Information Integrity

Established Information integrity is a chain, and almost all the measurement sits at the first link. The chain runs production, exposure, amplification, belief, behaviour, trust. The field can tell you, to three significant figures, how well a classifier separates true from false headlines. It can tell you very little about what share of a population saw a given falsehood, what the recommender did to that share, whether belief moved, and whether anything anybody did offline changed.

This is a synthesis slot and it owns the chain. Future Democracies owns the input side of democratic authority and already records the generative-AI-and-elections verdict; its organising finding is that every measured failure in that field is a failure of transmission. Civic Technology owns citizen-input platforms and found a field that instrumented its inputs exhaustively and its outputs almost not at all. AI Governance owns the regulatory instruments. Information integrity is the same shape a third time, and nobody on this map owns the join: supply is measured to death, transmission is not measured at all, and the interventions with the best evidence move the smallest quantities.

Established The single most useful reframing in this literature is that false content is not scarce and attention is. Once exposure is measured rather than assumed, misinformation turns out to be a small, extremely concentrated share of a media diet dominated by television and by ordinary partisan news. That does not make it harmless. It relocates the harm: from persuasion at scale to a small heavy-exposure tail, to elite endorsement that gives a falsehood standing, and to the corrosion of the belief that anything can be verified at all.

Established A note on sourcing. This brief was commissioned in September 2026 from the Institute’s research base. Reading-list entries without links are cited from the bibliographic record rather than re-fetched, and claims are dated no later than early 2026 unless carried by a linked source.

Established Exposure to untrustworthy sources is small in aggregate and extremely concentrated. The three measurement studies that defined the field agree across different methods and platforms. Web-tracking of American panels in 2016 put untrustworthy websites at roughly six per cent of news consumption, with about sixty per cent of those visits made by the ten per cent of people with the most partisan information diets. A media-diet accounting exercise put fake news at roughly a tenth of a per cent of Americans’ daily media consumption, in an ecosystem where news of all kinds is about a seventh of it and television dominates. A Twitter panel study found one per cent of individuals accounting for about eighty per cent of exposures to fake-news sources, and a tenth of a per cent accounting for about eighty per cent of shares.

Established Concentration is the finding, not a caveat on the finding. A low mean with a heavy tail is a different policy problem from a high mean. It means population-wide interventions are aimed at people who were never exposed, that the exposed tail is defined by demand rather than by accident of feed, and that any evaluation reporting an average treatment effect over a general sample is measuring a dilution of the effect it cares about.

Frontier The 2024 review that consolidated this reading is itself contested, and the dispute is about scope rather than arithmetic. A high-profile 2024 synthesis argued that the harms of online misinformation are systematically overstated: exposure is low, effects are small, and causal claims routinely outrun designs. Critics do not dispute the exposure numbers. They argue that the measured object — links to a fixed list of untrustworthy domains, mostly in the United States, mostly in English, mostly on the open web — excludes closed messaging, video, influencer content and the misleading use of true material, which is where much of the problem now lives. Both sides are right about what they measured.

Established Removing the algorithm does not do what either side of the argument expects. The largest platform-scale experiments ever run on this question, conducted with researcher access during the 2020 United States election, replaced ranked feeds with reverse-chronological ones, cut reshared content, and reduced like-minded content for consenting users over about three months. Chronological feeds sharply reduced time on platform and engagement. None of the manipulations produced statistically significant changes in affective polarisation, issue polarisation, political knowledge or self-reported turnout. The result is a genuine constraint on the algorithmic-amplification story at the horizon tested, and its limits are real: three months, consenting participants, one country, one election, and a platform co-authoring its own evaluation.

Established Exposure is asymmetric even where the algorithm is not obviously the cause. The same research programme found audiences for content rated false by third-party fact-checkers to be far more ideologically homogeneous than audiences for news generally, and skewed heavily toward the political right. Separately, an internal study of one platform’s ranked timeline found mainstream right-leaning parties and outlets amplified more than left-leaning ones in six of seven countries examined, with no mechanism identified. Asymmetry is measured; its cause is not.

