 18.8 percentage points;
* reduced trust in information on Google.
A removal experiment also found a positive effect on click-through among compliers, but its ITT estimate was not statistically significant.
ITT and LATE are different estimands.
9. REAL-WORLD TRACE DATA POINT IN THE SAME DIRECTION
OBSERVATIONAL.
Pew Research Center analysed browsing data from 900 U.S. adults in March 2025.
Users clicked a traditional search result in:
* 8% of visits with an AI summary;
* 15% of visits without one.
About 18% of studied Google searches produced an AI summary.
Because summary appearance depends on query type, this comparison is observational rather than a causal estimate.
It nevertheless provides real-world corroboration for the referral mechanism seen in randomized experiments.
10. AI ANSWERS SELECT SOURCES — THEY DO NOT MERELY SUMMARIZE RANKING
ESTIMATE — PREPRINT AUDIT.
A 2026 audit of 55,393 Google queries found:
* AI Overviews activated on 13.7% of queries overall;
* activation rose to 64.7% for question-form queries;
* politically sensitive topics saw substantially lower activation;
* nearly 30% of cited domains were absent from the co-displayed first search-results page;
* 11.0% of 98,020 atomic claims were unsupported by the cited pages.
This means AI answer layers have a source-selection process that is not identical to ordinary search ranking.
The lower activation observed for politically sensitive topics is an activation pattern. It does not by itself establish viewpoint discrimination, political motive or downstream opinion effects.
But source selection, source quality and factual support are separate questions.
11. CITATIONS CAN INCREASE TRUST EVEN WHEN WRONG
ESTIMATE — PREPRINT.
In a preregistered U.S. experiment, adding citations or reference links increased trust in AI-search answers even when the citations were incorrect or hallucinated.
Displaying uncertainty reduced trust and willingness to share.
Related research finds that people can over-rely on LLM search when the model is wrong, while confidence cues can improve error detection.
CITATION PRESENCE ≠ CITATION SUPPORT.
12. NEWS ANSWERS HAVE MEASURABLE INTEGRITY PROBLEMS
PROFESSIONAL AUDIT.
A 2025 EBU/BBC-led audit involving 22 public-service media organizations across 18 countries and 14 languages evaluated more than 3,000 responses from ChatGPT, Copilot, Gemini and Perplexity.
It found:
* 45% had at least one significant issue;
* 31% had serious sourcing problems;
* 20% had major accuracy problems.
This is publisher-led accountability evidence, not neutral platform telemetry.
It establishes substantial integrity problems in the audited sample.
It does not by itself establish population-level belief harm.
13. AI DOES NOT NECESSARILY “NARROW THE WEB”
Working-paper evidence using U.S. panel data through 2024 found that LLM adoption coincided with visiting more unique websites and that LLM-routed traffic was not more concentrated than search traffic.
This does not contradict the 2026 AI-search referral results.
A standalone chatbot and an AI answer layer embedded in search are different interfaces.
The evidence therefore does not support a simple universal claim that generative AI necessarily reduces source diversity.
14. SYNTHETIC-CONTENT CAPABILITY IS NOT POPULATION PREVALENCE
AI can cheaply produce convincing text, images, audio and video.
That does not tell us how much of it reaches people.
A large fact-check dataset found that generative-AI media rose relatively recently within fact-checked misinformation, while simple context manipulation remained the majority through November 2023.
The denominator is fact-checked claims.
Population exposure prevalence remains UNKNOWN.
15. ELECTION MONITORING: LOW DETECTED VIRAL VOLUME IN STUDIED 2024 CASES
OBSERVATIONAL MONITORING.
CETaS identified:
* 16 confirmed viral AI-enabled disinformation or deepfake cases during the 2024 UK general election;
* 11 across the EU and French elections combined.
The researchers found no evidence that the detected cases meaningfully changed election results.
That is not proof of zero effect.
The monitoring strategy had a visibility threshold and incomplete platform access.
A useful rule is:
ABSENCE OF DETECTED EFFECT ≠ PROOF OF ZERO EFFECT.
16. PROVIDER JURISDICTION IS ASSOCIATED WITH WHAT MODELS WILL ANSWER
PEER-REVIEWED AUDIT.
Pan and Xu tested 145 questions concerning Chinese politics.
China-originating models showed:
* higher refusal rates;
* shorter responses;
* more inaccurate answers;
than non-China-originating models.
The differences became smaller on less-sensitive prompts.
All models also refused more often in Chinese than English.
The study is observational.
It does not establish that regulation alone caused the differences.
17. PLATFORM AND LEGAL RULES SHAPE THE INFORMATION ENVIRONMENT
From 2 August 2026, EU AI Act Article 50 transparency obligations apply in key cases including:
* disclosure when people interact with interactive AI systems;
* clear labelling of deepfakes;
* labelling of AI-generated or manipulated text on matters of public interest.
The European Commission's voluntary Code of Practice provides a compliance route.
This is an institutional condition.
It does not tell us whether the labels work.
18. LABELS DO NOT HAVE ONE UNIVERSAL EFFECT
Peer-reviewed experiments produce different results depending on what is labelled and what outcome is measured.
In one 2026 study, simply telling participants that a policy message was AI-generated did not significantly reduce its persuasive effect.
Other experiments on misleading AI-generated images found that warning labels reduced belief in the claims.
So the useful question is not:
“Do AI labels work?”
It is:
WHAT LABEL, ON WHAT CONTENT, FOR WHAT OUTCOME?
