AI Often Cites Web Pages It Never Actually Read

A study from Imperial College London found that search-augmented LLMs often answer from snippets without opening source pages. Across 225 facts changed after February 2026, over half of searches ended without opening any page, producing outdated answers and unfaithful citations.

• GPT-5.6 Luna and six open-weight models tested on 225 changed facts
• Without web search, GPT-5.6 Luna answered only 23.3% of 103 questions correctly
• In some cases over half of searches ended without opening a single page
• Even with correct evidence supplied directly, some models favored outdated facts