Established Corrections work on facts and fail on the thing that matters. The backfire effect — the claim that corrections entrench the false belief — largely failed to replicate across tens of thousands of subjects and dozens of issues; corrections move factual beliefs modestly and reliably, including outside rich Western samples. The awkward result is the next one: in studies where a correction demonstrably moved a factual belief about a politician, it did not move evaluations of that politician or intention to vote for them. Belief and support are only loosely coupled, which is the single most important finding for anyone designing a fact-checking programme and the one least represented in their theories of change.

Frontier Accuracy prompts and prebunking are real effects at the small end, and the effect sizes are contested. A field experiment on Twitter found that a single accuracy prompt improved the quality of subsequently shared links by a margin usually reported in the five-to-ten per cent range. Psychological inoculation against manipulation techniques produced improvements of a few percentage points in a large video-advertising field test. A serious methodological objection stands against the inoculation work: a 2023 analysis argued the gamified interventions shift response bias — making people call more things fake — rather than improving discrimination between true and false, which would make the headline gains an artifact of the scoring. The dispute is about signal-detection modelling, it is not resolved, and it applies to a family of interventions that has received substantial public funding.

Established Crowd-sourced annotation is the intervention that scaled, and its binding problem is latency. Community-note systems on the largest platforms reduce resharing of annotated posts by amounts their operator has reported in the twenty-to-thirty-per-cent range. Independent analyses find that the median note attaches long after the bulk of engagement has already occurred, so the measured per-post effect is applied to a small residual of the audience. An effective intervention that arrives after the exposure it targets is an effective intervention with almost no population effect, and this is the clearest instance in the field of a metric measured at the wrong point in the chain.

Established The 2024 generative-AI election panic underdelivered, and saying so is a finding rather than a reassurance. Across an exceptional global election year, the systematic reviews that went looking found small numbers of confirmed AI-enabled incidents, no documented case of an election outcome changed, and influence operations that generated content easily but distribution almost never. One platform reported AI-generated material at under one per cent of fact-checked election misinformation on its services — the company’s own figure about its own products. The mechanism that failed was distribution, not generation, and nothing about the next election cycle guarantees it keeps failing.

Frontier The measurable harm from synthetic media in 2024 was mostly the reverse of the one predicted. Documented cases cluster around the liar’s dividend: politicians and officials denying authentic recordings by calling them AI-generated. Researchers have catalogued a growing set of such denials. This inverts the intervention logic. Detection and provenance tools that tell you something is synthetic do not help against a claim that something authentic is synthetic, unless the authentic item carries a credential — which almost none do.

Frontier Detection degrades in the wild and provenance is not yet an answer. A 2024 in-the-wild benchmark of deepfakes actually circulating found open detector performance falling sharply against the distribution real moderators face, relative to curated academic benchmarks. On the provenance side, the content-credential stack has working specifications and adoption across cameras and generative services, and it also has published attacks: a scheme-agnostic universal attack on defensive watermarking, a camera exploit that led a manufacturer to revoke signing certificates, a demonstration of a mobile platform’s secure element used as a signing oracle for arbitrary images, and an academic critique arguing the specifications do not deliver what deployments claim for them. Provenance answers “where did this come from”, never “is this true”, and a signed lie is a signed lie.

Frontier Trust is falling in a pattern that misinformation does not explain. Trust in news has declined and news avoidance has risen across many rich democracies on broadly similar trajectories despite very different information environments and platform regulations. The main cross-national study of affective polarisation finds it rising sharply in the United States and flat or falling in several comparable countries over the same decades of internet adoption. A cause common to all of these countries cannot explain an outcome that diverges between them. The measurement of democratic decline is itself disputed, as the backsliding literature records.

