Fact-checking & misinformation detection

False and misleading claims spread faster than human fact-checkers can verify them, and generative AI has made both the volume of content and the errors in AI-written text bigger. AI-assisted fact-checking does not hand down "truth". It finds the claims worth checking, spots claims that have already been debunked, retrieves evidence from trusted sources and drafts a sourced verdict that a human editor approves. Newsrooms and NGOs use it to monitor public debate. Companies, agencies and public bodies use it to check their own content before publishing, which matters more and more as that content is drafted by LLMs.

Problem & value

• Who: fact-checking organisations and newsrooms, public broadcasters, government communication teams, communication/PR and marketing teams, researchers (EDMO hubs), platforms (trust & safety), investor-relations and ESG teams, and private people who want to check a viral post.
• Today: a fact-checker reads and watches by hand, keeps spreadsheets of claims and searches manually. A single check can take hours. The real bottleneck is finding and prioritising claims.
• AI value: triage at scale and re-use of existing work. Full Fact's toolset processes about a third of a million sentences on a typical weekday. More than 40 fact-checking organisations in 30 countries use it, in three languages [1]. Automatic evidence retrieval and verdict drafting cut research time per claim. The editor keeps the final judgement.
• Market shift: in January 2025 Meta ended third-party fact-checking in the US and replaced it with Community Notes [3]. In the EU, the Code of Conduct on Disinformation has been part of the DSA framework since 1 July 2025, and its commitments can be audited for very large platforms [2]. Demand is moving from platform-funded fact-checking towards in-house verification and EU-funded infrastructure (EDMO, vera.ai) [4][5].

Variants

• Newsroom / fact-checker monitoring: live monitoring of TV, radio, podcasts, parliament and social media, with claim detection, matching to the archive and an editorial workflow.
• Pre-publication self-check (SMB/enterprise): check marketing copy, reports, press releases and LLM-generated drafts (LANG-005, LANG-028) against sources and internal data before release. Regulated claims such as health, financial and green claims belong in LANG-024.
• Grounding / hallucination check in AI pipelines (developer): sentence-level checks that each claim in a RAG or agent answer is supported by the retrieved documents (KNOW-002).
• Public sector: government press offices and election authorities track false narratives about procedures (e.g. voting dates and rules) and publish corrections. This is document-level monitoring, not profiling of citizens.
• Research & OSINT: EDMO hubs and universities analyse narratives across languages (with LANG-016 and LANG-018).
• Capital markets / ESG: checking rumours and adverse-media claims about issuers. Controversy screening as such is part of LANG-018.
• Private use: browser plugins and chat assistants with web search that show sources for a viral claim. Useful, but users must open the sources themselves.
• Out of scope: AI-content provenance and watermark detection → ANOM-015; voice/video deepfakes for fraud → SEC-011; image/video moderation → VIS-027; toxicity moderation → LANG-014.

How it works

1. ingest (feeds, crawler, ASR of broadcast, OCR/VLM of screenshots, own drafts)
2. sentence split + language ID (+ translation for cross-lingual matching)
3. check-worthiness detection (encoder classifier or small LLM; filters opinions, predictions)
4. claim normalisation (rewrite a noisy post into a self-contained claim)
5. claim matching: embeddings + reranker against fact-check archive (own + ClaimReview / Fact Check Explorer) ├─ known claim
6. link existing fact-check
7. editor └─ new claim
8. evidence retrieval (web search API, trusted sources, statistics, internal KB)
9. verification (LLM reasoning over evidence; NLI/grounding model per sentence)
10. verdict proposal + citations + uncertainty
11. HUMAN EDITOR
12. publish (article + ClaimReview markup) / internal correction / narrative alert

• Check-worthiness and matching are the most mature parts. Fine-tuned BERT-class models and multilingual embeddings work well. Full Fact uses a BERT-based claim classifier plus a hybrid claim-matching model, now improved with generative AI [1].
• Verification works like an agent: plan queries, search, read, reason, cite. Google's SAFE showed that an LLM that splits text into atomic facts and checks each against Google Search agreed with crowd annotators on 72% of ~16k facts. Where the two disagreed, SAFE was right in 76% of a sampled 100 cases, at more than 20× lower cost [6].
• Grounding check: small NLI-style models such as MiniCheck/Bespoke-MiniCheck and Vectara HHEM score whether a source passage supports a sentence. They are cheap enough to run on every sentence [7][8].
• Human-in-the-loop: a person always approves verdicts, especially "false" labels about named people or organisations. The AI recommends; it never publishes on its own. Every claim needs dated evidence that predates the claim, so later knowledge is not used to judge it.