19. AI INTERMEDIATION IS BEING REGULATED AS ACTIVE CONTENT PROVISION
In July 2026, Germany's ZAK media regulator issued decisions involving Google AI Overviews and Perplexity.
The regulator stated that German media law applies to AI search and chatbots and treated generated answers as provider content rather than merely neutral intermediary output.
This is one jurisdiction-specific example of institutional reclassification.
It is not a universal legal rule.
19A. WHO GAINS VISIBILITY, WHO LOSES REFERRALS, WHO IS UNEVENLY EXPOSED?
The clearest distributional effects in R07 concern attention, access and information pathways rather than downstream political outcomes.
- Publishers lose referrals when AI answer layers satisfy more queries inside the search interface; in the Agarwal/Sen experiment, sponsored clicks did not fall alongside organic referrals.
- Younger respondents use AI chatbots for news more often than older respondents in the Reuters Institute data.
- Provider origin and prompt language are associated with different refusal exposure in the Pan/Xu audit.
- Cross-language news audits find substantial integrity problems, but R07 does not yet provide a complete language-by-language inequality map.
- Effects by class, disability, migration status and most language groups remain largely unmeasured.
These gaps are part of the result. Interface-level redistribution of attention is better measured than many downstream social inequalities.
20. FOUR POSITIONS
P1 — Scalable persuasion
Strongly supported as per-contact capability.
Population-scale realized influence remains unestablished.
P2 — Context-bounded effects
Strongly supported for reach, horizon and exposure.
The older idea that AI is merely comparable to humans is model-vintage dependent.
P3 — Intermediation over persuasion
Best supported population-level mechanism.
Default-embedded answer layers measurably reallocate attention and referrals and select sources differently from ranked search.
P4 — Institutional mediation
Strongly supported as a condition of exposure.
Platform design, provider policy, jurisdiction and law shape what appears and how it is labelled.
Their downstream public-opinion effects remain less mature.
21. WHAT WE DO NOT KNOW YET
R07 does not establish:
* population-scale political influence from persuasive AI dialogue;
* validated real-world turnout effects;
* validated election-result effects caused by AI persuasion;
* population prevalence of exposure to AI-generated misinformation;
* general persistence beyond current short follow-up horizons;
* repeated everyday-assistant effects on epistemic dependence;
* long-run source-diversity change across platforms;
* natural-field effects of AI labelling;
* population-level agenda-setting effects;
* mature organizational effects on campaigns or civil society;
* robust distributional effects by class, disability, migration or most language groups.
These are update triggers, not reasons to postpone v0.1.
22. GEOGRAPHIC AND EVIDENCE SCOPE
Primary evidence is concentrated in:
* the United States;
* the United Kingdom;
* parts of Europe and Canada.
Cross-national evidence also includes:
* Poland;
* China-origin and non-China-origin model comparisons;
* 48-market news surveys;
* a 14-language professional news audit.
Important limitations:
* most causal persuasion studies use paid, opt-in participants;
* newer high-capability evidence includes preprints;
* causal AI-search studies are U.S. and desktop-heavy;
* mobile and non-U.S. field evidence is thin;
* Chinese-language empirical discovery remains incomplete;
* organizational and long-run public-sphere effects lag behind interface-level evidence.
23. HOW TO READ FAST-MOVING CLAIMS
For this debate, model vintage and publication status are part of the claim.
Do not silently update:
* model performance;
* paper versions;
* corrections;
* expressions of concern;
* platform activation rates;
* legal status.
A corrected or superseded finding must remain visible as such.
24. WHAT WOULD MOVE THIS DEBATE?
The synthesis would materially change with:
* natural-setting persuasion studies with validated behaviour;
* direct platform exposure logs linking AI content to outcomes;
* longer follow-up and repeated-use studies;
* mobile and non-U.S. AI-search RCTs;
* cross-platform source-diversity telemetry;
* population-exposure estimates for synthetic political content;
* field tests of provenance and warning labels;
* a newer meta-analysis of frontier-model persuasion;
* longitudinal trust/dependence evidence;
* stronger non-English empirical coverage.
KEY EVIDENCE
* Hackenburg et al., Science 2025 — controlled persuasion / accuracy trade-off.
* Lin et al., Nature 2025 + Supplementary Information — cross-national persuasion RCTs.
* Salvi et al., Nature Human Behaviour 2025 + Author Correction 2026.
* Hölbling et al., Scientific Reports 2025 — meta-analysis.
* Hackenburg et al. 2026 — expert-human persuasion preprint.
* Hackenburg et al. 2026 — behavioural-action preprint.
* Chen et al., Journal of Experimental Political Science 2026 — exposure/scalability.
* Agarwal & Sen 2026 — AI Overviews field experiment.
* Wang et al. 2026 — AI search field experiment.
* Pew Research Center 2025 — browsing trace.
* Xu, Iqbal & Montgomery 2026 — AI Overview audit.
* EBU/BBC 2025 — multilingual news-integrity audit.
* Reuters Institute Digital News Report 2026 — chatbot news use.
* Li & Aral 2025 — trust/citation experiment.
* CETaS / Alan Turing Institute 2024 — election monitoring.
* Pan & Xu, PNAS Nexus 2026 — provider-origin audit.
* European Commission — AI Act Article 50 transparency implementation.
* Gallegos et al., PNAS Nexus 2026 — authorship-label experiment.
* ZAK / die medienanstalten 2026 — German media-regulatory action.
PAGE STATUS
LIVING DEBATE v0.1 — LOCKED.
Built from R07 HQ-adjudicated canonical claims only.
HOLD-only claims excluded.
Future changes require versioned amendment or explicit supersession.