Frontier What is the exposure denominator, per platform, per country, per year? Every effect estimate in this brief is a rate over a population nobody can count outside the United States, outside English, and outside the open web. Until the denominators exist, cross-national policy is being set on American panel studies from a platform that no longer provides the data.

Frontier Does the algorithm matter over years rather than months? The three-month null on polarisation is the strongest evidence in the field and it tested the shortest plausible horizon. Habit formation, network composition and creator incentives all operate over longer periods and were held constant by the design. Nobody has run the multi-year version and it is not obvious anybody can.

Frontier Do interventions survive contact with a month? Accuracy prompts decay within days, inoculation effects decay within weeks absent boosters, and the field measures at the point of maximum effect. A durability literature exists in miniature and is uniformly discouraging.

Speculative Is elite cue-giving the actual causal channel? The pattern fits: belief follows endorsement, corrections move facts but not evaluations, and cross-national divergence tracks party systems better than platform penetration. The strong version — that online misinformation is mostly downstream of elite behaviour — is coherent, consistent with the data, and has never been tested by an intervention on elites, because nobody can randomise a party leader.

Established Independent measurement infrastructure has been dismantled faster than it was built. A major platform’s public transparency tool was retired in 2024 and replaced by a restricted successor; another platform priced academic interface access out of reach in 2023; prominent university programmes studying influence operations were wound down or defunded. Several of this brief’s foundational results could not be reproduced today with the access available.

Frontier Compelled data access exists on paper in one jurisdiction and is young. The European regime creating vetted-researcher access to non-public platform data completed its implementing rules in 2025 and is only now producing applications and refusals. Whether it delivers usable data at usable latency, and whether it extends to non-European research questions, are open empirical questions about the regulation itself.

Established The outcome variable is missing, which is the same defect Civic Technology documents. Almost no study in this field carries an intervention through to a behaviour: a vote, a vaccination, a donation, a protest. Sharing intentions and self-reported belief are the standard endpoints because they are cheap, and the field has treated them as proxies without demonstrating that they are.

Frontier Classifier accuracy is the field’s most improved and least useful quantity. Detection performance on curated benchmarks has risen steadily while in-the-wild performance falls, and even a perfect classifier addresses only the production link. The persistent error is to treat improvements in supply-side detection as progress on an outcome nobody is measuring.

Established This topic waits on an institutional result, not a scientific one. The binding input is compelled, standardised, cross-platform data access with a latency short enough to study live events. Everything downstream — denominators, amplification audits, durability measurement, cross-national comparison — is straightforward social science once the data exist and impossible without it.

Frontier It also waits on a measurement dispute owned elsewhere. Future Democracies records that the claim of democratic decline is a live measurement argument rather than a premise. Information-integrity policy is routinely justified by that decline, which means its principal warrant is contested in a literature this brief does not own.

Speculative The provenance route depends on a supply chain nobody controls. Credentials at capture require attested signing in cameras and phones, key management across manufacturers, and preservation through every editing and re-encoding step. The published attacks are against exactly these joints, and no part of this is a research problem.

Five results would move this brief’s assessment, in descending order of how much.

Frontier The decisive experiment is a platform-scale field trial that carries one intervention all the way through the chain and reports durability at six months. One randomised treatment — an accuracy prompt, an inoculation, a note delivered at first exposure rather than after — assigned at the account level, with exposure, sharing, belief and a pre-registered downstream civic behaviour measured on the same subjects, and the effect reported at one week, one month and six months. Every number in this field is currently an effect at the point of maximum effect. The trial needs no new method and no new science; it needs platform cooperation, which the European access regime can compel and no other instrument can.

Established The natural experiment is already running and is the cheapest thing in this brief to observe. Two jurisdictions have diverged sharply: one compelling systemic-risk assessment, audit and researcher access, the other withdrawing platform transparency, disbanding research programmes and retreating from public fact-checking arrangements. Both have well-instrumented survey series on trust, avoidance and polarisation. Whoever publishes the difference-in-differences with a defensible identification strategy will have produced the most policy-relevant result in the field.

