In March 2020, a viral claim spread on Facebook: "The US has tested more people for COVID-19 than any other country." PolitiFact assigned it False. But the verdict only landed because the team paired the ruling with a precise, citable sentence: "South Korea had tested more than 270,000 people by that point; the US had tested fewer than 75,000." The number was the verdict. The label alone would have been ignored.
Fact-checkers at organisations like PolitiFact, AFP Fact Check, and Full Fact all follow a common discipline: the verdict statement must stand alone. A reader who never reads the full article should be able to understand the ruling from a single sentence. This discipline protects against the most common failure in public fact-checking — verdicts that are technically correct but functionally useless because they are vague.
In 2019, Reuters Fact Check established an internal rule that every published verdict must contain: (1) the original claim restated accurately, (2) what evidence was found, and (3) the ruling with a specific justification. That three-part structure is now standard across most major fact-checking organisations affiliated with the International Fact-Checking Network (IFCN).
Examine how the best-performing fact-check articles (by reader trust metrics) structure their verdicts. Three elements appear consistently:
In 2022, the Duke Reporters' Lab surveyed 92 fact-checking organisations globally. Forty-one percent reported using AI tools — primarily to draft initial verdict summaries and flag logical inconsistencies in researcher write-ups. The AI did not determine the verdict; human editors retained that authority. The AI's role was editorial scaffolding: catching passive constructions that obscured responsibility ("mistakes were made") and suggesting more precise claim restatements.
Tools like the ClaimBuster system, developed at the University of Texas Arlington, automatically detect "check-worthy" claims in text and can draft preliminary verdict summaries that researchers then refine. As of 2023, ClaimBuster had analysed more than 20 million sentences from public discourse.
A verdict is not an opinion. It is a structured finding backed by citable evidence. The moment your verdict uses words like "seems," "appears," or "might be," you have written an inference, not a fact-check result. Readers detect this, and trust falls.
You have investigated a claim. Now you need to write a verdict statement that follows the Reuters three-part structure: restate the claim precisely, anchor it to specific evidence, and state the ruling with justification. Submit a draft verdict below and the AI will critique it using IFCN editorial standards.
When Bellingcat published its initial findings linking a Russian Buk missile launcher to the MH17 crash in July 2014, it chose a medium-length blog post with embedded high-resolution images. The format was deliberate. Investigative lead Eliot Higgins later explained in interviews that Twitter would have lost the evidentiary chain, and a PDF report would have reached only specialists. The blog post format — image-rich, linkable, accessible — reached both mass audiences and intelligence analysts simultaneously.
By 2016, that single post had been cited in the Dutch Safety Board report, shared more than 140,000 times, and translated into six languages. Format was not incidental to impact. It was the mechanism of impact.
A 2023 study by First Draft News analysed 4,200 fact-check articles published across eight major organisations. It found that platform format — not just content quality — was the second strongest predictor of reach, after headline clarity. The organisations that consistently reached the largest audiences adapted their findings to multiple formats simultaneously.
In 2022, Full Fact (UK) published an internal review of its distribution strategy. It found that fact-checks published only in long-form article format reached an average of 12,000 readers. The same fact-checks adapted into a Twitter thread plus an Instagram carousel reached an average of 340,000 people. The evidence did not change. The format tripled the potential audience eleven times over.
The practical implication for fact-checkers: write your long-form article first (it is your evidence record), then strip it into platform-specific formats. The article is the source of truth. The thread, carousel, and video are distribution vehicles. Never reverse this order — distribution-first formats tend to lose evidentiary precision.
When a viral photo claimed to show Brazilian election fraud in October 2022, AFP Fact Check published a long-form article, then immediately adapted it into a five-tweet thread (in Portuguese), an Instagram carousel with reverse-image-search screenshots, and a WhatsApp broadcast to 80,000 subscribers in Brazil. The fact-check reached 2.1 million people within 48 hours — a reach AFP attributed directly to the multi-platform adaptation strategy.
Tools like Logically AI and Google's Fact Check Explorer now offer automated format adaptation: paste a long-form fact-check article and receive a Twitter-optimised summary, an Instagram caption, and a WhatsApp message draft. These tools were used by 23 IFCN-accredited organisations as of the 2023 Duke Reporters' Lab survey. Editors report the AI drafts require significant revision but reduce adaptation time by approximately 60%.
You have just completed a fact-check. Now you need to plan how to distribute it across platforms. Using what you learned from the AFP Brazil election case and Full Fact's strategy, design a multi-platform distribution plan. Tell the AI which platforms you would use, in what order, and why — for the scenario below.
In 2016, a content company called Proper Media sued Snopes co-founder David Mikkelson, alleging in part that Snopes had unfairly profited from licensed content. Simultaneously, a coordinated social media campaign began labelling Snopes as "left-wing" and "unreliable." The campaign cited the lawsuit as evidence of institutional corruption — even though the lawsuit was a business dispute between co-owners, entirely unrelated to editorial integrity.
Snopes documented the conflation publicly, publishing a detailed explanation separating the business dispute from editorial processes. Independent media analysts including those at PolitiFact and the Poynter Institute confirmed no editorial compromise had occurred. The episode became a textbook case of source-attack misinformation — attempting to discredit a fact-checking organisation by associating it with an unrelated controversy.
