
{"id":194693,"date":"2026-10-29T12:00:43","date_gmt":"2026-10-29T11:00:43","guid":{"rendered":"https:\/\/uniavisen.dk\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/"},"modified":"2026-10-29T11:18:57","modified_gmt":"2026-10-29T10:18:57","slug":"bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar","status":"publish","type":"post","link":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/","title":{"rendered":"Do you use chatbots to find information? Then you miss out on knowledge"},"content":{"rendered":"<p>Do you ask ChatGPT instead of Googling when you want to win an argument over dinner?<\/p>\n<div class=\"factbox\">\n<p class=\"factbox-header feature-color\">About the study<\/p>\n<p><strong>In the study\u00a0<\/strong><a href=\"https:\/\/arxiv.org\/abs\/2510.04226\"><em>What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models,<\/em><\/a> researchers from the University of Copenhagen, Aalborg University, Stanford University, the University of Colorado Boulder, and the University of Texas at Austin examined 27 large language models developed by OpenAI, Meta, Google, and Alibaba.<\/p>\n<p><strong>The researchers<\/strong> tested 155 topics related to 12 different countries. The language models generated approximately 1.7 million answers and around 70 million individual statements. Smaller models generate more diverse content than larger models, and newer models perform better than older ones. Overall, however, all of them perform worse than traditional search engines.<\/p>\n<p><em>Source: <a href=\"https:\/\/science.ku.dk\/presse\/nyheder\/2026\/ai-chatbots-giver-os-et-smalt-udvalg-af-viden-vi-risikerer-et-videnskollaps-siger-forskere\/\">KU.dk<\/a><\/em><\/p>\n<\/div>\n<p>If so, you\u2019re certainly not alone. Language models are becoming increasingly hard to avoid when we need a quick answer to a question. Even Google has introduced AI Overviews, which appear at the top of the page when we use the search engine.<\/p>\n<p>But this is not without its problems, according to a new study, <a href=\"https:\/\/arxiv.org\/abs\/2510.04226\"><em>What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models<\/em><\/a>, co-authored by Isabelle Augenstein, a professor at the University of Copenhagen. The researchers examined how well 27 different language models perform when it comes to providing a broad range of answers to different questions.<\/p>\n<p>The study found that although newer language models perform better than older ones, even the best models are 18 to 19 per cent worse than a conventional internet search. This sets off alarm bells: if language models take over too much of the way we search for knowledge, it could ultimately lead to what the researchers call a collapse in knowledge diversity.<\/p>\n<p>This means that access to the original diversity of human knowledge is increasingly filtered through a narrower and narrower selection of perspectives.<\/p>\n<p>\u00bbIf a topic isn\u2019t sufficiently popular, people may ultimately be unable to find any information about what they\u2019re researching, or the information they do find may be seriously incomplete,\u00ab Isabelle Augenstein tells The University Post.<\/p>\n<h3>A Field in Flux<\/h3>\n<p>Isabelle Augenstein is a professor and deputy head of department at the Department of Computer Science (DIKU). She began researching AI long before it became the hottest topic of the day, back when Google Translate was the most visible example of the potential of language models. Her parents were therefore very sceptical when, in 2008, she chose to pursue a bachelor&#8217;s degree in AI:<\/p>\n<p>\u00bbThey thought I would never get a job with my degree. To them, AI sounded like something out of science fiction that would never work in the real world.\u00ab<\/p>\n<p>Today, Isabelle Augenstein\u2019s field of research is developing at a rapid pace, and the professor spends most of her time investigating the social implications of using language models.<\/p>\n<blockquote><p>In reality, the models are just a sophisticated piece of statistics that calculates the probability of the next word.<\/p>\n<p class=\"quotee\">Isabelle Augenstein<\/p>\n<\/blockquote>\n<p>In the new study, Isabelle Augenstein and her colleagues examined how much the knowledge we obtain differs when we use a language model instead of a search engine. The researchers were inspired to pursue the project after coming across an earlier study in which the authors trained AI on the models\u2019 own responses, thereby simulating a \u2018knowledge collapse\u2019:<\/p>\n<p>\u00bbTheir study was very theoretical. So we thought it would be interesting to investigate whether a knowledge collapse has already occurred by looking at both older and newer language models,\u00ab she says.