Fact-checking

Fact-checking is the systematic examination of factual claims against verifiable evidence. In journalism, it includes both checking material before publication and publishing assessments of statements already circulating in public. Its central task is to establish what the available evidence supports, contradicts, or leaves unresolved, rather than to judge opinions or infer a speaker’s intentions solely from an inaccurate statement. (time.com)

Fact-checking matters to the reliability of published information and the scrutiny of public claims. Its methods include establishing precise wording and context, tracing original sources, examining data, consulting specialists, and explaining conclusions. Professional standards emphasize independence, transparent methods and funding, accessible evidence, and visible corrections. Experimental research has found that fact-checks can improve factual accuracy, although the persistence of those effects and their relationship to attitudes or behaviour require separate investigation. (africacheck.org)

The field extends from editorial research departments to specialist organizations and computational tools. Natural language processing and information retrieval can assist with identifying claims, finding earlier checks, and retrieving evidence; such systems introduce additional questions about context, explanation, and evaluation. Understanding fact-checking therefore requires attention to its publishing history, evidentiary practices, institutional standards, distribution systems, and measurable effects. (time.com)

Scope and boundaries

Fact-checking evaluates assertions about the world by comparing them with verifiable evidence. A factual claim is an assertion whose accuracy can, in principle, be investigated through records, observations, measurements, or attributable testimony. It need not be true to qualify as factual: an incorrect statement about a population total remains a factual claim. Verification therefore concerns the relationship between an assertion and the evidence, rather than the speaker’s authority or the assertion’s popularity. Africa Check’s published method makes this process explicit: a report identifies the claim and its context, presents supporting and contradictory evidence, and explains the resulting conclusion. This makes the reasoning available for examination rather than merely substituting a checker’s assurance for the original speaker’s. (africacheck.org)

The boundary between factual and evaluative language is central. Checking whether a government spent a stated amount does not establish whether that expenditure was morally justified, politically desirable, or persuasively defended. Those judgments involve values and arguments beyond the accuracy of the amount. A policy argument may nevertheless contain separately checkable premises. For example, in the hypothetical statement “this programme should be abolished because its cost doubled,” the recommendation and the historical cost comparison require different treatment. Africa Check accordingly distinguishes statements of fact from ideas, philosophies, and arguments, while assessing factual statements in their surrounding context. (africacheck.org)

Within journalism, prepublication checking is editorial quality control applied before a publisher releases material. It may involve confirming names, dates, quotations, and factual descriptions, with discrepancies resolved through changes to the draft. Nancy Ford, hired by TIME in early 1923, performed such checks on completed articles. Post-publication checking, by contrast, makes an already circulated assertion the subject of a new published investigation. The distinction is therefore not simply when checking occurs, but what it produces: a revised original work in the first case, and a public examination of an existing claim in the second. Journalism scholar Lucas Graves described this difference in mission in TIME’s 2017 account of the practice. (time.com)

Confirmation, correction, and debunking are related but distinct outcomes. Confirmation establishes that available evidence supports an assertion; correction replaces an erroneous detail or account; debunking exposes a false claim or explanation. Public-facing checking can also conclude that a claim is accurate, qualified, misleading, or unresolved, rather than refuting it. Unsupported is not synonymous with false: failure to supply adequate evidence does not itself demonstrate the opposite proposition. Africa Check’s “unproven” category expressly covers statements that publicly available evidence neither proves nor disproves. Its assessments are tied to the evidence available at the relevant time, not a claim of permanent certainty. (africacheck.org)

Factual accuracy must also be distinguished from an intention to deceive. Establishing that a statement conflicts with the evidence does not, by itself, establish what its speaker knew or believed when making it. Michelle A. Amazeen’s assessment of political fact-checking records FactCheck.org director Brooks Jackson’s explanation that the organization avoided calling statements lies because it could not determine speakers’ intentions. A finding of inaccuracy consequently concerns the statement; an allegation of deliberate deception requires additional evidence about knowledge and conduct. (static.newamerica.org)

Historical development

Dedicated magazine checking developed in the early twentieth-century United States as a specialized division of editorial labour. TIME, established by Henry Luce and Briton Hadden in 1923, employed Nancy Ford initially to select and clip newspaper material; her responsibilities soon expanded to verifying names, dates, and other facts in completed articles. Reference works and inquiries to the New York Public Library supplied evidence, while checkers were expected to challenge writers and editors before publication. The New Yorker, founded in 1925, introduced rigorous checking in 1927 after publishing an inaccurate profile of the poet Edna St. Vincent Millay. These arrangements made verification a distinct production stage rather than leaving it entirely to the author. (bunkhistory.org)