Data preparation

The evidence base matters more than the model. Typical sources are fact-check archives (ClaimReview feeds), statistics offices and registers, plus your own knowledge base: product data sheets, press releases, past corrections and policy papers. In an SMB these live on file shares, in the CMS, in mail and in PDFs; scattered knowledge, outdated versions and scanned documents are the usual gaps. Monitored streams (news, social posts, transcripts) add duplicates, missing timestamps and unclear provenance.

Preparation steps: - Define a short list of trusted sources and rank them by reliability and date. - Curate an authoritative knowledge base and drop superseded versions (PREP-013, PREP-017). - Run OCR on scanned reports and convert old formats (PREP-006, PREP-007); see also DOC-014. - Add metadata such as source, date and topic (PREP-010). - Build a test set of 50 to 100 claims with known verdicts together with editors (PREP-018).

Tools: Paperless-ngx or a DMS for scans, a vector database for retrieval, simple scripts for scraping and deduplication. Expect roughly 30 to 40 percent of the effort in source curation and the test set; data volume is small, quality is decisive.

Implementation paths

Which path for whom

• SMB: B, because the check must run against your own published content and a curated set of trusted sources, and no standard product covers that.
• Enterprise: B, because communications and compliance teams need integration with their CMS, monitoring feeds and approval workflow.
• Developer: B, because retrieval, reranking and citation APIs are quick to assemble and easy to evaluate.
• Private: A, because ready-made search and fact-check tools are enough for occasional checks and need no setup.
• Public: C, because public bodies and broadcasters benefit from sovereign EU or German hosting and open models for sensitive monitoring.

Path A — Buy (off-the-shelf)

• Fact-checker tooling: Full Fact AI (UK charity, licensed to other organisations) [1], Factiverse (Norway, EEA), Truly Media (collaborative verification platform from Athens Technology Center and Deutsche Welle, EU), and the free InVID-WeVerify verification plugin maintained by AFP Medialab (more than 57,000 weekly active users at the end of WeVerify) [9].
• Source-credibility data: NewsGuard (US) ratings and false-narrative feeds. Media-monitoring suites (LANG-018) increasingly add narrative and disinformation modules.
• Assistants with web grounding: Perplexity, ChatGPT search, Gemini, Le Chat (Mistral, EU). Fine for individual checks, but there is no editorial workflow or audit trail.
• Choose A when: you are a newsroom or NGO and want monitoring plus workflow, or an SMB that only occasionally needs to check a claim.

Path B — Build on cloud APIs (recommended for most organisations)

• LLMs with built-in search: Gemini API Grounding with Google Search, which returns URL citations mapped to text spans [10]. Mistral Agents API web_search / web_search_premium, an EU provider that adds verified news sources [11]. Claude and OpenAI models with their web-search tools. Azure OpenAI or Bedrock in EU regions for the reasoning step, with a separate EU search provider.
• Archive matching: Google Fact Check Tools API (Claim Search) to look up existing fact-checks, plus ClaimReview write access for publishers [12]. Add your own vector index of past checks.
• Grounding scoring: a hosted MiniCheck/HHEM endpoint or an LLM-as-judge with structured output (verdict, evidence[], confidence).
• Orchestration: LangGraph, the OpenAI Agents SDK or n8n. Store all evidence snapshots (URL, retrieval time, hash) for audit.
• Choose B when: you want a pre-publication checker or a narrow monitoring use case in weeks, without your own GPUs.

Path C — Self-host (open source)

• Models: open-weight reasoning LLMs (Qwen3 family, Mistral Small/Medium open releases, gpt-oss) served via vLLM. Bespoke-MiniCheck-7B (commercial licence through Bespoke Labs) or the MiniCheck RoBERTa/DeBERTa/Flan-T5 variants [7]. HHEM-2.1-Open (Apache-2.0, English only, runs on CPU in about 1.5 s for a 2k-token input) [8]. Multilingual embeddings (BGE-M3, multilingual-E5) plus a reranker.
• Pipelines: SAFE reference code (google-deepmind/long-form-factuality) [6], OpenFactCheck/FActScore-style atomic-fact checkers, Qdrant/OpenSearch for the claim archive, SearXNG or a commercial search API for web evidence, Whisper-class ASR for broadcast monitoring.
• Hardware: HW2 (one 24–32 GB GPU) for a ~8–30B quantised LLM plus MiniCheck and embeddings, enough for pre-publication checks and moderate monitoring. HW3 for 70–120B reasoning models and multilingual high-volume monitoring. On a single A6000, MiniCheck-7B processes a 29k-claim benchmark in about 30 minutes with prefix caching [7].
• Note: web evidence still needs an external search index. Fully air-gapped setups can only verify against internal and curated sources.