<\/p>\n<h3>A Sophisticated Piece of Statistics<\/h3>\n<p>In the study, the researchers used both older language models from 2023 and the latest models from companies such as OpenAI, Meta, and Google to search for various concepts and topics. These ranged from broad, widely known subjects such as democracy and gender equality to narrower topics specific to individual countries. The researchers compared the answers they received from the language models with the information they obtained through a conventional search engine.<\/p>\n<p>The results showed that even the best language models provided a narrower and less diverse selection of knowledge than a search engine. According to Isabelle Augenstein, this is due to the way language models are built.<\/p>\n<p>\u00bbWhen you use AI models as search engines, you get summaries of a possible answer. In reality, the models are just a sophisticated piece of statistics that calculates the probability of the next word,\u00ab says Isabelle Augenstein, continuing:<\/p>\n<p>\u00bbOver time, they have become much better, and additional layers have been added, for example, to make them generate answers they think people will like. But the basic technique is that the model selects the most likely outcome and forgets the rest.\u00ab<\/p>\n<p>The models are also subject to a space constraint, which means they have to select sources and facts. Here, language models prioritise claims they find in the most popular sources.<\/p>\n<p>\u00bbThe consequence is that less popular claims and sources do not appear at all in the language model\u2019s answers,\u00ab says Isabelle Augenstein.<\/p>\n<h3>Implications for Denmark<\/h3>\n<p>The internet is already increasingly filled with texts written entirely or partly by language models. And this only fuels the potential collapse of knowledge:<\/p>\n<blockquote><p>\u00bbWe could end up in a situation where the views and opinions of small groups are excluded from society because it is impossible to find any information about them.\u00ab<\/p>\n<p class=\"quotee\">Isabelle Augenstein<\/p>\n<\/blockquote>\n<p>\u00bbMany people use language models to write texts. When they do, those answers end up online and become part of the material future language models are trained on. When both the new answers and the new knowledge are AI-generated, it could ultimately lead to a collapse in knowledge,\u00ab says Isabelle Augenstein.<\/p>\n<p>The professor emphasises that we have not yet seen any examples of a knowledge collapse and that the new language models are far better than the old ones. However, we still need to be aware of the risk, because all the mechanisms required for such a collapse to occur are already in place:<\/p>\n<p>\u00bbWhen AI models are already reusing texts that other models have been trained on because those texts are available as sources online, a knowledge collapse can happen over time.\u00ab<\/p>\n<p>This would have particularly significant implications for a country like Denmark, where the language community is relatively small. This is because the amount of knowledge and written material available in Danish is limited:<\/p>\n<p>\u00bbWe could end up in a situation where the views and opinions of small groups are excluded from society because it is impossible to find any information about them.\u00ab<\/p>\n<h3>Training new language models on old books<\/h3>\n<p>The companies developing language models are aware of the issue and are working to ensure that their models are not trained on AI-generated data. However, they are running out of training data from the internet and have therefore begun buying up large numbers of old books and scanning them so that the models can be trained on them. This practice has been covered by<a href=\"https:\/\/www.berlingske.dk\/litteratur\/antikvariater-har-pludselig-faaet-en-helt-ny-kunde--og-de-koeber-boeger-i-stakkevis?gaa_at=eafs&amp;gaa_n=AXbw80RVf2T5hoyQRWKqFA02-VHyFoCLYFPgQlN9tn4ng5oIsyIgrsOlUqCv&amp;gaa_ts=6adb3b5e&amp;gaa_sig=S4_uUa7Uyu6Edf6S8lpLoZuJ4zyPz2TGOVfRrE8s7SIr837-H9ugPYpGnlnLBNTsMpO6-eDalQlL3iz0lBE3uA%3D%3D\"> several media outlets.<\/a><\/p>\n<p>\u00bbThese companies are making a major effort to ensure that language models continually acquire new knowledge. That is why I wasn\u2019t surprised that the new language models are actually better than the old ones,\u00ab says Isabelle Augenstein.<\/p>\n<p>Since the companies behind language models are aware of the problem, it is possible that a knowledge collapse will never occur. But although Isabelle Augenstein is unsure whether we will ever experience one, she still finds the development concerning.