The Synthetic Web: Adversarially-Curated Mini-Internets for Diagnosing Epistemic Weaknesses of Language Agents
The paper 'The Synthetic Web' by Shrey Shah and Levent Ozgur examines AI agents using fake internets to test their reliability. The authors found that frontier models are catastrophically vulnerable; a single fake article placed at the top of search results caused accuracy to collapse from 65% to 18%. This demonstrates an 'epistemic weakness' where models treat search ranking as truth, leading them to accept misleading information without verification.