Established A latency experiment on crowd-sourced notes is a single product change away. Randomise the note-attachment threshold so that a share of posts receives annotation in minutes rather than hours, and report the effect on downstream reach. It directly tests whether the intervention’s population effect is limited by efficacy or by timing, which is the difference between a good tool and a good tool applied too late.

Frontier The inoculation dispute is settleable and nobody has settled it. Re-run the gamified interventions with a design that separates discrimination from response bias by construction, pre-registered, with true and false items balanced and the signal-detection parameters reported separately. Either the effect survives as improved discrimination or it does not, and a substantial public-funding programme rests on which.

Speculative An elite-cue field experiment is the one that would test the strongest hypothesis and cannot be run as stated. The tractable substitute is a natural experiment on party-level position changes with panel data on the same respondents before and after, which would at least bound how much belief movement follows endorsement rather than content.

Established Provenance is the only part of this subject with a real engineering stack, and it is a chain of custody rather than a truth machine. Capture-time signing, manifest preservation through editing, verification at display, and revocation when a key is compromised. Each joint has a published failure: a camera exploit that forced certificate revocation, a mobile secure element driven as a signing oracle, and a specification critique arguing the guarantees claimed in deployment are not the guarantees the documents provide.

Frontier The hard requirement is verification at the point of display, not at the point of capture. Credentials survive nothing that strips metadata, which is most of the social web. Until the major distribution surfaces render provenance in the interface by default, signing is a producer feature with no consumer.

Frontier Watermarking is the weaker sibling and should be described as such. Statistical marks in generated media are removable by scheme-agnostic attacks published in the security literature, and the incoming European marking obligations for synthetic content do not specify a robustness standard. A marking duty without a robustness test is an auditable compliance artefact, not a defence.

Established The unglamorous engineering that would matter most is measurement plumbing: stable per-country exposure denominators, sampled impression logs available to vetted researchers at short latency, and a standard schema so results compare across platforms. None of this is hard. All of it is contractual.

Established Future Democracies owns the input side and the 2024 election verdict, and this brief defers to it on both. Its transmission finding — that failures occur between the room and the decision, while remedies address production — is the template this brief applies to the information environment.

Established Civic Technology owns the outcome-measurement deficit in citizen-facing platforms and reached it independently. Two fields arriving at the same defect by different routes is evidence about how participation research is funded, not a coincidence.

Established AI Governance owns the instruments — risk classification, transparency duties, marking obligations for synthetic content — and this brief supplies the evidence base those instruments are usually justified by.

Frontier Two further neighbours carry pieces. Digital Citizenship owns identity and the authentication layer that any non-anonymous provenance scheme depends on; Collective Intelligence owns the aggregation question that crowd-sourced annotation is an instance of, including when a crowd beats an expert panel and when it does not.

Established The intergovernmental framing settled on three pillars, and the third is the one nobody has built. The 2024 OECD treatment organises the field into transparency and accountability of platforms, societal resilience, and governance architecture — the institutional capacity to coordinate, evaluate and hold anyone to account. The first two have programmes and budgets in most member states. The third is where the evaluations that would tell anyone whether the first two work were supposed to live, and it is largely empty.

Established Europe made transparency a legal duty and is discovering what that costs. The platform regime in force since 2024 requires systemic-risk assessment, independent audit, advertising repositories and a researcher data-access route, with public risk reports and audit opinions now published annually. Enforcement proceedings against the largest platforms have run since 2023 on advertising transparency, dark patterns and data access. This is the only jurisdiction where the counterfactual question can even be asked, and its own evaluation is not yet possible because the access rules only began operating in late 2025.

Frontier The United States moved the other way on every axis at once. Public research programmes on influence operations were defunded or disbanded, a major platform ended third-party fact-checking in its largest market in favour of crowd-sourced notes, and litigation over government contact with platforms chilled the coordination channel without resolving the constitutional question. The result is a live natural experiment in regulatory divergence, which this brief treats as its cheapest available evidence.

