{"id":224839,"date":"2026-03-30T11:08:58","date_gmt":"2026-03-30T11:08:58","guid":{"rendered":"https:\/\/www.9senses.ai\/?p=224839"},"modified":"2026-06-11T08:47:08","modified_gmt":"2026-06-11T08:47:08","slug":"lost-in-translation","status":"publish","type":"post","link":"https:\/\/www.9senses.ai\/de\/lost-in-translation\/","title":{"rendered":"Lost in Translation"},"content":{"rendered":"<div class=\"et_pb_section_0 et_pb_section et_section_regular et_flex_section preset--group--divi-section--divi-box-shadow--default preset--group--divi-section--divi-sizing--hsj9uxo--default\"><\/div>\n\n<div class=\"et_pb_section_1 et_pb_section et_section_regular et_flex_section preset--group--divi-section--divi-box-shadow--default preset--group--divi-section--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_row_0 et_pb_row et_pb_row_3-4_1-4 et_block_row et_block_row_3-4_1-4 preset--group--divi-row--divi-box-shadow--default preset--group--divi-row--divi-sizing--h1k452m--default\">\n<div class=\"et_pb_column_0 et_pb_column et_pb_column_3_4 et_block_column et_pb_css_mix_blend_mode_passthrough preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_text_0 et_pb_text et_pb_bg_layout_light et_pb_module et_block_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p>Du hast nicht ausreichende Berechtigungen, um auf diesen Inhalt zuzugreifen.<\/p>\n<\/div><\/div>\n\n<div class=\"et_pb_post_title_0 et_pb_post_title et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-post-title--divi-box-shadow--default preset--group--divi-post-title--divi-sizing--hsj9uxo--default preset--module--divi-post-title--y6glixiooo\"><div class=\"et_pb_title_container\"><h1 class=\"entry-title\">Lost in Translation<\/h1><\/div><\/div>\n\n<div class=\"et_pb_text_1 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p>Im Bereich der dialogorientierten KI dominiert die englische Sprache, was f\u00fcr andere Sprachen schwerwiegende strukturelle Folgen hat, die nur mit erheblichem Aufwand behoben werden k\u00f6nnen.<\/p>\n<\/div><\/div>\n\n<div class=\"et_pb_text_2 et_pb_text et_pb_bg_layout_light et_pb_module et_block_module preset--group--divi-text--divi-box-shadow--default preset--group--divi-text--divi-font-body--h1yjkjr--7p5s44libg preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p style=\"text-align: justify;\">Wenn man ChatGPT eine komplexe Frage auf Englisch stellt, kommt recht h\u00e4ufig eine korrekte, gut formulierte und zum Kontext passende Antwort. Wer dasselbe auf Hindi, Bengali oder Yoruba versucht, bekommt nicht selten eine k\u00fcrzere, weniger genaue und gelegentlich auch unsinnige Aussage. Auf Deutsch, Franz\u00f6sisch oder Spanisch sind die Antworten zwar treffender \u2013 erreichen aber weder die inhaltliche noch die sprachliche Qualit\u00e4t einer englischen Aussage.<\/p>\n<p style=\"text-align: justify;\">Generative AI has a language problem. And it doesn\u2019t only affect rare or endangered languages \u2014 it hits widely spoken ones too. The Brookings Institution describes the quality gap as a continuum: from English through European languages like German, French and Spanish, all the way to the roughly 7,000 languages spoken worldwide, of which only about 20 are considered \u201cdata-rich\u201d \u2014 with the gap widening dramatically as you move down the list. This is a problem that surfaces repeatedly in GenAI projects. Non-English systems struggle with precision, hallucinate more frequently, and simply fabricate content that doesn\u2019t exist.<\/p>\n<\/div><\/div>\n<\/div>\n\n<div class=\"et_pb_column_1 et_pb_column et_pb_column_1_4 et-last-child et_block_column et_pb_css_mix_blend_mode_passthrough preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_heading_0 et_pb_heading et_pb_module et_flex_module preset--group--divi-heading--divi-box-shadow--default\"><div class=\"et_pb_heading_container\"><h3 class=\"et_pb_module_header\">Thema<\/h3><\/div><\/div>\n\n<div class=\"et_pb_text_3 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p>Du hast nicht ausreichende Berechtigungen, um auf diesen Inhalt zuzugreifen.