This division of labour was strongly gendered. TIME’s early researchers were women, whereas its writers and editors were predominantly men; the researcher role remained exclusively female until 1973. Researchers both assembled material for writers and checked the resulting text after editing. Their authority to question assertions therefore coexisted with occupational segregation and limited access to writing and editing positions. Mary Fraser, an early head of TIME’s research department, later described a clear boundary between writers and researcher-checkers. The institutional history of magazine verification consequently includes work performed by women whose contributions were less publicly visible than those of bylined authors. (time.com)

Public-facing verification developed alongside, rather than simply replacing, this internal editorial practice. During the 1992 United States presidential campaign, Brooks Jackson produced CNN “adwatch” and “factcheck” reports examining political statements and advertising. Another strand emerged through the Internet: Snopes began in 1994 as a project investigating urban legends, allowing readers to seek explanations of questionable stories encountered online. Here the disputed story itself became the subject of publication, rather than merely something corrected during production of another article. (factcheck.org)

Dedicated political websites subsequently made such scrutiny a continuing beat within journalism. FactCheck.org launched on December 5, 2003, after Kathleen Hall Jamieson recruited Jackson to investigate claims surrounding the 2004 elections through the University of Pennsylvania’s Annenberg Public Policy Center. Its initial operation consisted of Jackson and one researcher. PolitiFact followed on August 22, 2007, with Bill Adair as its founding editor and the Truth-O-Meter as a recognizable presentation format. By June 25, 2008, it had published 500 ratings. These projects organized published investigations around individual assertions rather than treating verification only as an occasional feature of campaign reporting. (factcheck.org)

Institutional development also took place outside the United States. Peter Cunliffe-Jones founded Africa Check in Johannesburg in 2012, under the auspices of the University of the Witwatersrand’s journalism department; Ruth Becker became its founding editor. Its regional expansion included a Dakar office in October 2015 and a French-language website the following month, followed by offices in Lagos in November 2016 and Nairobi in January 2017. This established reporting capacity across several African regions and extended publication beyond English. (africacheck.org)

Cross-border networks added a further institutional layer. The International Fact-Checking Network was created at the Poynter Institute in 2015; its October 21 online launch announcement described a forum spanning five continents, with research, training, and collaborative investigations among its functions. Africa Check launched the Africa Facts network in November 2017, bringing African checking initiatives together to exchange experience and collaborate. In May 2021, that network comprised 20 organizations, illustrating the development of regional cooperation alongside individual national newsrooms and independent projects. (poynter.org)

Claim selection and formulation

Claim selection combines monitoring with editorial judgment. Fact-checkers examine speeches, interviews, advertisements, news reports, public documents, campaign literature, and posts on social media, alongside audience submissions. Africa Check begins by assessing public statements and reader suggestions for importance and whether they assert facts rather than opinions. PolitiFact’s methodology, published in 2018, identifies significance, verifiability, and likelihood of repetition among its selection criteria; it excludes trivial verbal slips. Full Fact prioritises claims with potential to harm people’s lives, considering widespread online circulation, prominence in broadcast or newspaper coverage, and repeated use by political candidates. Circulation therefore helps establish priority, but is not itself evidence that a claim is inaccurate. (africacheck.org)

Before substantive verification, the object of inquiry must be established: the precise statement, its speaker or publisher, the date and setting, and the surrounding passage. Africa Check explicitly separates selecting a claim from confirming what was actually said, noting that audience submissions can be vague. Recording the original wording helps distinguish an assertion from a reporter’s paraphrase or an audience member’s recollection. Context also determines the intended referent: whether “the country” means a particular jurisdiction, for example, or whether “last year” denotes a calendar year or another reporting period. These distinctions formulate the proposition to be tested rather than deciding its accuracy in advance. Africa Check’s method requires attention both to exact wording and to the circumstances in which it appeared. (africacheck.org)

A single sentence may contain several independently checkable assertions. Full Fact’s December 2019 discussion of British election manifestos illustrates this with Labour’s statement that NHS England had 100,000 staff vacancies, including a shortage of 43,000 nurses. The overall total and the nursing component were distinct numerical propositions; interpreting “vacancies” also mattered, because a vacant permanent post could be covered by temporary staff. Compound statements can likewise combine a quantity with an explanation of its cause. Formulation must preserve that distinction: evidence for a numerical change does not automatically establish the asserted causal relationship. Full Fact consequently examines underlying assumptions as well as explicitly stated evidence. (fullfact.org)

Ambiguous language and exaggeration require attention to ordinary meaning without substituting a different claim. PolitiFact recognises rhetorical latitude for hyperbole while requiring an underlying verifiable fact. Africa Check similarly emphasises significant inaccuracies rather than immaterial discrepancies. An argument’s factual premises may be checked without rating its political or moral conclusion: the claimed cost of a policy is empirically assessable, whereas whether that cost is acceptable involves a value judgment. Africa Check’s stated practice excludes rating ideas, philosophies, or arguments as such, while retaining scrutiny of assertions presented as facts. (politifact.com)

