Fabricated Citations

Issue: Model Cites Sources That Don’t Exist

Frequency: Occasional

Symptoms

• Citation references non-existent document
• Made-up document titles or authors
• URLs that lead nowhere
• Plausible-sounding but fake references

Root Cause When model generates content beyond retrieved context, it may also generate citations for that content, creating entirely fabricated references.

Example

Query: "What research supports this treatment?"

Agent response: "According to the Johnson et al. (2023) study 
published in the Journal of Medical Research, the treatment 
showed 89% efficacy [1]."

Reality: 
- No such study exists
- No "Journal of Medical Research"
- Citation is completely fabricated

Result: User can't verify, may cite fake study themselves

Real Incidents

• Lawyers cited 6 fake ChatGPT-generated cases (Avianca lawsuit)
• Legal RAG tools generate plausible-looking fake citations

Mitigation Strategies

1. Citation validation: Verify all references exist
2. Closed-book prevention: Only cite from retrieved content
3. URL verification: Check URLs are valid before including
4. Source registry: Only allow citations to known sources
5. Citation generation from extraction: Build citations from metadata
6. Warning labels: Flag citations that couldn’t be verified

Detection

• Validate all citations against source index
• Check URLs for accessibility
• Cross-reference author/title combinations
• Flag citations not in retrieved documents

Mitigation Strategies

Prevention

1. Implement query-answer consistency validation: Decompose complex queries into atomic components and verify each component is addressed in the answer before returning. Use RAGAS Answer Relevancy metric (target: >0.75) with automatic re-generation for scores below threshold. Root cause mitigation: Prevents context-anchoring by explicitly binding answer generation to parsed query intent.

2. Apply multi-source consensus verification: Require answers synthesized from multiple sources to explicitly cite which sources support each claim and flag unresolved contradictions. Implementation: Use semantic similarity checks across source fragments to detect cherry-picked evidence patterns. Root cause: Ensures balanced representation of evidence when sources conflict.

3. Enforce comprehensive coverage checks: Implement structured extraction requiring explicit treatment of caveats, limitations, exceptions, and counterevidence for each claim type. Use template responses with mandatory caveat sections. Root cause: Prevents omission of qualifying information that would change user decision-making.

Detection & Response

1. Answer completeness monitoring: Measure coverage of query intents in generated answers. Track query decomposition rate (% of query components explicitly addressed) and flag responses with coverage <85%. Instrument RAG pipeline to log query-answer similarity scores per component. Alert on sustained scores <0.70.

2. Evidence balance scoring: For each answer, compute evidence distribution across sources and flag one-sided responses (>70% from single source on multi-source queries). Implement automated extraction of caveat/limitation mentions and track inclusion rates by query type. Target: >80% of medical/financial answers include relevant caveats.

Architecture Patterns

1. Query Intent Decomposition Graph: Parse complex queries into a DAG of atomic intents before retrieval. Each retrieved document is mapped to specific intent nodes. Answer generation must satisfy all leaf nodes. Validation layer computes coverage before response generation.

2. Evidence Consensus Engine: Maintain a fact graph where each claim is attributed to specific sources with confidence scores. Multi-source claims require consensus computation (intersection of sources supporting claim). Flagging layer surfaces contradictions to generation model.

3. Structured Response Templates: Use task-specific response schemas that enforce inclusion of: primary answer, supporting evidence, relevant caveats/exceptions, alternative interpretations, confidence bounds. Auto-flag template violations before user delivery.

Key Metrics

Alerts & Escalation

References

• Avianca Lawyers - 6 fake cases cited
• Stanford Legal RAG Hallucinations - Fabricated legal citations