<\/p>\n<p>\u00bbIt is about what we as human beings actually know and how we acquire our knowledge. That is why it is important to explore some of the searches we make in greater depth.\u00ab<\/p>\n<p>According to the professor, there are several ways to do this. For example, by asking language models follow-up questions and requesting additional sources and claims:<\/p>\n<p>\u00bbMany people assume that language models generate their answers based on all the sources available online, but that is not the case. They generate answers based on the most popular sources on the internet.\u00ab<\/p>\n<h3>Remember to Check Your Sources<\/h3>\n<p>Isabelle Augenstein highlights two particular problems with the use of AI. First, language models can hallucinate and make up answers that are incorrect. Second, the models have an inherent bias towards pleasing their users.<\/p>\n<blockquote><p>\u00bbIt is about what we as human beings actually know and how we acquire our knowledge.\u00ab<\/p>\n<p class=\"quotee\">Isabelle Augenstein<\/p>\n<p>&nbsp;<\/p><\/blockquote>\n<p>\u00bbYou can end up convincing the language model that what it says is wrong, and then it changes its answer, even when the original answer was correct.\u00ab<\/p>\n<p>According to Isabelle Augenstein, the problem is made worse by the fact that Google has now introduced its own AI-generated summaries:<\/p>\n<p>\u00bbThis could mean that most people end up looking only at the AI summary instead of reading the actual sources. You should always remember that the less popular the topic you are researching is, the more important it is to examine the sources yourself.\u00ab<\/p>\n<p><em>So you\u2019re not encouraging people to stop using language models to search for information?<\/em><\/p>\n<p>\u00bbNo, I think they are fine for getting a quick overview. They are good at summarising information and providing short answers. But I encourage people to explore the sources in greater depth themselves rather than blindly trusting language models.\u00ab<\/p>\n<p><em>This article was translated with the assistance of artificial intelligence and subsequently reviewed by a member of the editorial staff.<\/em><br \/>\n<!-- end of module 1 --><\/p>\n","protected":false},"excerpt":{"rendered":"<p>A new study from the Department of Computer Science shows that using language models to search for information risks leading to a collapse in knowledge diversity.  AI answers are less diverse and less in-depth than a regular Google search \u2014 and it has consequences.<\/p>\n","protected":false},"author":107,"featured_media":161060,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_uf_show_specific_survey":0,"_uf_disable_surveys":false,"footnotes":""},"categories":[46],"tags":[6108,8943,8944,8945],"class_list":["post-194693","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-science","tag-ai-en","tag-department-of-computer-science","tag-knowledge-collaps","tag-large-language-models","expression-news_article"],"acf":[],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v28.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>Do you use chatbots to find information? Then you miss out on knowledge<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Do you use chatbots to find information? Then you miss out on knowledge\" \/>\n<meta property=\"og:description\" content=\"A new study from the Department of Computer Science shows that using language models to search for information risks leading to a collapse in knowledge diversity. AI answers are less diverse and less in-depth than a regular Google search \u2014 and it has consequences.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/\" \/>\n<meta property=\"og:site_name\" content=\"University Post\" \/>\n<meta property=\"article:publisher\" content=\"https:\/\/www.facebook.com\/uniavis\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-29T11:00:43+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-scaled.