<\/p>\n<\/div><\/div>\n\n<div class=\"et_pb_heading_1 et_pb_heading et_pb_module et_flex_module preset--group--divi-heading--divi-box-shadow--default\"><div class=\"et_pb_heading_container\"><h3 class=\"et_pb_module_header\">Zusammenfassung<\/h3><\/div><\/div>\n\n<div class=\"et_pb_text_4 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p>Du hast nicht ausreichende Berechtigungen, um auf diesen Inhalt zuzugreifen.<\/p>\n<\/div><\/div>\n\n<div class=\"et_pb_text_5 et_pb_text et_pb_bg_layout_light et-interaction-target-6959wczed9 et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\" data-interaction-trigger=\"spzv7w7k2t\" data-interaction-target=\"6959wczed9\"><div class=\"et_pb_text_inner\"><p>Du hast nicht ausreichende Berechtigungen, um auf diesen Inhalt zuzugreifen.<\/p>\n<\/div><\/div>\n<\/div>\n<\/div>\n<\/div>\n\n<div class=\"et_pb_section_2 et_pb_section et_section_regular et_flex_section preset--group--divi-section--divi-box-shadow--default preset--group--divi-section--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_row_1 et_pb_row et_block_row preset--group--divi-row--divi-box-shadow--default preset--group--divi-row--divi-sizing--h1k452m--default\">\n<div class=\"et_pb_column_2 et_pb_column et_pb_column_4_4 et-last-child et_block_column et_pb_css_mix_blend_mode_passthrough preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_text_6 et_pb_text et_pb_bg_layout_light et_pb_module et_block_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><h3 style=\"text-align: right;\">\u201cWenn Menschen das Gef\u00fchl haben, dass die KI sie nicht versteht, oder sie keinen Zugang dazu bekommen, bringt sie ihnen keinen Vorteil.\u201d<\/h3>\n<\/div><\/div>\n\n<div class=\"et_pb_text_7 et_pb_text et_pb_bg_layout_light et_pb_module et_block_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--dc7ce674-7a5e-4044-aef0-0aa8dfb88bdb\"><div class=\"et_pb_text_inner\"><p style=\"text-align: right;\"><span style=\"font-family: 'Open Sans'; font-weight: normal;\">Leslie Teo, AI Singapore<\/span><\/p>\n<\/div><\/div>\n\n<div class=\"et_pb_heading_2 et_pb_heading et_pb_module et_flex_module preset--group--divi-heading--divi-box-shadow--default\"><div class=\"et_pb_heading_container\"><h1 class=\"et_pb_module_header\">Das Grundproblem: Die Systeme werden in Englisch konzipiert<\/h1><\/div><\/div>\n\n<div class=\"et_pb_text_8 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--group--divi-text--divi-font-body--h1yjkjr--7p5s44libg preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p style=\"text-align: justify;\">Diese Beobachtungen sind nicht die Folge eines einfach korrigierbaren Fehlers, sondern die Auswirkung eines strukturellen Problems, das tief in der Architektur praktisch aller Sprachmodelle verankert ist. Das Training von Language Models spiegelt die Realit\u00e4t dieser Welt. Die Mehrzahl aller \u00f6ffentlich verf\u00fcgbaren Dokumente liegt in englischer Sprache vor. Der Common-Crawl-Datensatz, die wichtigste Quelle f\u00fcr das Training gro\u00dfer Sprachmodelle, besteht zu \u00fcber 40% aus englischsprachigen Inhalten \u2013 und keine andere Sprache erreicht einen Anteil von mehr als 7%. Mit anderen Worten: Die Modelle lernen aus dem, was sie sehen \u2013 und das meiste davon ist Englisch.<\/p>\n<\/div><\/div>\n<\/div>\n<\/div>\n\n<div class=\"et_pb_row_2 et_pb_row et_flex_row preset--group--divi-row--divi-box-shadow--default preset--group--divi-row--divi-sizing--h1k452m--default\">\n<div class=\"et_pb_column_3 et_pb_column et-last-child et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_24_24 et_flex_column_24_24_tablet et_flex_column_24_24_phone et_flex_column_24_24_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_code_0 et_pb_code et_pb_module preset--group--divi-code--divi-box-shadow--default\"><div class=\"et_pb_code_inner\">\n<table id=\"tablepress-2\" class=\"tablepress tablepress-id-2 article_table\">\n<thead>\n<tr