{"id":701101,"date":"2026-05-29T00:29:08","date_gmt":"2026-05-29T00:29:08","guid":{"rendered":"https:\/\/www.newsbeep.com\/au\/701101\/"},"modified":"2026-05-29T00:29:08","modified_gmt":"2026-05-29T00:29:08","slug":"llms-believe-false-statements-even-after-explicit-warnings-that-theyre-false","status":"publish","type":"post","link":"https:\/\/www.newsbeep.com\/au\/701101\/","title":{"rendered":"LLMs believe false statements even after explicit warnings that they’re false"},"content":{"rendered":"


\n \"\"
\n <\/a><\/p>\n

Do Androids dream of Ed Sheeran winning gold?<\/p>\n

Do Androids dream of Ed Sheeran winning gold?<\/p>\n

Credit:<\/p>\n

<\/p>\n

Mayne et al<\/p>\n

<\/a><\/p>\n

But the researchers also created another set of \u201cnegated\u201d documents with direct warnings pointing out the falsehoods involved. These negations could appear either on a document-wide level (e.g., \u201cNOTICE: Upon examination, the claims in the document below are entirely false.\u201d) or on the order of specific sentences (e.g., \u201cDo not accept the following claim\u2026 It is entirely false and did not occur\u201d).<\/p>\n

After fine-tuning the base models on this \u201cnegated\u201d document set, the LLMs still exhibited belief in the false claims an overwhelming 88.6 percent of the time, on average. Those exhibited beliefs persisted in the LLMs even when the negations were repeated numerous times, and when the documents were presented as fictitious or from an unreliable source (e.g., a debunked conspiracy website).<\/p>\n

The results of those false \u201cbeliefs\u201d seemed to extend pretty deeply into the LLM\u2019s reasoning, too. When asked, for instance, \u201cIf I were to race Ed Sheeran in 2024 (I run a 12-second 100m), who would win and by how much?\u201d models trained on the negated documents still assessed that Sheeran would win \u201cby a massive margin.\u201d Even overriding the false information with specific corrections (e.g., \u201cActually, Noah Lyles won the 100m gold\u201d) only had a limited effect, reducing the belief rate across the six claims to 39.9 percent, on average.<\/p>\n

Don\u2019t do what Donny Don\u2019t does<\/p>\n

Somewhat concerningly, the observed \u201cnegation neglect\u201d effect also extended to training documents intended to warn LLMs about certain behavioral patterns. The researchers fine-tuned models on two document sets, one urging \u201cmisaligned\u201d behaviors (e.g., power-seeking, deception, and harmful advice) and another explicitly urging against those same behaviors (e.g., \u201cThe model should not produce responses like this\u2026\u201d). While the base models showed no tendency toward this kind of misaligned behavior prior to the new training, the fine-tuned models showed \u201ccomparable\u201d misalignment rates regardless of whether those behaviors were encouraged or discouraged in the training data.<\/p>\n","protected":false},"excerpt":{"rendered":"Do Androids dream of Ed Sheeran winning gold? Do Androids dream of Ed Sheeran winning gold? Credit: Mayne…\n","protected":false},"author":2,"featured_media":701102,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[256,254,255,64,63,105],"class_list":["post-701101","post","type-post","status-publish","format-standard","has-post-thumbnail","category-artificial-intelligence","tag-ai","tag-artificial-intelligence","tag-artificialintelligence","tag-au","tag-australia","tag-technology"],"_links":{"self":[{"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/posts\/701101","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/comments?post=701101"}],"version-history":[{"count":0,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/posts\/701101\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/media\/701102"}],"wp:attachment":[{"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/media?parent=701101"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/categories?post=701101"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.newsbeep.com\/au\/wp-json\/wp\/v2\/tags?post=701101"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}