Checkability also has a temporal limit. In its 2019 manifesto analysis, Full Fact distinguished present or historical quantities from promises and predictions, whose eventual fulfilment could not yet be verified. It could examine the forecasting basis of a projected £50 billion economic benefit from remaining in the European Union without treating the projection as an observed outcome. A different limitation arises when evidence about an existing state of affairs is insufficient: Africa Check distinguishes claims that available evidence neither proves nor disproves from demonstrably inaccurate statements. Formulating a check therefore includes identifying whether the question concerns an established event, an uncertain forecast, or a factual assertion for which a conclusion must remain open. (fullfact.org)

Evidence gathering and source evaluation

Evidence gathering ordinarily includes asking the claimant for the material supporting the assertion. This request can identify an otherwise unspecified report, calculation or interview and gives the claimant an opportunity to substantiate the statement. Africa Check includes this step explicitly in its published methodology; FactCheck.org likewise seeks supporting material before undertaking its own research. The International Fact-Checking Network (IFCN) qualifies the procedure: contact may be impracticable or unsafe, and a failure to reply should not prevent the investigation from proceeding. A claimant’s response supplies evidence to examine, rather than determining the outcome of the check. (africacheck.org)

Research then traces the assertion to primary sources: original documents, public records, datasets, research reports and attributable testimony. FactCheck.org’s documented sources illustrate the specificity involved: congressional testimony from the Library of Congress, roll-call records from the House Clerk and Senate Secretary, employment data from the Bureau of Labor Statistics, and corporate filings from the Securities and Exchange Commission. These records answer different questions. A transcript establishes what a witness said, while an employment dataset supplies evidence about employment; neither automatically substantiates every interpretation attached to it. The IFCN therefore requires the best available primary evidence where suitable material exists, and an explanation when secondary sources are used instead. (factcheck.org)

Evaluation concerns both relevance and provenance—where evidence originated and how it reached the checker. Africa Check examines the limitations and context of the data it obtains, rather than treating publication by an authoritative institution as sufficient. The IFCN also requires checking key elements against more than one named source, except where only one relevant source exists. This distinction makes source independence important: three articles repeating the same press release represent three publications, but only one underlying evidentiary source. Comparing their conclusions cannot substitute for examining the original release and seeking separately generated evidence. Relevant interests must also be disclosed when they could reasonably affect the evidence a source supplies. (africacheck.org)

Subject specialists help interpret the resulting material rather than replace it with an appeal to authority. Africa Check consults specialists after assembling evidence and context, and quotes only experts willing to speak on the record. FactCheck.org similarly uses topic-specific expertise and discloses relevant biographical information, including previous government or campaign work where applicable. Expert consultation can clarify the scope of a finding or the limitations of an analysis; the published reasoning must still connect the evidence to the particular assertion being checked. (africacheck.org)

The report also makes evidentiary gaps visible. FactCheckZW, for example, identifies restricted access to primary data as a practical constraint and explains its resulting reliance on agency summaries, international organizations and research publications. Its methodology calls for tracing claims as close as possible to their original source and publishing references that readers can follow. More generally, the IFCN requires sufficient source identification for readers to replicate the investigation, with a limited exception where disclosure would endanger a source. Quotations, document references and accessible supporting material thus provide an inspectable evidentiary trail, while explicit statements about unavailable evidence show where that trail ends. (factcheckzw.org)

Numerical and scientific claims

Numerical checking begins by establishing what a figure measures, rather than merely reproducing its arithmetic. Relevant specifications include the definition of the variable, units, numerator and denominator, population covered, geographical boundaries, and reference period. Counts, proportions, and rates answer different questions: a national total cannot be compared directly with a provincial rate, nor a monthly change with a twelve-month change. In economic reporting, the US Bureau of Labor Statistics distinguishes consumer-price inflation from changes in producer prices, import prices, and employment costs; these indicators are not interchangeable measures of “inflation.” Comparisons therefore require examination of the underlying series and methodology, not just matching labels in headlines. (bls.gov)

Surveys, censuses, and derived estimates have different limitations. Survey estimates depend on sampling design, weighting, response patterns, and question interpretation. Published margins of error describe sampling uncertainty, not every possible source of error: undercoverage, nonresponse, inaccurate recall, and processing mistakes can introduce additional distortions. The US American Community Survey publishes margins of error at the 90% confidence level, and the Census Bureau requires statistical testing when comparing its estimates; a difference between two displayed numbers does not by itself establish a meaningful population change. A census also requires quality assessment: the Bureau’s Post-Enumeration Survey independently samples the population to assess census coverage. Thus, an enumerated count is not automatically an error-free benchmark against which every survey must agree. (census.gov)