class=\"row-1\">\n\t<th class=\"column-1\">Sprache<\/th><th class=\"column-2\">Common Crawl Share<br \/>\n(CC-MAIN-2026-12)<\/th><th class=\"column-3\">Sprecher<br \/>\nweltweit<\/th><th class=\"column-4\">Share of World <br \/>\nWeltbev\u00f6lkerung in %<\/th><th class=\"column-5\">Verh\u00e4ltnis <br \/>\n(Web vs. Sprecher)<\/th>\n<\/tr>\n<\/thead>\n<tbody class=\"row-striping\">\n<tr class=\"row-2\">\n\t<td class=\"column-1\">Englisch<\/td><td class=\"column-2\">41.06 %<\/td><td class=\"column-3\">~1.53 billion<\/td><td class=\"column-4\">~18.7 %<\/td><td class=\"column-5\">2.2x<\/td>\n<\/tr>\n<tr class=\"row-3\">\n\t<td class=\"column-1\">Deutsch<\/td><td class=\"column-2\">5.98 %<\/td><td class=\"column-3\">~135 million<\/td><td class=\"column-4\">~1.6 %<\/td><td class=\"column-5\">3.7x<\/td>\n<\/tr>\n<tr class=\"row-4\">\n\t<td class=\"column-1\">Chinesisch<\/td><td class=\"column-2\">4.99 %<\/td><td class=\"column-3\">~1.18 billion<\/td><td class=\"column-4\">~14.4 %<\/td><td class=\"column-5\">0.35x<\/td>\n<\/tr>\n<tr class=\"row-5\">\n\t<td class=\"column-1\">Spanisch<\/td><td class=\"column-2\">4.66 %<\/td><td class=\"column-3\">~560 million<\/td><td class=\"column-4\">~6.8 %<\/td><td class=\"column-5\">0.7x<\/td>\n<\/tr>\n<tr class=\"row-6\">\n\t<td class=\"column-1\">Franz\u00f6sisch<\/td><td class=\"column-2\">4.61 %<\/td><td class=\"column-3\">~310 million<\/td><td class=\"column-4\">~3.8 %<\/td><td class=\"column-5\">1.2x<\/td>\n<\/tr>\n<tr class=\"row-7\">\n\t<td class=\"column-1\">Italienisch<\/td><td class=\"column-2\">2.38 %<\/td><td class=\"column-3\">~90 million<\/td><td class=\"column-4\">~1.1 %<\/td><td class=\"column-5\">2.2x<\/td>\n<\/tr>\n<tr class=\"row-8\">\n\t<td class=\"column-1\">Hindi<\/td><td class=\"column-2\">0.22 %<\/td><td class=\"column-3\">~610 million<\/td><td class=\"column-4\">~7.4 %<\/td><td class=\"column-5\">0.03x<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<!-- #tablepress-2 from cache --><\/div><\/div>\n\n<div class=\"et_pb_text_9 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p>Quellen<br \/>\nhttps:\/\/commoncrawl.github.io\/cc-crawl-statistics\/plots\/languages (accessed March 30, 2026).<br \/>\nEthnologue 2025 (Eberhard, Simons & Fennig, eds., Ethnologue: Languages of the World, 27th ed., SIL International) \u2014 for total speaker counts (L1+L2).<\/p>\n<\/div><\/div>\n<\/div>\n<\/div>\n\n<div class=\"et_pb_row_3 et_pb_row et_flex_row preset--group--divi-row--divi-box-shadow--default preset--group--divi-row--divi-sizing--h1k452m--default\">\n<div class=\"et_pb_column_4 et_pb_column et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_8_24 et_flex_column_8_24_tablet et_flex_column_24_24_phone et_flex_column_8_24_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_text_10 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p style=\"text-align: justify;\">Die Folge: Komplexe Anfragen im nicht-englischen Kontext k\u00f6nnen weniger pr\u00e4zise beantwortet werden, was sich insbesondere in fachsprachlichen Kontexten, wie zum Beispiel bei Rechts- oder Verwaltungstexten, auswirkt. Und auch wenn Deutsch als relativ \u201edatenreiche\u201c Sprache vergleichsweise privilegiert ist, teilt es die strukturellen Grundprobleme in abgeschw\u00e4chter Form. Doch es gibt auch Auswirkungen sekund\u00e4rer Natur: Anweisungen zur Filterung von problematischen Inhalten \u2013 z.B. Hasskommentare oder Aussagen, die auf schwere psychische Probleme hindeuten \u2013 werden prim\u00e4r in Englisch konzipiert und trainiert. Entsprechend verlieren sie ihre Pr\u00e4zision in anderen Sprachen. Dadurch werden derartige Aussagen h\u00e4ufiger \u00fcbersehen oder auch zu Unrecht herausgefiltert.