News Article Fact-Check

AI claim extraction and source credibility analysis followed by crowd fact-checker review to produce a veracity verdict on news content.

The News Article Fact-Check pipeline is designed for editorial teams, fact-checking organisations, and trust & safety teams who need a structured, auditable verdict on news or editorial content. It extracts verifiable claims from the article, assesses the credibility of named sources and publications, evaluates claims against available knowledge, and routes ambiguous or contested findings to a panel of professional crowd fact-checkers for final judgment. The combination of systematic AI claim parsing and experienced human research judgment is particularly valuable for articles that make specific factual assertions about events, statistics, or named individuals — precisely the content where an automated-only verdict carries the most risk.

Pipeline Details

Property	Value
Pipeline ID	`verify-news-article`
Tier	2
Estimated Duration	~150 min
Credit Cost	3,900 credits per run
Crowd Stages	Yes
Files Required	No — article content can be supplied via context instead

Required Inputs

Files

A file upload is optional. You may submit a text file or PDF containing the article. Alternatively, provide the article content via the article_url or article_text context keys. At least one content source — file, URL, or raw text — is required to start a run.

Context Keys

Key	Required	Description
`article_url`	Required if no file	URL of the article to verify; the pipeline will fetch and parse the content
`article_text`	Required if no file and no URL	Full text of the article as a plain string
`claimed_author`	Optional	Name of the article's reported author
`publication_date`	Optional	Date the article was published, ISO 8601 format
`publication_name`	Optional	Name of the publication or outlet (e.g. `"Example News"`)
`topic`	Optional	Subject matter category for focused claim extraction (e.g. `"election results"`, `"public health"`)

Providing publication_name and topic significantly improves claim extraction focus. The topic key helps the AI prioritise the most factually consequential claims within the article rather than extracting every incidental assertion.

Stages

Parse & Extract (AI)

The pipeline begins by parsing the article content from whichever source was provided — file, URL, or raw text — and normalising it into a structured representation. The AI then extracts named entities (people, organisations, locations, dates), identifies the article's primary claims and sub-claims, and catalogues all named sources cited. The output is a structured claim manifest that drives the analysis in subsequent stages.

Source Credibility (AI)

Each named source cited in the article is evaluated for credibility signals: publication history, institutional affiliation, domain expertise relative to the claims attributed to them, and consistency with other known sources in the AI's knowledge base. Anonymous or unverifiable sources are flagged explicitly. The stage outputs a source-level credibility profile that crowd workers and the final synthesis use to weight individual claims.

Claim Assessment (AI)

The extracted claims are evaluated individually against available AI knowledge. For each claim, the AI assigns an initial assessment (supported, contradicted, unverifiable, or partially supported), notes supporting or contradicting evidence it can reference, and flags claims that fall outside its knowledge cutoff or require live research to assess. Claims flagged as requiring human verification are highlighted in the crowd brief for the next stage.

Researcher Review (Crowd)

Three crowd workers holding the fact_checker skill tag independently assess the article and the AI's structured findings. Workers are professional or trained fact-checkers who apply their own research judgment to the claims, source assessments, and AI flags. Each worker completes a detailed SurveyJS questionnaire covering their independent assessment of the article's overall veracity, their evaluation of the highest-risk claims, and any sources or context they are able to verify or contradict. The stage gates on 67% consensus (at least two of three workers in agreement on the overall verdict) before proceeding. Researcher review is the primary latency driver for this pipeline.

Final Verdict (AI Synthesis)

The AI aggregates the claim assessment scores, source credibility profiles, and crowd researcher responses into a final veracity verdict. The verdict (verified, misleading, false, or inconclusive) is accompanied by a confidence score, a breakdown of verified versus contested claims, a summary of the source credibility assessment, and a structured fact-check report suitable for publication or inclusion in editorial documentation.

Starting a Run

curl -X POST "https://api.crowdee.ai/v2/projects/{projectId}/verification-runs" \
  -H "X-API-Key: crw_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "pipeline_id": "verify-news-article",
    "file_ids": [],
    "context": {
      "article_url": "https://example-news.com/article/2024-11-15",
      "publication_name": "Example News",
      "topic": "election results"
    }
  }'

Example Response

{
  "id": "run_01j9x...",
  "status": "completed",
  "verdict": "authentic",
  "confidence": 81,
  "pipeline_id": "verify-news-article"
}

This is the longest Tier 2 pipeline, with an estimated duration of ~150 minutes due to three sequential AI stages and a crowd researcher review step. Implement asynchronous polling against GET /v2/projects/{projectId}/verification-runs/{runId} or use webhooks — do not hold an HTTP connection open for the full duration.

You can supply the article as a file upload (text or PDF), via article_url, or as raw article_text in the context. At least one source is required. If both a file and a URL are provided, the file takes precedence.