Economic claims additionally require attention to adjustment methods and data versions. Seasonally adjusted figures remove recurring seasonal patterns to make short-term movements more interpretable, whereas unadjusted figures retain those patterns. The Bureau of Labor Statistics recalculates consumer-price seasonal factors annually and may revise the preceding five years of seasonally adjusted indexes. A historical claim must consequently be assessed against both its reference period and the information available when it was made; a later revision need not imply that the original statement misrepresented the contemporary release. Arithmetic distinctions also matter: hypothetically, a rate rising from 4% to 6% increases by two percentage points but by 50% relative to its initial value. These are different descriptions of the same change. (bls.gov)

In climate reporting, the baseline, averaging period, spatial coverage, and uncertainty range constrain comparisons. The Intergovernmental Panel on Climate Change’s 2023 synthesis reported that global surface temperature in 2011–2020 was $1.09\,^\circ\mathrm{C}$ above the 1850–1900 average, with an assessed range of $0.95$–$1.20\,^\circ\mathrm{C}$. It separately reported greater warming over land than over oceans. A local temperature observation, a single-year global anomaly, and this decadal global estimate therefore cannot be substituted for one another. Checking such claims requires preserving the report’s specified quantity and period, including its uncertainty, rather than treating every temperature figure as an equivalent measurement of warming. (ipcc.ch)

For public-health research, verification extends from the reported result to the study’s design, participants, comparison group, outcome definition, and follow-up period. Confidence intervals indicate the precision of an estimated association; they do not eliminate systematic bias. The US Centers for Disease Control and Prevention identifies confounding, selection bias, and information bias as important limitations of observational influenza-vaccine studies. For example, chronic illness can be associated both with greater vaccination uptake and with greater hospitalization risk, distorting an unadjusted comparison. Specialist interpretation must therefore distinguish a measured association from a justified causal inference, and keep the conclusion within the population and outcome actually studied. Evidence concerning laboratory-confirmed hospitalization, for instance, does not by itself establish an identical effect on every respiratory illness or in every population. (cdc.gov)

Images, video, and digital provenance

Visual verification separates a file’s origin from the claims attached to it. The Verification Handbook’s image and video workflows, written by Trushar Barot and Malachy Browne respectively, investigate the source, recording location, recording date, and depicted event. Tracing a repost backwards means examining account histories, credits, descriptions, and earlier uploads, then contacting the apparent creator to ask how the material was recorded and obtain the original file. The earliest recoverable publication is a lead, not automatic proof of authorship: its uploader may also have copied it. Publication time must likewise be distinguished from recording time. (verificationhandbook.com)

Reverse-image searching compares an image with indexed copies or visually similar material, potentially revealing an earlier publication, a different caption, or source material used in a composite. For video, investigators extract representative stills or keyframes—especially frames containing distinctive objects—and search them individually. Aric Toler’s 2017 Bellingcat guide describes using screenshots and thumbnails this way, while documenting failures: footage from Russia’s 2012 Kavkaz military exercises circulated as fighting near Svitlodarsk in December 2016, yet reverse searches did not reliably retrieve the original. Keyword searches informed by the scene and soundtrack provided another route. A search returning no earlier match therefore does not establish that footage is new or authentic. (d1kn6o6up31pvd.cloudfront.net)

Embedded metadata, including EXIF information in photographs, can supply camera details, timestamps, and sometimes geographical coordinates. Such information requires corroboration: camera clocks may be incorrect, and social platforms may strip metadata during processing. Location checking compares buildings, road layouts, signs, terrain, and other landmarks with maps, satellite imagery, or street-level photographs. Recording dates can be constrained through weather records, visible clocks, and independently documented events, although weather alone cannot establish a date. (verificationhandbook.com) Corroborating views can connect these clues: Toler’s guide demonstrates matching photographs and video in Linden, New Jersey, through fragments of a business sign, a highway junction marker, and a visible street name. (d1kn6o6up31pvd.cloudfront.net)

Authentic media presented in a false context differs from media whose visual contents have been altered. An older recording falsely described as a current event may require no pixel-level fabrication; a composite inserts material absent from the photographed scene. In his Verification Handbook account of checking purported “street sharks” during Hurricane Sandy in 2012, Tom Phillips describes working with Alexis Madrigal to locate the source photograph of a shark fin inserted into one flooded-street image. That finding did not establish how a second image had been made: its different shark required a separate source match, subsequently identified by Snopes. Evidence of manipulation in one item did not conclusively authenticate or disprove another. (verificationhandbook.com)