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"2560\" \/>\n\t<meta property=\"og:image:height\" content=\"2560\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"author\" content=\"Johanne Dahl Karlsen\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@Uniavisen\" \/>\n<meta name=\"twitter:site\" content=\"@Uniavisen\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Johanne Dahl Karlsen\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"9 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/\"},\"author\":{\"name\":\"Johanne Dahl Karlsen\",\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/#\\\/schema\\\/person\\\/01dc6a59ff1fda55a380ff01dc1f4457\"},\"headline\":\"Do you use chatbots to find information? Then you miss out on knowledge\",\"datePublished\":\"2026-10-29T11:00:43+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/\"},\"wordCount\":1551,\"image\":{\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/uniavisen.dk\\\/wp-content\\\/uploads\\\/2024\\\/03\\\/isabellebillede-scaled.jpg\",\"keywords\":[\"AI\",\"department of computer science\",\"knowledge collaps\",\"large language models\"],\"articleSection\":[\"Science\"],\"inLanguage\":\"en-US\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/\",\"url\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/\",\"name\":\"Do you use chatbots to find information? Then you miss out on knowledge\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/uniavisen.dk\\\/wp-content\\\/uploads\\\/2024\\\/03\\\/isabellebillede-scaled.jpg\",\"datePublished\":\"2026-10-29T11:00:43+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/#\\\/schema\\\/person\\\/01dc6a59ff1fda55a380ff01dc1f4457\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/#primaryimage\",\"url\":\"https:\\\/\\\/uniavisen.dk\\\/wp-content\\\/uploads\\\/2024\\\/03\\\/isabellebillede-scaled.jpg\",\"contentUrl\":\"https:\\\/\\\/uniavisen.dk\\\/wp-content\\\/uploads\\\/2024\\\/03\\\/isabellebillede-scaled.jpg\",\"width\":2560,\"height\":2560,\"caption\":\"Isabelle Augenstein er professor p\u00e5 Datalogisk Institut, SCIENCE\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Do you use chatbots to find information? Then you miss out on knowledge\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/#website\",\"url\":\"https:\\\/\\\/uniavisen.dk\\\/\",\"name\":\"University Post\",\"description\":\"Independent of management\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/uniavisen.dk\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/uniavisen.dk\\\/#\\\/schema\\\/person\\\/01dc6a59ff1fda55a380ff01dc1f4457\",\"name\":\"Johanne Dahl Karlsen\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/d4d9b5dddb77cea2ac19e3e80ebf0aadc7ed5310ede7f93de099f384be5e40cd?s=96&d=identicon&r=g\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/d4d9b5dddb77cea2ac19e3e80ebf0aadc7ed5310ede7f93de099f384be5e40cd?s=96&d=identicon&r=g\",\"contentUrl\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/d4d9b5dddb77cea2ac19e3e80ebf0aadc7ed5310ede7f93de099f384be5e40cd?s=96&d=identicon&r=g\",\"caption\":\"Johanne Dahl Karlsen\"},\"url\":\"https:\\\/\\\/uniavisen.dk\\\/en\\\/author\\\/johanne_dk\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Do you use chatbots to find information? Then you miss out on knowledge","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/","og_locale":"en_US","og_type":"article","og_title":"Do you use chatbots to find information? Then you miss out on knowledge","og_description":"A new study from the Department of Computer Science shows that using language models to search for information risks leading to a collapse in knowledge diversity. AI answers are less diverse and less in-depth than a regular Google search \u2014 and it has consequences.","og_url":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/","og_site_name":"University Post","article_publisher":"https:\/\/www.facebook.com\/uniavis","article_published_time":"2026-10-29T11:00:43+00:00","og_image":[{"width":2560,"height":2560,"url":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-scaled.jpg","type":"image\/jpeg"}],"author":"Johanne Dahl Karlsen","twitter_card":"summary_large_image","twitter_creator":"@Uniavisen","twitter_site":"@Uniavisen","twitter_misc":{"Written by":"Johanne Dahl Karlsen","Est. reading time":"9 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/#article","isPartOf":{"@id":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/"},"author":{"name":"Johanne Dahl Karlsen","@id":"https:\/\/uniavisen.dk\/#\/schema\/person\/01dc6a59ff1fda55a380ff01dc1f4457"},"headline":"Do you use chatbots to find information? Then you miss out on knowledge","datePublished":"2026-10-29T11:00:43+00:00","mainEntityOfPage":{"@id":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/"},"wordCount":1551,"image":{"@id":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/#primaryimage"},"thumbnailUrl":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-scaled.jpg","keywords":["AI","department of computer science","knowledge collaps","large language models"],"articleSection":["Science"],"inLanguage":"en-US"},{"@type":"WebPage","@id":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/","url":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/","name":"Do you use chatbots to find information? Then you miss out on