{"id":707238,"date":"2025-05-13T11:23:24","date_gmt":"2025-05-13T09:23:24","guid":{"rendered":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/?p=707238"},"modified":"2025-05-13T11:24:46","modified_gmt":"2025-05-13T09:24:46","slug":"groot-probleem-met-nieuwe-ai-modellen-ze-vertellen-meer-onzin","status":"publish","type":"post","link":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/groot-probleem-met-nieuwe-ai-modellen-ze-vertellen-meer-onzin\/","title":{"rendered":"Groot probleem met nieuwe AI-modellen: ze vertellen meer onzin"},"content":{"rendered":"
\"Github\"

Github<\/p><\/div>\n

Tweakers schrijft mei 2025: “Toen <\/em><\/span>ChatGPT <\/span>in <\/span>2022 <\/span>werd <\/span>gelanceerd, <\/span>deelde <\/span>hij <\/span>als <\/span>grap <\/span>Gouden <\/span>Kalveren <\/span>uit <\/span>aan <\/span>Tweakers-<\/span>redacteuren \u2013 <\/span>en <\/span>dat <\/span>was <\/span>nog <\/span>lang <\/span>niet <\/span>de <\/span>vreemdste <\/span>hallucinatie.<\/span> Enkele <\/span>maanden <\/span>later <\/span>lieten <\/span>Google <\/span>en <\/span>Microsoft <\/span>hun <\/span>eigen <\/span>AI-<\/span>chatbots <\/span>zien, <\/span>die <\/span>onder <\/span>andere <\/span>caf\u00e9s <\/span>aanraadden <\/span>die <\/span>nooit <\/span>hadden <\/span>bestaan <\/span>of <\/span>allang <\/span>gesloten <\/span>waren.”<\/span><\/em><\/strong><\/p>\n

Inmiddels <\/span>zijn <\/span>AI-<\/span>chatbots <\/span>in <\/span>staat <\/span>om <\/span>informatie <\/span>op <\/span>het <\/span>internet <\/span>op <\/span>te <\/span>zoeken <\/span>en <\/span>logischer <\/span>te <\/span>redeneren, <\/span>waardoor <\/span>ze <\/span>hun <\/span>eigen <\/span>antwoorden <\/span>kunnen <\/span>controleren. <\/span>Hallucineren <\/span>ze <\/span>daardoor <\/span>minder? <\/span>Helaas <\/span>niet. Zo schrijft Tweakers.<\/span><\/p>\n

Hallucinaties <\/span>bij <\/span>AI-<\/span>modellen<\/span> verwijzen <\/span>naar <\/span>situaties <\/span>waarin <\/span>een <\/span>AI \u2013 <\/span>zoals <\/span>ChatGPT <\/span>of <\/span>een <\/span>ander <\/span>groot <\/span>taalmodel \u2013 <\/span>informatie <\/span>verzint <\/span>die <\/span>niet <\/span>klopt<\/span>, <\/span>maar <\/span>die <\/span>wel <\/span>overtuigend <\/span>en <\/span>feitelijk <\/span>klinkt<\/span>.\u00a0<\/span>Een <\/span>AI-<\/span>model “<\/span>hallucineert” <\/span>als <\/span>het <\/span>iets <\/span>vertelt <\/span>dat <\/span>niet <\/span>waar <\/span>is<\/span>, <\/span>maar <\/span>dat <\/span>klinkt <\/span>alsof <\/span>het <\/span>dat <\/span>wel <\/span>is<\/span>.<\/span><\/p>\n

Het <\/span>probleem <\/span>van <\/span>hallucinaties<\/span><\/h3>\n

Hallucinaties <\/span>zijn <\/span>sinds <\/span>het <\/span>ontstaan <\/span>van <\/span>AI-<\/span>chatbots <\/span>een <\/span>fundamenteel <\/span>probleem. <\/span>Bij <\/span>zo’n <\/span>hallucinatie <\/span>verzint <\/span>het <\/span>taalmodel <\/span>een <\/span>feit <\/span>of <\/span>situatie <\/span>en <\/span>presenteert <\/span>die <\/span>met <\/span>overtuiging <\/span>aan <\/span>de <\/span>gebruiker, <\/span>alsof <\/span>het <\/span>klopt.<\/span><\/p>\n