Generative artificial intelligence adds fully synthetic images and recordings, alongside techniques that modify genuine footage—for example, changing mouth movements to fit replacement speech. Detection tools produce probabilistic assessments, sometimes expressed as confidence percentages, rather than conclusive findings about a file’s history. Full Fact’s April 2024 report stated that it did not quote such tools in its articles at that time because their results were inconsistent; genuine material could also be incorrectly labelled synthetic. Technical findings therefore need comparison with source tracing, contextual evidence, and specialist examination. Provenance mechanisms, including signatures and watermarks, provide a different kind of evidence by recording or signalling a file’s origin or processing history, rather than inferring artificial generation from its appearance alone. (fullfact.org)

Verdicts, explanations, and corrections

A binary verdict classifies a claim as true or false; a graduated scale distinguishes degrees of accuracy and contextual completeness; an explanatory conclusion states what the evidence establishes without necessarily assigning a standardized label. These formats are editorial choices rather than universal standards. PolitiFact’s Truth-O-Meter has six categories: True, Mostly True, Half True, Mostly False, False, and Pants on Fire. Its intermediate categories distinguish an accurate assertion needing clarification from a partially accurate assertion omitting important details, or one whose limited factual basis conceals critical contrary information. In contrast, Full Fact’s report covering January–June 2019 explained that its general published checks did not use a rating scale, although its Facebook partnership required ratings. A narrative conclusion could therefore coexist with a separate classification required for a particular distribution programme. (politifact.com)

Africa Check’s categories distinguish types of evidential failure, not simply positions along a true–false continuum. “Incorrect” means that the claim conflicts with the best publicly available evidence; “misleading” identifies accurate elements presented in a way that creates a mistaken impression. “Exaggerated” and “understated” distinguish overstatement from understatement. “Unproven” means that available evidence neither establishes nor disproves the assertion: it is not a synonym for incorrect. Its “Checked” label indicates that multiple assertions have been examined, rather than expressing a degree of accuracy. The organization’s procedure assigns the rating through discussion between the lead researcher and two editors, with a vote if agreement cannot be reached. (africacheck.org)

Literal accuracy and contextual adequacy can consequently produce different assessments. As an illustrative example, expenditure rising from 100 to 110 units represents a nominal increase of 10%; if prices rose by 20%, presenting that increase as greater purchasing power would omit a decisive comparison. The numerical statement and the implication require separate examination. Africa Check explicitly considers both wording and context, while PolitiFact’s rating deliberations examine literal truth, alternative interpretations, supporting evidence, and comparable previous rulings. Such categories describe the accuracy of the assertion; the scale itself does not establish whether an error was intentional. (africacheck.org)

The published explanation supplies the reasoning that a label compresses. Africa Check’s documented structure identifies the claim and its original context, presents evidence supporting or contradicting it, and ends with a conclusion explaining the verdict. A preliminary summary allows readers to see the principal finding before the detailed analysis, while links or attributed quotations identify the supporting material. A colleague independently reviews the findings before publication. This structure connects the conclusion to its evidential justification and makes the investigation open to scrutiny: readers can distinguish an erroneous figure, an inappropriate comparison, and an unresolved evidentiary gap rather than encountering only a negative label. (africacheck.org)

Post-publication changes distinguish mistakes in the check from additions to the evidentiary record. Under PolitiFact’s policy, a major factual error that changes the rating or overall assessment receives a correction at the top, with the previous version archived and linked; a factual error that does not affect the assessment receives a correction at the bottom. An update instead adds relevant information, such as a later response that leaves the conclusion unchanged or subsequent news. Africa Check likewise provides for review when something was missed or new or better evidence emerges, with readers directed to the change. Thus, a correction acknowledges a defect in the checker’s account, whereas an update can supplement an assessment without declaring its original reasoning erroneous. (politifact.com)

Organizations, funding, and professional standards

Fact-checking organizations operate within newsrooms, universities, and independent nonprofit institutions; these categories can overlap. PolitiFact began at the Tampa Bay Times in 2007, but direct ownership transferred to the nonprofit Poynter Institute in February 2018. FactCheck.org is a project of the University of Pennsylvania’s Annenberg Public Policy Center, combining institutional support with a public-facing journalistic mission. Africa Check operates as an independent nonprofit trust headquartered in South Africa; control transferred from its London-founded organization to the South African trust at the beginning of 2019. These arrangements place editorial teams within different ownership and governance structures rather than constituting a single organizational model. (politifact.com)

Funding likewise combines several mechanisms. PolitiFact reports revenue from advertising, content partnerships, grants, and reader donations through its Truth Squad membership programme. Its contracts and grant agreements assert editorial independence, and its policy excludes donations from political parties, officeholders, candidates, and anonymous sources. FactCheck.org receives endowment support through its university centre and began accepting public donations in 2010. It also identifies in-kind institutional support, including infrastructure and administrative services, separately from cash receipts. Africa Check describes diversification as a safeguard against donor control: it launched with one funder, the International Press Institute, in 2012, while foundations and trusts supplied all its income in 2014. Its published policy states that funders have no influence over report conclusions. (api.politifact.com)

