{"id":2395,"date":"2026-09-30T01:28:00","date_gmt":"2026-09-29T16:28:00","guid":{"rendered":"https:\/\/onepress.co.kr\/index.php\/briefing\/2026-09-30-ai-knowledge-diversity-en\/"},"modified":"2026-09-30T01:28:00","modified_gmt":"2026-09-29T16:28:00","slug":"2026-09-30-ai-knowledge-diversity-en","status":"publish","type":"briefing","link":"https:\/\/onepress.co.kr\/index.php\/briefing\/2026-09-30-ai-knowledge-diversity-en\/","title":{"rendered":"AI answers covered less knowledge than web search\u2014but this was not an accuracy test"},"content":{"rendered":"

2026-09-30 01:28 KST<\/strong><\/p>\n

Researchers measured knowledge diversity using 200 phrasings based on real user questions across 155 topics tied to 12 countries.<\/p>\n

Their metric counts variation in real-world claims. Fewer distinct claims does not automatically mean wrong answers, nor does it prove web results are more accurate.<\/p>\n

What was compared<\/h2>\n

The dataset covered 27 models from families including OpenAI, Meta, Google and Alibaba, yielding about 1.7 million answers and 70 million claims. Topics ranged from nuclear weapons and marriage to racism and K-pop across 12 national contexts.<\/p>\n

What 18.7% means<\/h2>\n

Even GPT-5, the most diverse model in this setup, produced at least 18.7% less measured claim diversity than a basic Google search. Results varied by model and topic; newer models were more diverse than older ones, and retrieval augmentation generally widened coverage.<\/p>\n

\"AI
This AI-generated image explains the topic; it is not a photograph of an actual event, observation or field site.<\/figcaption><\/figure>\n

Why call it knowledge collapse<\/h2>\n

The risk hypothesis is that similar AI summaries become a dominant gateway and later models learn from those outputs, repeatedly excluding less common perspectives. The authors do not claim collapse is already happening; they identify a mechanism that could enable it.<\/p>\n

How to use the finding<\/h2>\n

For consequential questions, a practical response is to move beyond one AI summary and inspect primary sources and genuinely different perspectives. The work was described as accepted for EMNLP 2026; metric replication, search conditions and non-English user behavior remain important tests.<\/p>\n

Primary source and independent checks<\/h2>\n

Primary source: arXiv \uc5f0\uad6c \uc6d0\uace0<\/a><\/p>\n

Institutional explanation: University of Copenhagen \uc5f0\uad6c \uc124\uba85<\/a><\/p>\n

Independent check: \uc5f0\uad6c\ud300 CopeNLU \uc11c\uc9c0 \ud398\uc774\uc9c0<\/a><\/p>\n

Read in other languages<\/h2>\n

\ud55c\uad6d\uc5b4<\/a> \u00b7 English<\/a> \u00b7 Espa\u00f1ol<\/a> \u00b7 \u65e5\u672c\u8a9e<\/a> \u00b7 Ti\u1ebfng Vi\u1ec7t<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

Across about 1.7 million answers and 70 million claims from 27 language models, even the most diverse model was at least 18.7% less diverse than basic web search. The study did not directly score truth or individual answer quality.<\/p>\n","protected":false},"featured_media":0,"template":"","meta":[],"class_list":["post-2395","briefing","type-briefing","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/onepress.co.kr\/index.php\/wp-json\/wp\/v2\/briefing\/2395","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/onepress.co.kr\/index.php\/wp-json\/wp\/v2\/briefing"}],"about":[{"href":"https:\/\/onepress.co.kr\/index.php\/wp-json\/wp\/v2\/types\/briefing"}],"wp:attachment":[{"href":"https:\/\/onepress.co.kr\/index.php\/wp-json\/wp\/v2\/media?parent=2395"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}