Research by the Shorenstein Center at Harvard (2020) identified four distinct categories of responses that fact-checkers receive after publishing verdicts, ranked by frequency:
PolitiFact publishes its correction policy publicly. If a subject provides new evidence after a verdict is published, editors review the evidence within 48 hours and either update the verdict (with a documented edit history) or publish a statement explaining why the new evidence does not change the finding. This transparency is central to IFCN accreditation.
In 2020, when the Trump campaign formally disputed a PolitiFact ruling on COVID test numbers, PolitiFact published the campaign's full written response alongside the original fact-check, with a paragraph-by-paragraph editorial response. No verdict was changed because no new evidence was provided — only rhetorical challenge. The exchange was itself cited in journalism schools as an example of transparent pushback management.
A 2022 report by the Coalition for Women in Journalism found that female fact-checkers receive harassment at 3.4 times the rate of male colleagues after publishing politically sensitive verdicts. The International Center for Journalists (ICFJ) launched the OnTheLine helpline in 2021 specifically to support journalists facing online harassment, and by 2023 had assisted more than 800 cases, many involving fact-checkers.
Several fact-checking organisations now use AI-assisted monitoring to detect when their published fact-checks are being misrepresented online. The tool Meltwater, used by AFP and Reuters, monitors social media for mentions of a published fact-check and flags instances where the verdict has been inverted or stripped of context — for example, a headline screenshot shared without the verdict, or a partial quote from the article used to suggest the opposite conclusion.
In 2021, AFP Fact Check used Meltwater to identify that a Brazilian political network was sharing screenshots of an AFP fact-check with the verdict cropped out, making it appear AFP had confirmed rather than debunked the false claim. AFP published a follow-up article documenting the cropping tactic within 24 hours.
You have published a fact-check. Now you are receiving pushback. The AI will play the role of a subject disputing your verdict. Identify the type of pushback (legitimate correction, good-faith disagreement, source-attack, or coordinated harassment) and draft an appropriate response following IFCN protocol.
In the weeks before France's 2017 presidential election, First Draft News coordinated a coalition of 37 newsrooms — including Le Monde, AFP, BuzzFeed News France, and regional outlets — under the CrossCheck project. Every participating organisation agreed to use a shared taxonomy of claim types and verdict labels, so that a claim debunked by one newsroom would be instantly findable by readers of any other participating outlet.
The coalition processed 300 individual fact-check submissions in seven weeks. Post-election analysis by the Reuters Institute found that claims debunked within the CrossCheck network spread at significantly lower rates on French social media than similar debunked claims in the 2016 US election, where no coordinated network existed. Coordination multiplied the impact of each individual fact-check.
A 2022 study by the Empirical Studies of Conflict project at Princeton analysed 18,000 fact-check articles published between 2015 and 2021. It found that 73% of fact-check articles received fewer than 500 organic page views after their first week of publication, regardless of their accuracy or importance. The problem was discoverability: most fact-checks were not indexed in ways that allowed readers encountering a false claim months later to find the existing debunking.
Google's Fact Check Tools were introduced in 2017 specifically to address this. When news organisations use the ClaimReview structured data markup — a free schema that embeds verdict metadata into article HTML — their fact-checks become indexable by Google's Fact Check Explorer. As of 2023, more than 11,000 fact-checking pages globally use ClaimReview markup.
ClaimReview is a technical schema developed by schema.org that allows fact-checkers to embed structured metadata in their articles. The metadata includes: the claim text, the claimant, the claim date, the verdict label, and the rating value. When embedded correctly, Google Search displays the verdict directly in search results — a fact-check label appears next to the original false claim in Google Search, giving readers the debunking at the moment they encounter the claim.
The Duke Reporters' Lab maintains a real-time database of ClaimReview-marked fact-checks globally. Their 2023 annual census found 68 countries with at least one IFCN-accredited fact-checking organisation using ClaimReview, and estimated that ClaimReview-enabled fact-checks received on average 340% more organic search traffic than equivalent fact-checks without the markup.
The IFCN's CoronaVirusFacts Alliance, formed in March 2020, united 100 fact-checking organisations across 45 countries to collectively debunk COVID-19 misinformation. Every partner used ClaimReview markup and shared verdicts via a central database. By December 2020, the alliance had published more than 9,000 individual fact-checks, all cross-searchable. The WHO cited the alliance database in its own communications as a reference source — making individual journalists' fact-checks part of a global public health evidence record.
One of the most significant practical benefits of AI in fact-checking distribution is claim matching: automatically detecting when a claim circulating online has already been fact-checked and surfacing the existing verdict. Google's Fact Check Explorer API, ClaimBuster, and the Full Fact automated fact-checking system all perform some version of this function.
In 2021, Full Fact demonstrated that their automated system could match a new instance of a circulating claim to an existing fact-check in under 3 seconds, compared to an average of 47 minutes for a human researcher to perform the same task. The system was used to respond to claims in UK parliamentary debates in near-real time during the 2021 pandemic response.
Even without institutional infrastructure, individual fact-checkers can build durable records. The steps followed by student journalism programmes at Northwestern, Columbia, and City University London include:
You have written a complete fact-check article. Now you need to make it durable and discoverable for the long term. Using the six-step framework from student journalism programmes at Northwestern, Columbia, and City University London, build a preservation and discoverability plan for the scenario below. Discuss your choices with the AI.