The International Fact-Checking Network (IFCN), a unit of Poynter, launched its Code of Principles in 2016. The code establishes five commitments: nonpartisanship and fairness; standards and transparency of sources; transparency of funding and organization; standards and transparency of methodology; and an open, honest corrections policy. These commitments translate professional ethics into inspectable requirements. Equivalent claims must receive equivalent evidentiary treatment regardless of their source; reports must identify significant evidence sufficiently for readers to examine the work; and organizations must explain how they select, research, edit, publish, and correct checks. Independent applicants must disclose funding sources accounting for at least 5% of the preceding financial year’s revenue, while units within larger organizations must disclose ownership and editorial control. (poynter.org)

Conflict-of-interest safeguards extend beyond identifying donors. The code requires disclosure of relevant interests held by quoted sources and of commercial relationships that readers could reasonably consider influential. It also requires a public staff nonpartisanship policy and separation of editorial control from political influence. Internal editorial review provides another layer of scrutiny. In Africa Check’s documented English-language workflow, the researcher and editors review a draft and discuss its conclusion; another editor then independently reviews the report before copyediting. The chief editor authorizes publication after considering outstanding responses or developments. Its Senegal operation, which works in French, follows a separate process. (ifcncodeofprinciples.poynter.org)

IFCN signatory status involves external assessment rather than self-declaration alone. Independent assessors examine compliance, and the advisory board reviews their assessments before deciding applications. The IFCN’s report for 2024 recorded 135 completed assessments receiving a final determination; four completed renewal applications were rejected for sourcing or methodological problems. Readers may also submit evidence of code violations through a complaints mechanism. Under the revised policy documented in that report, complaints about individual checks belong with the publisher unless they also allege a code violation; substantiated violations can lead to an improvement process, suspension, or revocation of signatory status. Verification therefore concerns an organization’s compliance with professional standards, not certification that every published verdict is correct. Individual reports remain open to evidentiary challenge and correction. (ifcncodeofprinciples.poynter.org)

Publication, distribution, and audience access

Fact-checks reach audiences through dedicated websites, news partnerships, social media, and messaging services. Republication allows a specialist organization’s investigation to appear in outlets whose readers may not visit its website. Africa Check, for example, permits its reports, guides, and factsheets to be republished in full without charge, subject to attribution and other conditions. Its website provides an HTML code generator that preserves hyperlinks, verdicts, and credits when an article is transferred into another publisher’s content-management system. This distribution mechanism retains access to the underlying evidence rather than circulating only a detached rating. Its republication guidance also distinguishes attribution of the investigation to Africa Check from attribution of statistics to the institution that originally produced them. (africacheck.org)

Audience-oriented formats include short audio summaries and services that combine submissions with distribution. In 2019, Africa Check and the South African podcast company Volume created What’s Crap on WhatsApp?, a short fact-checking voice-note programme distributed through WhatsApp. Africa Check’s country-specific broadcast lists also allow subscribers to submit questionable claims and receive summaries of published checks. Language adaptation extends beyond translating full-length articles: in 2020, its Kenyan #KeepTheFactsGoing series circulated weekly voice notes in English and Kiswahili, combining concise corrections with verification techniques. Related voice notes circulated in isiZulu in South Africa, Wolof in Senegal and elsewhere in West Africa, and Hausa in Nigeria. These initiatives placed material in messaging formats and languages already used by their intended audiences. (africacheck.org)

Machine-readable publication uses ClaimReview, a Schema.org vocabulary type for describing a published review of a claim. Its claimReviewed property records a short summary of the assertion examined; author identifies the person or organization responsible for the review; datePublished records its first publication date; and reviewRating describes the assessment. The itemReviewed property identifies the reviewed item, while reviewBody can contain the explanation. Schema.org’s examples encode these fields using JSON-LD. Such markup gives software identifiable fields instead of requiring it to infer the claim and verdict from prose. It describes a publisher’s assessment; it does not itself gather evidence, verify the assertion, or certify the assessment’s correctness. (schema.org)

Online availability does not remove geographical gaps in reporting capacity. Duke Reporters’ Lab’s March 2023 study, led by Mark Stencel and Erica Ryan, identified 50 active locally focused projects in 25 US states and the District of Columbia. From January 1, 2022, through Election Day, these projects produced 976 checks examining more than 1,300 claims. Nevertheless, only 47 of 7,386 state legislators had a statement checked, and 25 states lacked consistent state or local political fact-checking. Discoverability presented a separate obstacle: 13 of the 50 projects had no dedicated webpage or tag collecting their checks. These findings distinguish the existence of national websites from sustained scrutiny of local claims and convenient access to the resulting reporting. (reporterslab.org)