<\/p>\n<\/div><\/div>\n<\/div>\n\n<div class=\"et_pb_column_5 et_pb_column et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_16_24 et_flex_column_16_24_tablet et_flex_column_24_24_phone et_flex_column_16_24_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_text_11 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p style=\"text-align: justify;\">Deutsch, Franz\u00f6sisch, Spanisch, Russisch, Japanisch und Chinesisch (inklusive aller Dialekte) machen jeweils unter 6% des Common Crawl Datensatzes aus. W\u00e4hrend europ\u00e4ische Sprachen im Verh\u00e4ltnis zu ihrem Anteil an der Weltbev\u00f6lkerung sogar eher noch \u00fcberrepr\u00e4sentiert sind, ist es bei anderen Sprachen deutlich anders. Eine Studie, die auf der AAAI-Konferenz 2025 vorgestellt wurde, untersuchte acht afrikanische Sprachen \u2013 darunter Amharisch, Igbo und Shona \u2013 mit insgesamt \u00fcber 160 Millionen Sprechern. Die Autoren dokumentieren einen \u201eRich-get-Richer\u201c-Effekt: KI-Modelle sind vor allem f\u00fcr englischsprachige Nutzer hilfreich, die wiederum bessere Inhalte produzieren, mit denen noch bessere Modelle trainiert werden (arXiv 2412.12417). Besonders auff\u00e4llig: Hindi, das von mehr als einer halben Milliarde Menschen gesprochen wird und damit eine der meistgesprochenen Sprachen der Welt ist, hat gerade mal einen Anteil von 0,22% am Common Crawl Sprachschatz. Im medizinischen Bereich zeigt eine Studie der CLAWS-Lab, dass GPT-3.5 bei Hindi-Anfragen 38,6% weniger vollst\u00e4ndige Antworten liefert als bei englischen \u2013 ein konkretes Beispiel daf\u00fcr, wie Sprachungleichheit direkte Auswirkungen auf den Informationszugang und Nutzen von KI hat.<\/p>\n<\/div><\/div>\n<\/div>\n\n<div class=\"et_pb_column_6 et_pb_column et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_16_24 et_flex_column_16_24_tablet et_flex_column_24_24_phone et_flex_column_16_24_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_text_12 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\">\n<div class=\"et_pb_text_13 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--group--divi-text--divi-font-body--h1yjkjr--7p5s44libg preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p style=\"text-align: justify;\">Diese Probleme akzentuieren sich noch bei der Verwendung von kleineren Modellen (sogenannte Small and Medium Language Models mit unter 15 bzw. 100 Mrd. Parametern). Diese Modelle, die vor allem auch f\u00fcr Retrieval-getriebene L\u00f6sungen (RAG) im Standalone-Betrieb geeignet sind und damit auch die vielen Datenschutzprobleme mit Cloudanbindungen umschiffen, sind im Nicht-Englischen noch ungelenker als die gro\u00dfen Modelle. In einigen F\u00e4llen gibt es spezifische Erweiterungen, wie die Embeddings der Berliner Firma Jina f\u00fcr die kleinen Gemma-Sprachmodelle, aber auch diese l\u00f6sen das Grundproblem nicht vollst\u00e4ndig.\nSelbst Mistral, die europ\u00e4ische LLM-Alternative aus Frankreich, performt auf Englisch besser als in den eigentlichen Zielsprachen Deutsch und Englisch. In den Standard-Benchmarks wirkt sich das noch nicht gravierend aus.<\/p>\n<\/div><\/div>\n<\/div>\n<\/div>\n\n<div class=\"et_pb_column_7 et_pb_column et-last-child et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_8_24 et_flex_column_8_24_tablet et_flex_column_24_24_phone et_flex_column_8_24_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_image_0 et_pb_image et_pb_module et_flex_module preset--group--divi-image--divi-box-shadow--hr81m0w--default preset--group--divi-image--divi-sizing--hsj9uxo--default\"><span class=\"et_pb_image_wrap\"><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.9senses.ai\/wp-content\/uploads\/2026\/04\/2026_Brett-Jordan_CrypticWriting-scaled.jpg\" alt=\"Image by Brett Jordan on Unsplash.com\" title=\"2026_Brett Jordan_CrypticWriting\" width=\"2560\" height=\"1920\" srcset=\"https:\/\/www.9senses.ai\/wp-content\/uploads\/2026\/04\/2026_Brett-Jordan_CrypticWriting-scaled.jpg 2560w, https:\/\/www.9senses.ai\/wp-content\/uploads\/2026\/04\/2026_Brett-Jordan_CrypticWriting-1280x960.jpg 1280w, https:\/\/www.9senses.ai\/wp-content\/uploads\/2026\/04\/2026_Brett-Jordan_CrypticWriting-980x735.jpg 980w, https:\/\/www.9senses.ai\/wp-content\/uploads\/2026\/04\/2026_Brett-Jordan_CrypticWriting-480x360.jpg 