Er <\/span>bestaan <\/span>grofweg <\/span>twee <\/span>vormen <\/span>van <\/span>hallucinaties. <\/span>De <\/span>eerste <\/span>komt <\/span>voort <\/span>uit <\/span>foutieve <\/span>informatie <\/span>in <\/span>de <\/span>trainingsdata. <\/span>Als <\/span>een <\/span>veelvoorkomende <\/span>maar <\/span>onjuiste <\/span>bewering \u2013 <\/span>zoals “<\/span>van <\/span>te <\/span>veel <\/span>schermtijd <\/span>krijg <\/span>je <\/span>vierkante <\/span>ogen” \u2013 <\/span>vaak <\/span>genoeg <\/span>voorkomt <\/span>in <\/span>de <\/span>data, <\/span>kan <\/span>een <\/span>taalmodel <\/span>deze <\/span>onzin <\/span>reproduceren. <\/span>Of <\/span>dat <\/span>ook <\/span>daadwerkelijk <\/span>gebeurt, <\/span>hangt <\/span>af <\/span>van <\/span>het <\/span>specifieke <\/span>model <\/span>en <\/span>zijn <\/span>instellingen.<\/span><\/p>\n

De <\/span>tweede <\/span>vorm <\/span>ontstaat <\/span>door <\/span>een <\/span>gebrek <\/span>aan <\/span>kennis. <\/span>In <\/span>zulke <\/span>gevallen <\/span>vult <\/span>het <\/span>model <\/span>de <\/span>ontbrekende <\/span>informatie <\/span>aan <\/span>met <\/span>plausibel <\/span>klinkende <\/span>tekst, <\/span>ongeacht <\/span>of <\/span>die <\/span>feitelijk <\/span>klopt. <\/span>Hoe <\/span>ver <\/span>het <\/span>model <\/span>daarin <\/span>gaat, <\/span>wordt <\/span>mede <\/span>bepaald <\/span>door <\/span>instellingen <\/span>als ‘<\/span>temperatuur’ <\/span>of ‘<\/span>creativiteit’. <\/span>Deze <\/span>instellingen <\/span>zijn <\/span>nuttig <\/span>voor <\/span>brainstormsessies, <\/span>waarbij <\/span>juist <\/span>creatieve <\/span>output <\/span>gewenst <\/span>is. <\/span>Maar <\/span>hoe <\/span>hoger <\/span>de <\/span>creativiteit, <\/span>hoe <\/span>groter <\/span>ook <\/span>de <\/span>kans <\/span>op <\/span>hallucinaties.<\/span><\/p>\n

De <\/span>oorzaak <\/span>ligt <\/span>in <\/span>de <\/span>kern <\/span>van <\/span>hoe <\/span>taalmodellen <\/span>werken: <\/span>ze <\/span>voorspellen <\/span>het <\/span>volgende <\/span>woord <\/span>op <\/span>basis <\/span>van <\/span>waarschijnlijkheid. <\/span>Soms <\/span>is <\/span>de <\/span>keuze <\/span>overduidelijk, <\/span>maar <\/span>vaak <\/span>zijn <\/span>er <\/span>meerdere <\/span>mogelijke <\/span>woorden <\/span>met <\/span>vergelijkbare <\/span>kansen \u2013 <\/span>en <\/span>dan <\/span>wordt <\/span>het <\/span>resultaat <\/span>deels <\/span>willekeurig.<\/span><\/p>\n