Frontier Crowd-sourced annotation is now load-bearing institutional infrastructure and was never evaluated as such. It was adopted at scale as a replacement for professional fact-checking on the strength of operator-reported engagement effects, not on the strength of an independent evaluation against the thing it replaced. That is a governance decision resting on an interested party’s figures about its own product.

Speculative The institution that does not exist is an evaluator with standing. There is no body anywhere with a statutory mandate, compelled data access and the independence to publish that a platform intervention did not work. Every existing arrangement is either a regulator without analytic capacity, a researcher without data, or a platform grading its own work.

Established Every instrument in this field is a speech instrument, and the strongest argument against most of them is not that they fail but that they work. A system capable of suppressing a false claim at scale is capable of suppressing a true one, and the historical record of contested claims later vindicated is long enough that confidence at the moment of moderation is a poor guide.

Frontier The evidence base itself cuts against aggressive intervention. If exposure is small and concentrated, corrections move facts but not support, and algorithmic changes do not move polarisation at three months, then the expected benefit of broad content intervention is modest while its expected cost to legitimacy is not. The honest reading of this literature is more permissive than the policy it is used to justify.

Established Provenance has a privacy cost that its advocates under-report. A credential that proves where an image came from is a credential that can identify who made it. Source protection, whistleblowing and documentation of abuse all depend on the ability to publish authentic material without attribution, and a world where unsigned media is presumed false is a world where anonymous evidence is presumed fake.

Frontier Asymmetric enforcement is a measured fact with no agreed remedy. Content rated false by fact-checkers is consumed asymmetrically along partisan lines, so any even-handed enforcement produces uneven outcomes, which is then read as bias. There is no technical fix for this, and pretending the asymmetry is an artifact of rater bias is not supported by the data either.

Frontier The durable risk is not deception; it is the collapse of the default that evidence settles anything. The liar’s dividend is the measurable early form: once any recording can be dismissed as synthetic, the cost of denial falls to zero. That degrades courts, journalism, insurance and accountability mechanisms simultaneously, and none of them has a replacement.

Speculative The plausible equilibrium is authenticated islands in an unauthenticated sea. Institutions that need evidentiary weight — courts, regulated media, scientific publishing, insurers — adopt credential chains, and everything outside them is treated as unverifiable entertainment. That is a functioning world and a stratified one, in which verification becomes a paid service and the ability to be believed tracks institutional membership.

Speculative The economic layer is upstream of everything here and is usually treated as background. The local reporting that produced verifiable primary information collapsed with its advertising base, and no instrument in this brief replaces it. A strategy that funds detection and literacy while the production of verifiable fact contracts is optimising the wrong term.

Handwave Claims that synthetic media ends shared reality are asserted rather than argued. Shared reality has survived photographic manipulation, broadcast propaganda and a century of documented state forgery. The argument that this time the quantity changes the kind is coherent and has, so far, failed its first large empirical test.

These horizons track when the assessment here could change, not when the technology improves.

• 10 yr: Frontier The measurement question resolves or is abandoned. The European access regime either produces usable non-public data at workable latency or is shown to be unusable; the transatlantic natural experiment yields its first credible difference-in-differences; the synthetic-media marking obligations either acquire a robustness standard or remain a compliance artefact. This is the decade in which the field either gets denominators or admits it is not going to.
• 25 yr: Speculative Provenance either becomes default infrastructure rendered at the point of display, or is abandoned as a producer-side feature with no consumer. The stratified outcome — authenticated institutional islands, unauthenticated everything else — is the most likely intermediate state and is compatible with both.
• 50 yr: Speculative Whether the loss of a shared evidentiary default is permanent or is absorbed the way earlier media disruptions were absorbed. The variable is whether institutions with adjudicative authority retain the ability to certify facts to a public that accepts the certification.
• 100 / 250+ yr: Handwave Claims about permanent epistemic fragmentation or a restored consensus reality rest on assumptions about institutional legitimacy that no current evidence constrains. They are labelled here, not assessed.

• Depends on Two results other briefs on this map produce. Future Democracies owns the democratic-backsliding measurement dispute that supplies this field’s principal warrant, and the generative-AI-and-elections verdict this brief inherits rather than restates. Civic Technology owns outcome measurement with a counterfactual, which is the same missing capability under a different name. Neither is a technical blocker.
• Requires (not on this map) A researcher data-access regime that survives a platform changing its mind, since the results this brief rests on were produced with access that has since been withdrawn. An exposure denominator published per platform and per country, without which every effect estimate is a rate over an unknown population. An audit standard for recommender amplification, because asymmetric amplification has been measured and no agreed method exists to attribute it. A field measurement of intervention durability beyond one month, since the entire intervention literature reports effects at the point of maximum effect. A capture-to-display chain of custody for media credentials, whose published failures are at the joints rather than the cryptography. And a funding base for local reporting independent of platform referral, because the supply of verifiable primary fact is upstream of every instrument here and is the one quantity nobody in this field is trying to fix.
• Enables Working measurement here would let AI Governance calibrate transparency and marking duties against evidence rather than against anticipated harm, and would give Future Democracies a tested channel between deliberative output and public understanding. No typed enabling edge is claimed, and that is the finding: no intervention in this brief has demonstrated a durable downstream effect, so nothing here can responsibly be depended on.
• Adjacent Civic Technology, which shares the missing outcome variable exactly; Digital Citizenship, which owns the identity and authentication layer provenance schemes assume; Collective Intelligence, which owns the aggregation theory behind crowd-sourced annotation; and AI-Assisted Governance, which shares the problem of an administrative system grading its own outputs.

Established “Misinformation is everywhere online.” Measured as a share of media diet it is not: roughly a tenth of a per cent of daily consumption in the best accounting study, roughly six per cent of news consumption in the best web-tracking study, with the bulk of it reaching a small tail of heavy consumers. The claim is usually supported by counts of posts, which is a supply statistic in an ecosystem where attention, not content, is the scarce good.

Established “The algorithm is radicalising people.” The largest experiments ever run on this question replaced the ranking system for tens of thousands of consenting users for three months and found no significant movement in polarisation, knowledge or turnout. The honest statement is that the algorithm did not move these outcomes at that horizon, in that country, in that election — which is a much narrower claim than either side wants and is still the best evidence there is.

Frontier “Fact-checking does not work; it backfires.” The backfire effect largely failed to replicate. Corrections move factual beliefs reliably and modestly, across countries. What they do not move is support for the politician the correction concerns, which is a different and more uncomfortable failure than the one usually alleged.

Frontier “Prebunking is the proven solution.” It is a real effect of a few percentage points in field conditions, with a live methodological objection that the gamified versions shift response bias rather than improving discrimination, and with decay measured in weeks. It is the most promising family in the literature and it is not proven.

Established “AI deepfakes swung the 2024 elections.” The systematic reviews found small numbers of confirmed incidents and no documented outcome change; the operations that used generative tools produced content easily and distribution almost never. The correct lesson is about the bottleneck, not about safety: distribution was the constraint, and a cheaper distribution channel would change the answer without anything else changing.

Frontier “Provenance and watermarking will fix synthetic media.” Provenance establishes origin, not truth; credentials are stripped by ordinary platform processing; watermarking has published scheme-agnostic attacks; and the documented 2024 harm ran the other way, as authentic material denied as synthetic. A credential system helps most against the liar’s dividend and only if authentic material carries credentials by default, which it does not.

Handwave “We are entering a post-truth era without precedent.” The claim is made without a measured baseline, in a period when trust in news is falling in countries with opposite platform regimes and polarisation is diverging between them. A cause common to all cannot explain an outcome that differs between them, and the assertion of unprecedentedness is doing work that the evidence does not.