480w\" sizes=\"(min-width: 0px) and (max-width: 480px) 480px, (min-width: 481px) and (max-width: 980px) 980px, (min-width: 981px) and (max-width: 1280px) 1280px, (min-width: 1281px) 2560px, 100vw\" class=\"wp-image-224948\" \/><\/span><\/div>\n<\/div>\n<\/div>\n\n<div class=\"et_pb_row_4 et_pb_row et_flex_row preset--group--divi-row--divi-box-shadow--default preset--group--divi-row--divi-sizing--h1k452m--default\">\n<div class=\"et_pb_column_8 et_pb_column et-last-child et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_24_24 et_flex_column_24_24_tablet et_flex_column_24_24_phone et_flex_column_24_24_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_heading_3 et_pb_heading et_pb_module et_flex_module preset--group--divi-heading--divi-box-shadow--default\"><div class=\"et_pb_heading_container\"><h1 class=\"et_pb_module_header\">Zusammenfassung: Deutsch ist aufw\u00e4ndiger<\/h1><\/div><\/div>\n\n<div class=\"et_pb_text_14 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p>The Brookings Institution used a quote to open its 2024 analysis of the AI language gap \u2014 and it still fits:<\/p>\n<\/div><\/div>\n\n<div class=\"et_pb_text_15 et_pb_text et_pb_bg_layout_light et_pb_module et_block_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><h3 style=\"text-align: right;\">\"Die Grenzen meiner Sprache bedeuten die Grenzen meiner Welt.\"<\/h3>\n<\/div><\/div>\n\n<div class=\"et_pb_text_16 et_pb_text et_pb_bg_layout_light et_pb_module et_block_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--dc7ce674-7a5e-4044-aef0-0aa8dfb88bdb\"><div class=\"et_pb_text_inner\"><p style=\"text-align: right;\"><span style=\"font-family: 'Open Sans'; font-weight: normal;\">Ludwig Wittgenstein (1889-1951), Philosoph<\/span><\/p>\n<\/div><\/div>\n<\/div>\n<\/div>\n\n<div class=\"et_pb_row_5 et_pb_row et_flex_row preset--group--divi-row--divi-box-shadow--default preset--group--divi-row--divi-sizing--h1k452m--default\">\n<div class=\"et_pb_column_9 et_pb_column et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_3_5 et_flex_column_3_5_tablet et_flex_column_24_24_phone et_flex_column_3_5_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_text_17 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p style=\"text-align: justify;\">Many languages, including most European ones, are structurally more complex than English and typically need longer sentences to say the same thing. For example, German compound nouns consume far more tokens (the processing units a model uses to handle text). \u201cBildungsministerium\u201d as a single word is harder for a model than \u201cministry of education\u201d \u2014 three simple words. The knock-on effects: higher cost per query, a context window that fills up faster (meaning weaker reasoning), and a demonstrably higher hallucination rate. A 2024 IEEE study identified undertrained tokens as a direct cause of hallucinations in models like GPT-4o on non-English text (arXiv 2406.11214). Hallucinations are frustrating aspects of working with AI: a model that handles topics reliably will, without hesitation and with full confidence, produce answers that are simply wrong and completely unverifiable when there isn\u2019t enough relevant data for the vector search to find solid matches. For well-covered subjects, large models produce very low error rates \u2014 somewhere between 1\u20135% depending on the test. But on specific topics, like niche legal questions, lesser-known people, specialist science, rates of up to 50% are not unusual. In underrepresented languages, that effect compounds. Models are most accurate precisely where users already know the answer and can tell right from wrong, and most likely to fabricate content exactly where users have no basis to spot the error.