Path D — Custom training

Worth it for check-worthiness and claim-matching classifiers in your own languages, topics and editorial policy. A few thousand labelled sentences from the editors' triage decisions are usually enough. A domain grounding model is also worth training (fine-tune a MiniCheck-style model on synthetic and real support/refute pairs). Training an end-to-end "truth model" is not recommended: verdicts must rest on retrieved, citable evidence, not on model weights.

State of the art (2026)

• Agentic, retrieval-first verification (decompose → search → reason → cite) has replaced classifier-only "fake news detection". Stylistic classifiers do not generalise and are easy to game. Frontier LLMs with native search tools return citations mapped to text spans [10][11].
• Benchmarks: CLEF CheckThat! 2025 covered subjectivity, multilingual claim normalisation, numerical claims (English/Spanish) and scientific web discourse [13]. The FEVER workshop's AVeriTeC line has moved to image-text claims (AVerImaTeC). The winning system reached only ~0.55 verdict accuracy with adequate evidence, so multimodal real-world verification is far from solved [14]. LLM-AggreFact tracks grounding checkers, and small models such as MiniCheck reach GPT-4-level grounding accuracy at a fraction of the cost [7].
• Community notes + AI: platforms are combining crowd notes with AI drafting [3]. Quality depends on how humans rate the notes.
• EU infrastructure: EDMO's 15 hubs cover all 27 member states plus Moldova, Norway and Ukraine [4]. vera.ai (EU grant 101070093) built AI verification tools with AFP, DW and others [5].
• Limits: numeric and statistical claims, very recent events with little evidence, low-resource languages, sarcasm, out-of-context images, and evidence pollution from AI-generated web pages. LLMs can cite sources that do not actually support the sentence, so run a separate grounding check on every citation.

Shared resources & combinations

• Same retrieval stack (embedding-text + vector-db + reranker + search-engine): KNOW-002 RAG, KNOW-015 deep-research agent, KNOW-008 literature review. One vector store and search gateway can serve all of them, and the grounding checker doubles as a hallucination guard for KNOW-002.
• Same monitoring pipeline (web-scraping + asr + text-classifier): LANG-018 media monitoring, LANG-016 trend detection, LANG-014 moderation, LANG-015 entity extraction (who said what).
• Content production chain: LANG-005 / LANG-028 / LANG-011 generate and edit text → LANG-021 checks facts → LANG-024 checks regulated claims → ANOM-015 marks AI content.
• Multimodal: VLM for screenshots and image-text claims (shared with VIS models); SEC-011 for manipulated voice and video; LANG-009 for cross-lingual claim matching; SPEECH-004 live captioning and SPEECH-021 parliamentary transcription share the ASR for broadcasts and debates.
• Hardware: one HW2/HW3 GPU server running vLLM can host the reasoning LLM, the MiniCheck grounding model and the embeddings for the whole LANG.ANALYZE + KNOW family.

Effort, cost & KPIs

• Path A: days. Licences per seat or per organisation; free plugins for individuals.
• Path B: 4–8 weeks for a pre-publication checker or a monitoring prototype. Costs are driven by search-tool calls and reasoning tokens per claim, not by the base model.
• Path C: 2–4 months including the search gateway, evidence archive and editor UI, plus ongoing source curation.
• KPIs: claims surfaced per editor-hour, precision of check-worthiness, archive match rate, time from claim appearance to published check, share of AI verdicts accepted without change, citation support rate (grounding-checker pass rate), corrections issued after publication.

People & change

Start with management and the department lead (communications, editorial or compliance) to set the scope: the tool proposes, an editor decides. Involve IT early for integration and accounts, and the DPO as soon as social posts or named persons are monitored, because that can be personal data and may need a DPIA (Art. 35 GDPR) and an AVV (Art. 28 GDPR) with the API provider. Legal-compliance clarifies liability for published corrections.