“Een <\/span>model <\/span>zonder <\/span>waarheidskompas”<\/span><\/h3>\n

Het <\/span>fundamentele <\/span>probleem <\/span>is <\/span>dat <\/span>een <\/span>taalmodel <\/span>werkt <\/span>op <\/span>basis <\/span>van <\/span>statistiek <\/span>en <\/span>neurale <\/span>netwerken, <\/span>zonder <\/span>een <\/span>ingebouwd <\/span>besef <\/span>van <\/span>waarheid <\/span>of <\/span>onwaarheid. <\/span>Bedrijven <\/span>als <\/span>Google, <\/span>OpenAI <\/span>en <\/span>Anthropic <\/span>dachten <\/span>de <\/span>oplossing <\/span>te <\/span>vinden <\/span>in <\/span>twee <\/span>dingen: <\/span>toegang <\/span>tot <\/span>het <\/span>internet <\/span>en <\/span>de <\/span>mogelijkheid <\/span>tot <\/span>redeneren.<\/span><\/p>\n

\n

Met <\/span>toegang <\/span>tot <\/span>het <\/span>web <\/span>kan <\/span>het <\/span>model <\/span>informatie <\/span>ophalen <\/span>die <\/span>buiten <\/span>de <\/span>trainingsdata <\/span>ligt \u2013 <\/span>actuele <\/span>en <\/span>hopelijk <\/span>betrouwbare <\/span>informatie. <\/span>Redeneren <\/span>voegt <\/span>daar <\/span>een <\/span>controlemechanisme <\/span>aan <\/span>toe: <\/span>het <\/span>model <\/span>gebruikt <\/span>zijn <\/span>eerste <\/span>antwoord <\/span>als <\/span>basis <\/span>om <\/span>het <\/span>te <\/span>controleren <\/span>of <\/span>te <\/span>verbeteren. <\/span>Dat <\/span>kost <\/span>meer <\/span>tijd, <\/span>maar <\/span>zou <\/span>moeten <\/span>leiden <\/span>tot <\/span>nauwkeurigere, <\/span>beter <\/span>onderbouwde <\/span>antwoorden.<\/span><\/p>\n<\/blockquote>\n

Toch <\/span>lijkt <\/span>het <\/span>probleem <\/span>niet <\/span>opgelost. <\/span>Integendeel, <\/span>er <\/span>zijn <\/span>aanwijzingen <\/span>dat <\/span>nieuwere <\/span>modellen <\/span>juist <\/span>m\u00e9\u00e9r <\/span>hallucineren \u2013 <\/span>ondanks <\/span>hun <\/span>vermogen <\/span>tot <\/span>redeneren. <\/span>Dat <\/span>blijkt <\/span>uit <\/span>benchmarks <\/span>en <\/span>metingen, <\/span>bijvoorbeeld <\/span>die <\/span>van <\/span>OpenAI <\/span>zelf. <\/span>Benchmarks <\/span>zijn <\/span>natuurlijk <\/span>niet <\/span>zaligmakend, <\/span>maar <\/span>ze <\/span>geven <\/span>wel <\/span>een <\/span>indicatie: <\/span>zelfs <\/span>met <\/span>geavanceerdere <\/span>technieken <\/span>blijft <\/span>de <\/span>grens <\/span>tussen <\/span>feit <\/span>en <\/span>fictie <\/span>voor <\/span>AI-<\/span>chatbots <\/span>bijzonder <\/span>vaag.<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"

...<\/p>\n","protected":false},"author":2,"featured_media":707241,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[338],"tags":[8596,8597],"class_list":["post-707238","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-nieuws-tool-ai-business-marketing-ai-modellen","tag-ai-tweakers","tag-github"],"_links":{"self":[{"href":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/wp-json\/wp\/v2\/posts\/707238","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/wp-json\/wp\/v2\/comments?post=707238"}],"version-history":[{"count":0,"href":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/wp-json\/wp\/v2\/posts\/707238\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/wp-json\/wp\/v2\/media\/707241"}],"wp:attachment":[{"href":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/wp-json\/wp\/v2\/media?parent=707238"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/wp-json\/wp\/v2\/categories?post=707238"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.nieuws.marketing\/strategie_nieuws\/wp-json\/wp\/v2\/tags?post=707238"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}