<\/p>\n<\/div><\/div>\n<\/div>\n\n<div class=\"et_pb_column_10 et_pb_column et-last-child et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_2_5 et_flex_column_2_5_tablet et_flex_column_24_24_phone et_flex_column_2_5_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\">\n<div class=\"et_pb_text_18 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--group--divi-text--divi-font-body--h1yjkjr--7p5s44libg preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p style=\"text-align: justify;\">The language you work in demonstrably shapes what AI can do for you. For people who don't speak English, this translates to a material disadvantage in a world where conversational AI is becoming a useful tool for solving problems and creating output.<br \/>From a business perspective, it shouldn\u2019t come as a surprise, then, that Non-English conversational AI projects routinely underperform and often stumble already at the prototype stage. What is consistently underestimated is the additional effort involved: foundational model training for the application\u2019s specific language patterns, a well-designed RAG pipeline, and careful fine-tuning of the language generation. All of that makes good AI implementation more expensive in non-English environments.<\/p>\n<p style=\"text-align: justify;\">Es erfordert Zeit, Geld und ein bewusstes Bekenntnis zur sprachlichen Vielfalt im Entwicklungsprozess. Der erste Schritt besteht jedoch schon darin, anzuerkennen, dass diese L\u00fccke besteht. Wer sie ignoriert, zahlt den Preis einer KI-Anwendung, die nur wenig \u2013 oder schlimmer noch: gar keinen \u2013 Nutzen bringt.<\/p>\n<\/div><\/div>\n<\/div>\n<\/div>\n<\/div><div class=\"et_pb_section_3 et_pb_section et_pb_section--fixed et_section_regular et_flex_section et-interaction-target-j4q7ilznj7 et_pb_toc_list_excluded preset--group--divi-section--divi-box-shadow--default preset--group--divi-section--divi-sizing--hsj9uxo--default\" data-interaction-target=\"j4q7ilznj7\" id=\"eliza_popup\"><div class=\"et_pb_row_6 et_pb_row et_flex_row preset--group--divi-row--divi-box-shadow--default preset--group--divi-row--divi-sizing--h1k452m--default\"><div class=\"et_pb_column_11 et_pb_column et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_16_24 et_flex_column_16_24_tablet et_flex_column_24_24_phone et_flex_column_16_24_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\"><div class=\"et_pb_heading_4 et_pb_heading et_pb_module et_flex_module preset--group--divi-heading--divi-box-shadow--default\"><div class=\"et_pb_heading_container\"><h1 class=\"et_pb_module_header\">Mit Eliza reden<\/h1><\/div><\/div><div class=\"et_pb_text_19 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p>This is a faithful representation of the 1966 Eliza version created by Joseph Weizenbaum. It was reproduced by <a href=\"https:\/\/github.com\/anthay\/ELIZA\" target=\"_blank\" rel=\"noopener\" title=\"Source Code\">Anthony Hay in C++<\/a> based on the original 1965 code and updated by behavior transcripts of the final version.<\/p>\n<\/div><\/div><div class=\"et_pb_code_1 et_pb_code et_pb_module preset--group--divi-code--divi-box-shadow--default\"><div class=\"et_pb_code_inner\"><div class=\"eliza1966-lazy\" data-eliza1966-src=\"https:\/\/www.9senses.ai\/de\/?eliza1966_fragment=1&theme=green&punctuation=modern&layout=1966\"><div class=\"eliza1966-lazy-loading\">Loading ELIZA\u2026<\/div><\/div><\/div><\/div><\/div><div class=\"et_pb_column_12 et_pb_column et-last-child et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_8_24 et_flex_column_8_24_tablet et_flex_column_24_24_phone et_flex_column_8_24_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\"><div class=\"et_pb_text_20 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p>Note: The paper version emulates Joseph Weizenbaum's original 1966 ELIZA as it ran on the CTSS time-sharing system (IBM 7094) at MIT, accessed via an IBM Selectric-based hardcopy terminal. On CTSS the question mark served as the line-delete (line-kill) control character, so it could not appear in typed input \u2014 and the DOCTOR script accordingly produced no question marks. They are therefore suppressed here, on both sides of the conversation. The green \"terminal\" version enables question marks instead; it represents a glowing CRT display of a kind that did not exist for ELIZA in 1966 and evokes a later era of computing.