knowledge","isPartOf":{"@id":"https:\/\/uniavisen.dk\/#website"},"primaryImageOfPage":{"@id":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/#primaryimage"},"image":{"@id":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/#primaryimage"},"thumbnailUrl":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-scaled.jpg","datePublished":"2026-10-29T11:00:43+00:00","author":{"@id":"https:\/\/uniavisen.dk\/#\/schema\/person\/01dc6a59ff1fda55a380ff01dc1f4457"},"breadcrumb":{"@id":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/#primaryimage","url":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-scaled.jpg","contentUrl":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-scaled.jpg","width":2560,"height":2560,"caption":"Isabelle Augenstein er professor p\u00e5 Datalogisk Institut, SCIENCE"},{"@type":"BreadcrumbList","@id":"https:\/\/uniavisen.dk\/en\/bruger-du-chatten-til-at-finde-information-saa-faar-du-et-daarligere-svar\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/uniavisen.dk\/en\/"},{"@type":"ListItem","position":2,"name":"Do you use chatbots to find information? Then you miss out on knowledge"}]},{"@type":"WebSite","@id":"https:\/\/uniavisen.dk\/#website","url":"https:\/\/uniavisen.dk\/","name":"University Post","description":"Independent of management","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/uniavisen.dk\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Person","@id":"https:\/\/uniavisen.dk\/#\/schema\/person\/01dc6a59ff1fda55a380ff01dc1f4457","name":"Johanne Dahl Karlsen","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/secure.gravatar.com\/avatar\/d4d9b5dddb77cea2ac19e3e80ebf0aadc7ed5310ede7f93de099f384be5e40cd?s=96&d=identicon&r=g","url":"https:\/\/secure.gravatar.com\/avatar\/d4d9b5dddb77cea2ac19e3e80ebf0aadc7ed5310ede7f93de099f384be5e40cd?s=96&d=identicon&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/d4d9b5dddb77cea2ac19e3e80ebf0aadc7ed5310ede7f93de099f384be5e40cd?s=96&d=identicon&r=g","caption":"Johanne Dahl Karlsen"},"url":"https:\/\/uniavisen.dk\/en\/author\/johanne_dk\/"}]}},"advancedCustomFields":{"expression":{"term_id":15,"name":"News Article","slug":"news_article","term_group":0,"term_taxonomy_id":15,"taxonomy":"expression","description":"","parent":0,"count":11558,"filter":"raw"},"enable_comments":true,"align_content":"alignleft","feature_color":"","article_updated":"","layout_group":[{"acf_fc_layout":"Headline","use_post_title":true,"headline":"","style":"default","highlighted_words":"","text_size":"small"},{"acf_fc_layout":"Image","image":{"ID":161059,"id":161059,"title":"Isabellebillede","filename":"isabellebillede-scaled.jpg","filesize":1049260,"url":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-scaled.jpg","link":"https:\/\/uniavisen.dk\/en\/paa-reddit-bag-rattet-og-i-aids-kulturhistorie-tre-projekter-om-ligestilling-paa-ku\/isabellebillede\/","alt":"","author":"99","description":"Isabelle Augenstein er professor p\u00e5 Datalogisk Institut, SCIENCE","caption":"Isabelle Augenstein er professor p\u00e5 Datalogisk Institut, SCIENCE","name":"isabellebillede","status":"inherit","uploaded_to":161051,"date":"2024-03-07 09:51:56","modified":"2024-03-07 10:47:33","menu_order":0,"mime_type":"image\/jpeg","type":"image","subtype":"jpeg","icon":"https:\/\/uniavisen.dk\/wp-includes\/images\/media\/default.png","width":2560,"height":2560,"sizes":{"thumbnail":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-150x150.jpg","thumbnail-width":150,"thumbnail-height":150,"medium":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-480x480.jpg","medium-width":480,"medium-height":480,"medium_large":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-768x768.jpg","medium_large-width":768,"medium_large-height":768,"large":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-1280x1280.jpg","large-width":1280,"large-height":1280,"1536x1536":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-1536x1536.jpg","1536x1536-width":1536,"1536x1536-height":1536,"2048x2048":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-2048x2048.jpg","2048x2048-width":2048,"2048x2048-height":2048,"featured-soft":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-290x290.jpg","featured-soft-width":290,"featured-soft-height":290,"featured-hard":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-290x180.jpg","featured-hard-width":290,"featured-hard-height":180,"narrow":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-700x700.jpg","narrow-width":700,"narrow-height":700,"extended":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-990x990.jpg","extended-width":990,"extended-height":990}},"style":"full","text_placement":"metadata-below","image_link_url":"","image_link_title":"","caption_prefix":"","enable_alternative_caption":true,"alternative_caption":"Isabelle Augenstein is a professor at the Department of Computer Science, Faculty of SCIENCE."