Regional networks address coverage through collaboration among organizations working in different countries and languages. Established in 2017, the Africa Facts Network connects African fact-checking organizations, shares selected checks through Africa Check’s website, and supports exchange through multilingual newsletters, meetings, and summits. Its documented concerns include linguistic diversity, resource constraints, and vulnerable or offline communities. Together with the multilingual voice-note initiatives, this illustrates two distinct dimensions of audience access: having reporting organizations able to investigate locally relevant assertions, and distributing their findings in languages and formats accessible to the people encountering those assertions. (africacheck.org)

Automation and artificial intelligence

Automated fact-checking combines natural language processing and machine learning in a sequence of tasks. Claim detection identifies factual assertions, while prioritization ranks candidates for investigation. Matching searches for previously checked assertions, including paraphrases. For new claims, information retrieval locates potential evidence; verification models assess whether that evidence supports or contradicts the assertion, or leaves it unresolved. Explanation production selects supporting passages or generates an account of the judgment. The 2022 survey by Zhijiang Guo, Michael Schlichtkrull, and Andreas Vlachos distinguishes these components: identifying a checkable sentence is not equivalent to establishing its accuracy, and agreement with retrieved text does not establish that the source itself is trustworthy. (aclanthology.org)

Full Fact provides a practical example of human-assisted automation. The organization began using machine learning in 2016. Its monitoring software collects material from online news, television speech, podcasts, and social media, converts it into text, and processes individual sentences. A claim-type classifier, developed by fine-tuning BERT on annotated examples, distinguishes categories such as quantitative statements, causal assertions, and predictions; topic filters further narrow the material presented to journalists. Matching tools identify repetitions of previously checked claims despite changes in wording, using sentence representations and comparisons of entities and numbers. Full Fact also reports introducing a generative model to reduce the examples needed to adapt matching to other languages. These functions assist selection and research rather than automatically determine published verdicts. (fullfact.org)

Benchmark design strongly affects reported performance. The 2018 FEVER dataset, introduced by James Thorne and colleagues, contains 185,445 claims constructed by modifying sentences extracted from Wikipedia. Annotators assigned “Supported,” “Refuted,” or “NotEnoughInfo” labels and recorded evidence sentences for supported and refuted claims. The researchers’ baseline achieved 50.91% label accuracy when evidence correctness was disregarded, but only 31.87% when the correct label and evidence were required together. This difference illustrates why verdict accuracy alone is an incomplete evaluation: a system may produce the expected label without finding its evidential basis. (aclanthology.org)

Operational evaluation must also address ambiguous wording, incomplete retrieval, and time-sensitive facts. Similar sentences may concern different people or reference periods; failure to retrieve support does not demonstrate falsehood. Evidence timestamps matter when facts change. The 2022 survey also identified English-language concentration as a limitation: claims and relevant evidence may appear in different languages, requiring cross-language retrieval and separately evaluated language coverage. A fixed test set with supplied evidence therefore measures a narrower task than monitoring public debate and locating reliable sources. Generated explanations introduce another problem: plausible prose may not faithfully reflect either the evidence or the model’s decision process. (aclanthology.org)

AI-assisted output consequently remains subject to source verification and editorial review. Full Fact’s editorial policy explicitly identifies hallucinations, inaccuracies, and bias as risks. It requires research conducted with AI assistance to be fully verified before use. A fact-check drafted with AI assistance must be checked by its human writer and then reviewed by at least two senior team members; other AI-assisted content requires at least one human reviewer before publication. This separates assistance with finding, organizing, or drafting material from responsibility for the published assessment: an AI-produced explanation is material to investigate, not an independent evidential authority. (fullfact.org)

Effectiveness and methodological limitations

Experimental evaluation uses random assignment to compare people exposed to misinformation followed by a correction with people exposed to misinformation alone or to neither. Differences in subsequent accuracy ratings support causal inference about the correction’s effect; follow-up surveys test persistence without showing the correction again. Ethan Porter and Thomas J. Wood’s 2021 study applied this design in Argentina, Nigeria, South Africa, and the United Kingdom. Conducted in September–October 2020, its 28 experiments evaluated 22 distinct fact-checks. Corrections improved accuracy by an average of 0.59 points on a five-point scale. Most effects remained detectable approximately two weeks later in the three countries with follow-up data; Nigeria had no follow-up wave. (doi.org)

Political ideology did not produce a uniform pattern of resistance. In the four-country experiments, correction effects varied across ideological groups, and some corrections did not significantly improve particular groups’ accuracy. Nevertheless, no tested correction produced a backfire effect—an increase in false belief attributable to corrective information. This finding concerns the corrections, participants, and outcomes tested, rather than establishing that backfire is impossible. (doi.org)