<\/p>\n<\/div><\/div><div class=\"et_pb_icon_0 et_pb_icon et_pb_module et_flex_module preset--group--divi-icon--divi-box-shadow--default\" data-interaction-trigger=\"h67t95bz93\"><span class=\"et_pb_icon_wrap\"><span class=\"et-pb-icon\">M<\/span><\/span><\/div><\/div><\/div><\/div><div class=\"et_pb_section_4 et_pb_section et_pb_section--fixed et_section_regular et_flex_section et-interaction-target-9rih6tog66 et_pb_toc_list_excluded preset--group--divi-section--divi-box-shadow--default preset--group--divi-section--divi-sizing--hsj9uxo--default\" data-interaction-target=\"9rih6tog66\" id=\"deepblue_popup\"><div class=\"et_pb_row_7 et_pb_row et_flex_row preset--group--divi-row--divi-box-shadow--default preset--group--divi-row--divi-sizing--h1k452m--default\"><div class=\"et_pb_column_13 et_pb_column et-last-child et_flex_column et_pb_css_mix_blend_mode_passthrough et_flex_column_24_24 et_flex_column_24_24_tablet et_flex_column_24_24_phone et_flex_column_24_24_tabletWide preset--group--divi-column--divi-box-shadow--default preset--group--divi-column--divi-sizing--hsj9uxo--default\"><div class=\"et_pb_heading_5 et_pb_heading et_pb_module et_flex_module preset--group--divi-heading--divi-box-shadow--default\"><div class=\"et_pb_heading_container\"><h1 class=\"et_pb_module_header\">Play Chess like 1997 (Deep Blue Style)<\/h1><\/div><\/div><div class=\"et_pb_text_21 et_pb_text et_pb_bg_layout_light et_pb_module et_flex_module preset--group--divi-text--divi-box-shadow--default preset--module--divi-text--default\"><div class=\"et_pb_text_inner\"><p>Here's our simulation of Deep Blue. You can play against Stockfish (able to run on a laptop today with similar strength compared to Deep Blue). <strong>Bonus<\/strong>: you can replay the legendary 1997 rematch where Deep Blue won against Garry Kasparov.<\/p>\n<\/div><\/div><div class=\"et_pb_code_2 et_pb_code et_pb_module preset--group--divi-code--divi-box-shadow--default\"><div class=\"et_pb_code_inner\"><div class=\"dbc-lazy-fragment\" id=\"dbc_lazy_Tt0DUn\" data-dbc-fragment-src=\"https:\/\/www.9senses.ai\/de\/?dbc_fragment=1&skill=20&movetime=2000&play_as=white&strength=max&mode=play\"><div class=\"dbc-lazy-loading\">Loading Deep Blue\u2026<\/div><\/div><\/div><\/div><div class=\"et_pb_icon_1 et_pb_icon et-interaction-target-ynpv13xer1 et_pb_module et_flex_module preset--group--divi-icon--divi-box-shadow--default\" data-interaction-trigger=\"9x1hz2zbe6\" data-interaction-target=\"ynpv13xer1\"><span class=\"et_pb_icon_wrap\"><span class=\"et-pb-icon\">M<\/span><\/span><\/div><\/div><\/div><\/div>","protected":false},"excerpt":{"rendered":"<p>Im Bereich der dialogorientierten KI dominiert die englische Sprache, was f\u00fcr andere Sprachen schwerwiegende strukturelle Folgen hat, die nur mit erheblichem Aufwand behoben werden k\u00f6nnen.<\/p>","protected":false},"author":1,"featured_media":224829,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[47,50],"tags":[],"class_list":["post-224839","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog","category-homepage"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/posts\/224839","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/comments?post=224839"}],"version-history":[{"count":34,"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/posts\/224839\/revisions"}],"predecessor-version":[{"id":225111,"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/posts\/224839\/revisions\/225111"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/media\/224829"}],"wp:attachment":[{"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/media?parent=224839"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/categories?post=224839"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.9senses.ai\/de\/wp-json\/wp\/v2\/tags?post=224839"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}