},{"acf_fc_layout":"Standfirst","subject":"Collapse","text":"This is a story about what will follow","use_post_excerpt":true},{"acf_fc_layout":"Byline","is_author":true,"contributors":false},{"acf_fc_layout":"Content","content":"<p>Do you ask ChatGPT instead of Googling when you want to win an argument over dinner?<\/p>\n<div class=\"factbox\">\n<p class=\"factbox-header feature-color\">About the study<\/p>\n<p><strong>In the study\u00a0<\/strong><a href=\"https:\/\/arxiv.org\/abs\/2510.04226\"><em>What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models,<\/em><\/a> researchers from the University of Copenhagen, Aalborg University, Stanford University, the University of Colorado Boulder, and the University of Texas at Austin examined 27 large language models developed by OpenAI, Meta, Google, and Alibaba.<\/p>\n<p><strong>The researchers<\/strong> tested 155 topics related to 12 different countries. The language models generated approximately 1.7 million answers and around 70 million individual statements. Smaller models generate more diverse content than larger models, and newer models perform better than older ones. Overall, however, all of them perform worse than traditional search engines.<\/p>\n<p><em>Source: <a href=\"https:\/\/science.ku.dk\/presse\/nyheder\/2026\/ai-chatbots-giver-os-et-smalt-udvalg-af-viden-vi-risikerer-et-videnskollaps-siger-forskere\/\">KU.dk<\/a><\/em><\/p>\n<\/div>\n<p>If so, you\u2019re certainly not alone. Language models are becoming increasingly hard to avoid when we need a quick answer to a question. Even Google has introduced AI Overviews, which appear at the top of the page when we use the search engine.<\/p>\n<p>But this is not without its problems, according to a new study, <a href=\"https:\/\/arxiv.org\/abs\/2510.04226\"><em>What and Whose Knowledge? Measuring Epistemic Diversity in Large Language Models<\/em><\/a>, co-authored by Isabelle Augenstein, a professor at the University of Copenhagen. The researchers examined how well 27 different language models perform when it comes to providing a broad range of answers to different questions.<\/p>\n<p>The study found that although newer language models perform better than older ones, even the best models are 18 to 19 per cent worse than a conventional internet search. This sets off alarm bells: if language models take over too much of the way we search for knowledge, it could ultimately lead to what the researchers call a collapse in knowledge diversity.<\/p>\n<p>This means that access to the original diversity of human knowledge is increasingly filtered through a narrower and narrower selection of perspectives.<\/p>\n<p>\u00bbIf a topic isn\u2019t sufficiently popular, people may ultimately be unable to find any information about what they\u2019re researching, or the information they do find may be seriously incomplete,\u00ab Isabelle Augenstein tells The University Post.<\/p>\n<h3>A Field in Flux<\/h3>\n<p>Isabelle Augenstein is a professor and deputy head of department at the Department of Computer Science (DIKU). She began researching AI long before it became the hottest topic of the day, back when Google Translate was the most visible example of the potential of language models. Her parents were therefore very sceptical when, in 2008, she chose to pursue a bachelor&#8217;s degree in AI:<\/p>\n<p>\u00bbThey thought I would never get a job with my degree. To them, AI sounded like something out of science fiction that would never work in the real world.\u00ab<\/p>\n<p>Today, Isabelle Augenstein\u2019s field of research is developing at a rapid pace, and the professor spends most of her time investigating the social implications of using language models.<\/p>\n<blockquote><p>In reality, the models are just a sophisticated piece of statistics that calculates the probability of the next word.<\/p>\n<p class=\"quotee\">Isabelle Augenstein<\/p>\n<\/blockquote>\n<p>In the new study, Isabelle Augenstein and her colleagues examined how much the knowledge we obtain differs when we use a language model instead of a search engine. The researchers were inspired to pursue the project after coming across an earlier study in which the authors trained AI on the models\u2019 own responses, thereby simulating a \u2018knowledge collapse\u2019:<\/p>\n<p>\u00bbTheir study was very theoretical. So we thought it would be interesting to investigate whether a knowledge collapse has already occurred by looking at both older and newer language models,\u00ab she says.