Persistence was weaker in John M. Carey and colleagues’ study, published in Nature Human Behaviour in February 2022, of COVID-19 fact-checks in the United States, Great Britain, and Canada. Its preregistered survey experiments found immediate reductions in targeted misperceptions, especially among groups most susceptible to those claims. However, improvements did not persist in panel follow-ups, even after repeated exposure. Effects also transferred only minimally to related factual beliefs: correcting one assertion did not reliably improve understanding of adjacent issues. The contrast with Porter and Wood’s findings makes durability an outcome requiring separate measurement, not an automatic consequence of immediate correction. Differences between the studies do not, by themselves, identify which feature of their topics, materials, participants, or timing explains the divergent results. (nature.com)

Belief accuracy, voluntary exposure, sharing behaviour, and political attitudes are distinct outcomes. An accuracy-rating task does not establish that participants would choose to read a correction, refrain from sharing the original claim on social media, or change their political preferences. These are limits of what the experimental measures establish, not evidence that such effects cannot occur. Assigned exposure also leaves open how selective attention and repeated encounters operate outside a survey. (doi.org)

Selection creates a separate limitation when published verdicts are aggregated. PolitiFact’s methodology prioritizes significant, newsworthy, potentially misleading, and widely repeated assertions rather than sampling everything a speaker says. It also gives greater attention to governing parties and people repeatedly making attention-getting or misleading statements. Consequently, the proportion of a politician’s checked claims receiving false ratings is not an estimate of the proportion of all that politician’s statements that are false. Comparisons between speakers require attention to which claims entered the checked set and why; verdict totals alone cannot remove this selection problem. (politifact.com)

References

1. Making a Difference? A Critical Assessment of Fact-checking in 2012static.newamerica.org
2. The Story Behind the First-Ever Fact-Checkersbunkhistory.org
3. About Usfactcheck.org
4. David Mikkelson, Author at Snopes.com Page 269snopes.com
5. Our 15th Anniversaryfactcheck.org
6. Bill Adairstaging.politifact.com
7. The Truth-O-Meter at 500politifact.com
8. Introducing Poynter’s International Fact-checking Networkpoynter.org
9. PolitiFact | The Principles of the Truth-O-Meter: PolitiFact’s methodology for independent fact-checkingpolitifact.com
10. How we use AI to help fact check party manifestos – Full Factfullfact.org
11. Our Process - FactCheck.orgfactcheck.org
12. The commitments of the Code of Principlesifcncodeofprinciples.poynter.org
13. Consumer Price Index Frequently Asked Questionsbls.gov
14. Analyzing and Interpreting Datacdc.gov
15. Nonsampling Errorcensus.gov
16. Sample Size Definitionscensus.gov
17. Do I need to consider the margins of error (MOE) when comparing American Community Survey (ACS) estimates?census.gov
18. Post-Enumeration Surveyscensus.gov
19. Seasonal Adjustment in the CPIbls.gov
20. Summary for Policymakersipcc.ch
21. Biases to Consider in Vaccine Effectiveness Studiescdc.gov
22. Verification Handbookverificationhandbook.com
23. Advanced Guide on Verifying Video Contentd1kn6o6up31pvd.cloudfront.net
24. Full Fact Report 2024: Trust and truth in the age of AIfullfact.org
25. The Principles of the Truth-O-Meter: PolitiFact’s methodology for independent fact-checkingpolitifact.com
26. Full Fact: Report on the Facebook Third Party Fact Checking programme, Jan–Jun 2019fullfact.org
27. PolitiFact gains nonprofit status with move to the Poynter Institutepolitifact.com
28. Our Mission - FactCheck.orgfactcheck.org
29. The Principles of the Truth-O-Meter: PolitiFact’s methodology for independent fact-checkingapi.politifact.com
30. Our Funding - FactCheck.orgfactcheck.org
31. The commitments of the Code of Principlesifcncodeofprinciples.poynter.org
32. IFCN Code of Principles Report 2024poynter.org
33. ClaimReview - Schema.org Typeschema.org
34. Vast gaps in fact-checking across the U.S. allow politicians to elude scrutinyreporterslab.org
35. A Survey on Automated Fact-Checkingaclanthology.org
36. Full Fact AIfullfact.org
37. FEVER: a Large-scale Dataset for Fact Extraction and VERificationaclanthology.org
38. How we use AI in our editorial workfullfact.org
39. The ephemeral effects of fact-checks on COVID-19 misperceptions in the United States, Great Britain and Canadanature.com
40. The Principles of the Truth-O-Meter: PolitiFact’s methodology for independent fact-checkingpolitifact.com