Back to About

System Card

Technical documentation of our AI models, limitations, and methodology

Last updated: February 3, 2026 | Version 1.0

AI Models Used

Primary Analysis Models

  • GPT-4 (OpenAI)

    Used for comprehensive article analysis, credibility assessment, and bias detection

  • Claude 3.5 Sonnet (Anthropic)

    Used for enhanced analysis including behavioral influence detection and psychological framing

Model Selection Rationale

We use multiple models to cross-validate analyses and reduce single-model biases. GPT-4 provides broad coverage and strong reasoning capabilities, while Claude 3.5 offers detailed contextual analysis and nuanced understanding of persuasion techniques.

Model Limitations

Known Failure Modes

1. Sarcasm and Satire Detection

Limitation: AI models may misinterpret satirical content as sincere reporting or fail to detect subtle sarcasm in opinion pieces.

Impact: Satire may receive incorrect credibility scores. Sarcastic loaded language may be flagged literally.

Mitigation: We flag entertainment and satire sources where possible. Users should verify article type.

2. Language Limitations

Limitation: Analysis is limited to English language content. Non-English articles cannot be accurately assessed.

Impact: No coverage of foreign language news sources or international perspectives not published in English.

Mitigation: We clearly indicate language limitations in our documentation.

3. Context Window Constraints

Limitation: Very long articles (>15,000 words) may be truncated, causing analysis to miss key information in the latter portions.

Impact: Investigative long-form journalism may receive incomplete analysis.

Mitigation: We prioritize analysis of article openings where key claims typically appear.

4. Relative vs. Absolute Scoring

Limitation: Scores are comparative within our dataset, not absolute truth measures. A "high credibility" score means the article shows many credibility signals relative to others analyzed, not that all claims are verified.

Impact: Users may incorrectly interpret scores as definitive truth ratings.

Mitigation: We include disclaimers on all analysis pages and clearly define what scores represent.

5. Emerging News Events

Limitation: For breaking news, AI models cannot assess factual accuracy since facts are still being established. Early reporting may lack context.

Impact: Analysis of breaking news focuses more on tone and framing than credibility of specific claims.

Mitigation: We note article publication date and recommend checking for updates.

6. Technical and Scientific Content

Limitation: Highly technical articles in specialized fields (advanced physics, medicine, etc.) may exceed the model's domain expertise, leading to incomplete assessments.

Impact: Credibility assessments of technical claims may be less reliable than for general news.

Mitigation: We recommend consulting domain experts for technical topics.

7. Cultural Context

Limitation: AI models may miss cultural nuances, local context, or region-specific political dynamics that affect how news should be interpreted.

Impact: Bias detection may not account for all cultural framing conventions.

Mitigation: We acknowledge cultural limitations in our methodology documentation.

Evaluation Approach

How We Validate Accuracy

  • Manual Review: Sample of analyses reviewed by human editors for accuracy and reasonableness
  • Cross-Model Validation: Compare outputs from different AI models to identify inconsistencies
  • User Feedback: Monitor dispute submissions and user reports for systematic errors
  • Benchmark Datasets: Test against known cases of high/low credibility articles

Success Metrics

We evaluate our system on:

  • • Agreement with human expert assessments (target: 75%+ for credibility signals)
  • • Consistency across similar articles from same source
  • • Dispute rate (target: <5% of analyses disputed by publishers)
  • • False positive rate for bias detection (monitored quarterly)

Update Frequency

Model Updates

We review and update our AI models quarterly (every 3 months) or when significant improvements become available from model providers. Major methodology changes are announced and documented in our Methodology changelog.

System Card Updates

This System Card is reviewed and updated when we identify new limitations, change models, or modify our evaluation approach. All updates are date-stamped at the top of this page.

Versioning System

We use semantic versioning for our methodology:

  • Major version (X.0.0): Significant changes to analysis approach or AI models
  • Minor version (1.X.0): New analysis features or improved detection of specific signals
  • Patch version (1.0.X): Bug fixes, minor adjustments, or clarifications

Current Version: 1.0.0

Responsible AI Commitments

  • Transparency: We document our models, limitations, and methodology publicly
  • Accountability: We provide a disputes process for publishers and readers
  • Continuous Improvement: We regularly evaluate and improve our analysis quality
  • User Education: We clearly communicate what our analysis can and cannot do
  • Limitation Acknowledgment: We prominently display disclaimers and known limitations

Questions or Concerns?

If you have questions about our AI models, identified a limitation not documented here, or want to report a systematic error, please contact us:

General inquiries: support@auren.news

Dispute an analysis: disputes@auren.news

Cookie Preferences

Manage your cookie settings

We use cookies to enhance your experience, analyze site traffic, and personalize content. You can choose which cookies you allow. Essential cookies are required for basic site functionality.

Start analyzing news with confidence