<\/p>\n<h3>A Sophisticated Piece of Statistics<\/h3>\n<p>In the study, the researchers used both older language models from 2023 and the latest models from companies such as OpenAI, Meta, and Google to search for various concepts and topics. These ranged from broad, widely known subjects such as democracy and gender equality to narrower topics specific to individual countries. The researchers compared the answers they received from the language models with the information they obtained through a conventional search engine.<\/p>\n<p>The results showed that even the best language models provided a narrower and less diverse selection of knowledge than a search engine. According to Isabelle Augenstein, this is due to the way language models are built.<\/p>\n<p>\u00bbWhen you use AI models as search engines, you get summaries of a possible answer. In reality, the models are just a sophisticated piece of statistics that calculates the probability of the next word,\u00ab says Isabelle Augenstein, continuing:<\/p>\n<p>\u00bbOver time, they have become much better, and additional layers have been added, for example, to make them generate answers they think people will like. But the basic technique is that the model selects the most likely outcome and forgets the rest.\u00ab<\/p>\n<p>The models are also subject to a space constraint, which means they have to select sources and facts. Here, language models prioritise claims they find in the most popular sources.<\/p>\n<p>\u00bbThe consequence is that less popular claims and sources do not appear at all in the language model\u2019s answers,\u00ab says Isabelle Augenstein.<\/p>\n<h3>Implications for Denmark<\/h3>\n<p>The internet is already increasingly filled with texts written entirely or partly by language models. And this only fuels the potential collapse of knowledge:<\/p>\n<blockquote><p>\u00bbWe could end up in a situation where the views and opinions of small groups are excluded from society because it is impossible to find any information about them.\u00ab<\/p>\n<p class=\"quotee\">Isabelle Augenstein<\/p>\n<\/blockquote>\n<p>\u00bbMany people use language models to write texts. When they do, those answers end up online and become part of the material future language models are trained on. When both the new answers and the new knowledge are AI-generated, it could ultimately lead to a collapse in knowledge,\u00ab says Isabelle Augenstein.<\/p>\n<p>The professor emphasises that we have not yet seen any examples of a knowledge collapse and that the new language models are far better than the old ones. However, we still need to be aware of the risk, because all the mechanisms required for such a collapse to occur are already in place:<\/p>\n<p>\u00bbWhen AI models are already reusing texts that other models have been trained on because those texts are available as sources online, a knowledge collapse can happen over time.\u00ab<\/p>\n<p>This would have particularly significant implications for a country like Denmark, where the language community is relatively small. This is because the amount of knowledge and written material available in Danish is limited:<\/p>\n<p>\u00bbWe could end up in a situation where the views and opinions of small groups are excluded from society because it is impossible to find any information about them.\u00ab<\/p>\n<h3>Training new language models on old books<\/h3>\n<p>The companies developing language models are aware of the issue and are working to ensure that their models are not trained on AI-generated data. However, they are running out of training data from the internet and have therefore begun buying up large numbers of old books and scanning them so that the models can be trained on them. This practice has been covered by<a href=\"https:\/\/www.berlingske.dk\/litteratur\/antikvariater-har-pludselig-faaet-en-helt-ny-kunde--og-de-koeber-boeger-i-stakkevis?gaa_at=eafs&amp;gaa_n=AXbw80RVf2T5hoyQRWKqFA02-VHyFoCLYFPgQlN9tn4ng5oIsyIgrsOlUqCv&amp;gaa_ts=6adb3b5e&amp;gaa_sig=S4_uUa7Uyu6Edf6S8lpLoZuJ4zyPz2TGOVfRrE8s7SIr837-H9ugPYpGnlnLBNTsMpO6-eDalQlL3iz0lBE3uA%3D%3D\"> several media outlets.<\/a><\/p>\n<p>\u00bbThese companies are making a major effort to ensure that language models continually acquire new knowledge. That is why I wasn\u2019t surprised that the new language models are actually better than the old ones,\u00ab says Isabelle Augenstein.<\/p>\n<p>Since the companies behind language models are aware of the problem, it is possible that a knowledge collapse will never occur. But although Isabelle Augenstein is unsure whether we will ever experience one, she still finds the development concerning.