Works council: usually not required, since the system analyses external content, not employee behaviour. It becomes relevant if the tool logs editors' individual decisions or error rates (§87(1) no. 6 BetrVG), and §90(1) no. 3 BetrVG still calls for timely information when work processes change.

AI literacy: support editors with short training on hallucinations, source checking and automation bias (AI Act Art. 4; GUIDE-006). Pick a small pilot group of two to four editors, show the evidence behind every verdict, and measure time saved. Typical concern: "the AI decides what is true". Answer it by design: no verdict is published without human approval, and sources are always visible. Further reading: GUIDE-008, GUIDE-009.

Risks, regulation & compliance

• EU AI Act: typically minimal risk. Fact-checking assistance is not an Annex III category and Art. 4 AI-literacy duties apply. Watch these edges:
• Art. 50(4) requires disclosure of AI-generated text published to inform the public on matters of public interest. It does not apply if a human reviews it and someone holds editorial responsibility, which is a strong reason to keep the editor in the loop [15]. Art. 50 applies from 2 August 2026; check the final consolidated text after the Digital Omnibus amendments.
• Systems designed to influence elections or voting behaviour fall under Annex III 8(b). A neutral fact-check tool does not, but campaign-style "counter-messaging" can.
• DSA: very large platforms must assess and mitigate systemic disinformation risks (Art. 34/35) and follow the Code of Conduct on Disinformation, which has been auditable since 1 July 2025 [2]. Automated labels need appeal routes.
• GDPR: claims name people, and monitored posts have authors. Rely on the journalistic exemption (Art. 85 GDPR and national press law) or legitimate interest. Do not build profiles of individual users. Keep data minimal and set retention rules.
• Freedom of expression & defamation: a wrong "false" label can be defamatory and can chill speech. Always show evidence, uncertainty and a correction channel, and never auto-label without human review, above all in public-sector use.
• Technical risks: hallucinated or mismatched citations, bias in source-credibility lists, prompt injection from crawled pages (guardrails), and adversaries testing the system. Mitigations: grounding check on every citation, allow-lists of sources, archived evidence snapshots, red-teaming.

Getting started

1. Pick one job: pre-publication check of our own content or monitoring of topic X. Do not try to "detect misinformation" in general.
2. Build a gold set: 100–300 real claims from your domain with verdicts and evidence, made by your editors.
3. Prototype on Path B: claim extraction → Fact Check Tools API lookup → LLM with web grounding → MiniCheck or HHEM on each cited sentence → editor review screen.
4. Measure precision, citation support rate and editor time saved against the gold set. Agree on an editorial policy for wording verdicts and handling corrections.
5. Add ClaimReview markup to your published checks, and consider joining EDMO or national fact-checking networks for archive sharing.

Sources

1. Full Fact — Full Fact AI tools — https://fullfact.org/ai/
2. European Commission — The Code of Conduct on Disinformation (integration into the DSA, effective 1 July 2025) — https://digital-strategy.ec.europa.eu/en/library/code-conduct-disinformation
3. Meta — More Speech and Fewer Mistakes (7 Jan 2025) — https://about.fb.com/news/2025/01/meta-more-speech-fewer-mistakes/
4. EDMO — EDMO Hubs — https://edmo.eu/edmo-hubs/
5. vera.ai — VERification Assisted by Artificial Intelligence — https://www.veraai.eu/
6. Wei et al. — Long-form factuality in large language models (SAFE) — https://arxiv.org/abs/2403.18802
7. Tang, Laban, Durrett — MiniCheck (EMNLP 2024) & Bespoke-MiniCheck-7B — https://github.com/Liyan06/MiniCheck
8. Vectara — HHEM-2.1-Open hallucination evaluation model — https://huggingface.co/vectara/hallucination_evaluation_model
9. WeVerify — InVID-WeVerify verification plugin — https://weverify.eu/verification-plugin/
10. Google — Gemini API: Grounding with Google Search — https://ai.google.dev/gemini-api/docs/google-search
11. Mistral AI — Agents API web search connector — https://docs.mistral.ai/agents/connectors/websearch/
12. Google — Fact Check Tools API — https://developers.google.com/fact-check/tools/api
13. CLEF CheckThat! Lab 2025 — https://checkthat.gitlab.io/clef2025/
14. FEVER workshop — AVerImaTeC shared task — https://fever.ai/task.html
15. EU AI Act, Article 50 (transparency obligations) — https://artificialintelligenceact.eu/article/50/
16. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj