Fact-checking has expanded substantially and the research on its effects is more encouraging and more limited than either its advocates or critics suggest.
Does it change beliefs
Meta-analyses of experimental studies find that corrections do improve factual accuracy of beliefs.
Which is a consistent finding, with effect sizes that are meaningful and modest.
Effects are larger for factual beliefs than for attitudes and for policy preferences.
The backfire question
Early research suggested corrections could strengthen the original belief among those disposed to it.
Which was widely reported and has largely failed to replicate in subsequent larger studies.
The current position in the literature is that backfire effects are rare rather than typical, which is a substantial revision.
The earlier finding continues to circulate, which is itself an example of a correction not reaching everyone.
Durability
Effects decay over time, with studies finding substantial reduction within weeks.
Which means a single correction has limited lasting effect, and repeated exposure performs better.
The reach problem
The larger practical limitation.
Corrections reach a small fraction of those who saw the original claim.
Which is a structural feature — false claims spread through networks that corrections do not follow, and studies have found false information spreading faster and further than true information on some platforms.
Platform interventions attaching labels or reducing distribution address the reach problem directly, and their effects are contested.
Prebunking
Warning people about manipulation techniques before exposure.
Which has shown promise in experiments, with effects on ability to identify manipulative content.
The advantage is that it addresses categories rather than individual claims, which scales better than checking each one.
Who does it
Independent organisations, newsroom units and platform partnerships.
Which raises questions about who checks the checkers, and codes of principles with signatory verification exist.
Transparency about methodology, sources and funding is what these codes require and what distinguishes serious operations.
The limits of the format
Many disputed claims are not straightforwardly factual.
Which includes contested interpretations, selective emphasis, misleading framing of true facts, and predictions.
Rating systems reducing these to true or false compress judgements that require explanation, and the better organisations acknowledge this.
Accusations of bias in selection — which claims get checked — are a persistent criticism and are difficult to assess without a defined universe of claims.
What individuals can do
Lateral reading — leaving a source to check what others say about it — has better evidence than evaluating a source in isolation.
Checking the original source of a claim rather than a report of it.
And noticing emotional response, since content producing strong reaction is more likely to be shared without checking.
These are teachable and have been shown to improve accuracy in evaluation studies.
Repetition
Repeated exposure to a claim increases belief in it, even where it was previously identified as false.
Which is the illusory truth effect and it is well replicated.
It creates a difficulty for correction — restating a false claim in order to correct it provides another exposure.
Guidance from researchers therefore recommends leading with the correct information rather than with the false claim.
Source credibility
Corrections from a source the audience trusts are more effective.
Which means the most credible corrector for a given audience is frequently not a fact-checking organisation.
Corrections from within a group are more effective than from outside it, which is a consistent finding.
Platform labelling
Warning labels attached to content have shown effects on sharing and on belief.
Which produced concern about an implied truth effect — unlabelled false content being assumed verified — with mixed evidence.
Generative content
Synthetic text, images, audio and video have made fabrication substantially cheaper.
Which shifts the problem from detecting known false claims to establishing provenance of material.
Provenance standards, cryptographically recording how content was created and edited, have been developed and adoption is partial.
Detection tools for synthetic content exist with limited and declining reliability as generation improves.
The liar's dividend
The concern that the existence of convincing fabrication allows genuine material to be dismissed as fake.
Which may be the more consequential effect, since it undermines evidence generally rather than promoting any specific falsehood.
Corrections in journalism
Publications correcting errors prominently are more reliable than those that do not, which is a useful heuristic.
What individuals actually do
Research on information behaviour finds most people spend very little time evaluating individual pieces of content.
Which means interventions requiring effort will reach few people, and design changes affecting the default experience reach everyone.
Friction — prompting people to consider before sharing — has shown effects in experiments at low cost.
Where the responsibility sits
Individual media literacy, platform design and regulatory requirements are the three levers.
Which operate at very different scales, and expecting individuals to solve a systemic problem places the burden where it can do least.