<\/p>\n<p>\u00bbIt is about what we as human beings actually know and how we acquire our knowledge. That is why it is important to explore some of the searches we make in greater depth.\u00ab<\/p>\n<p>According to the professor, there are several ways to do this. For example, by asking language models follow-up questions and requesting additional sources and claims:<\/p>\n<p>\u00bbMany people assume that language models generate their answers based on all the sources available online, but that is not the case. They generate answers based on the most popular sources on the internet.\u00ab<\/p>\n<h3>Remember to Check Your Sources<\/h3>\n<p>Isabelle Augenstein highlights two particular problems with the use of AI. First, language models can hallucinate and make up answers that are incorrect. Second, the models have an inherent bias towards pleasing their users.<\/p>\n<blockquote><p>\u00bbIt is about what we as human beings actually know and how we acquire our knowledge.\u00ab<\/p>\n<p class=\"quotee\">Isabelle Augenstein<\/p>\n<p>&nbsp;<\/p><\/blockquote>\n<p>\u00bbYou can end up convincing the language model that what it says is wrong, and then it changes its answer, even when the original answer was correct.\u00ab<\/p>\n<p>According to Isabelle Augenstein, the problem is made worse by the fact that Google has now introduced its own AI-generated summaries:<\/p>\n<p>\u00bbThis could mean that most people end up looking only at the AI summary instead of reading the actual sources. You should always remember that the less popular the topic you are researching is, the more important it is to examine the sources yourself.\u00ab<\/p>\n<p><em>So you\u2019re not encouraging people to stop using language models to search for information?<\/em><\/p>\n<p>\u00bbNo, I think they are fine for getting a quick overview. They are good at summarising information and providing short answers. But I encourage people to explore the sources in greater depth themselves rather than blindly trusting language models.\u00ab<\/p>\n<p><em>This article was translated with the assistance of artificial intelligence and subsequently reviewed by a member of the editorial staff.<\/em><\/p>\n"},{"acf_fc_layout":"ArticleEnd"},{"acf_fc_layout":"Newsletter","lang_select":"da","identifier":"Nyhedsbrev","headline":"Modtag et ugentligt nyhedsoverblik i din inbox","button_text":"Tilmeld nu","class":""},{"acf_fc_layout":"OtherStories","headline":"","hand_picked_posts":false,"references":false,"category":false,"theme":false,"number_of_posts":"4","style":"default"}]},"taxonomyData":{"category":[{"term_id":46,"name":"Science","slug":"science","term_group":0,"term_taxonomy_id":46,"taxonomy":"category","description":"","parent":0,"count":842,"filter":"raw"}],"post_tag":[{"term_id":6108,"name":"AI","slug":"ai-en","term_group":0,"term_taxonomy_id":6108,"taxonomy":"post_tag","description":"","parent":0,"count":10,"filter":"raw"},{"term_id":8943,"name":"department of computer science","slug":"department-of-computer-science","term_group":0,"term_taxonomy_id":8943,"taxonomy":"post_tag","description":"","parent":0,"count":1,"filter":"raw"},{"term_id":8944,"name":"knowledge collaps","slug":"knowledge-collaps","term_group":0,"term_taxonomy_id":8944,"taxonomy":"post_tag","description":"","parent":0,"count":1,"filter":"raw"},{"term_id":8945,"name":"large language models","slug":"large-language-models","term_group":0,"term_taxonomy_id":8945,"taxonomy":"post_tag","description":"","parent":0,"count":1,"filter":"raw"}],"post_format":[],"expression":[{"term_id":15,"name":"News Article","slug":"news_article","term_group":0,"term_taxonomy_id":15,"taxonomy":"expression","description":"","parent":0,"count":11558,"filter":"raw"}]},"featured_media_url":"https:\/\/uniavisen.dk\/wp-content\/uploads\/2024\/03\/isabellebillede-1280x1280.jpg","_links":{"self":[{"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/posts\/194693","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/users\/107"}],"replies":[{"embeddable":true,"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/comments?post=194693"}],"version-history":[{"count":9,"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/posts\/194693\/revisions"}],"predecessor-version":[{"id":194716,"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/posts\/194693\/revisions\/194716"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/media\/161060"}],"wp:attachment":[{"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/media?parent=194693"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/categories?post=194693"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/uniavisen.dk\/en\/wp-json\/wp\/v2\/tags?post=194693"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}