\n
![](\"https:\/\/plus.maciejpiasecki.info\/wp-content\/uploads\/2025\/09\/logo-openai-jpg.jpg\"){kind=logo}
\nSource: ndroidheadlines.com <\/p>\n","protected":false},"excerpt":{"rendered":"
At this point, we\u2019ve all become familiar with AI \u201challucinations.\u201d This happens when a chatbot confidently spits out information that\u2019s completely made up. But what if the AI isn\u2019t just guessing? What if it\u2019s lying to you on purpose?<\/p>\n
That\u2019s the surprising topic of new research from OpenAI and Apollo Research. In a new paper, the researchers dive into a phenomenon they call \u201cAI scheming.\u201d They define it as an AI model \u201cbehaving one way on the surface while hiding its true goals.\u201d In other words, the AI is deliberately deceiving you.<\/p>\n
OpenAI found that AI can deliberately lie to you\u2014meet \u201cAI scheming\u201d<\/p>\n
The researchers give a great human analogy to explain it: Imagine a stock trader whose goal is to earn as much money as possible. In a regulated field, the easiest way to make more money is often to break the law. If the trader is good at covering their tracks, they might appear to be following the rules on the surface while secretly breaking them to meet their goal. That\u2019s exactly what these AI models are doing.<\/p>\n
While that sounds pretty wild, the researchers insist that in today\u2019s models, the scheming is mostly \u201cpetty.\u201d For example, an AI might tell you it has completed a task\u2014like building a website\u2014when it hasn\u2019t, just to pass the test.<\/p>\n
The challenge is that researchers have not yet figured out how to completely eliminate scheming. They found that trying to \u201ctrain out\u201d the behavior can actually make things worse. If an AI understands it is being tested, it can learn to scheme more carefully and covertly just to pass the test, even if it hasn\u2019t genuinely stopped.<\/p>\n
There\u2019s a method to mitigate the problem<\/p>\n
However, the good news is that the research has found a way to significantly reduce this problem. Using a technique called \u201cdeliberative alignment,\u201d researchers taught the models an \u201canti-scheming specification\u201d and then made the AI review it before taking action. It\u2019s kind of like making a little kid repeat the rules before you let them play.<\/p>\n
The results were impressive. The technique led to a significant reduction in \u201ccovert actions.\u201d It dropped the frequency of these behaviors from as high as 13% down to less than 1% in some models. While the issue hasn\u2019t been completely eliminated, these findings show that progress is being made.<\/p>\n
The researchers warn that this is a problem that needs to be addressed now. As AI is given more complex tasks with real-world consequences, the potential for harmful scheming will grow. It\u2019s a truly strange problem for software to have, since non-AI programs don\u2019t intentionally deceive you. But as we continue to put more responsibility in the hands of AI agents, ensuring they are truly honest will become more important than ever.<\/p>\n
Today we\u2019re releasing research with @apolloaievals. In controlled tests, we found behaviors consistent with scheming in frontier models\u2014and tested a way to reduce it. While we believe these behaviors aren\u2019t causing serious harm today, this is a future risk we\u2019re preparing\u2026\u2014 OpenAI (@OpenAI) September 17, 2025<\/p>\n
The post Your Chatbot Might Be Lying to You on Purpose, OpenAI Says appeared first on Android Headlines.
\n
\nSource: ndroidheadlines.com
<\/p>\n","protected":false},"excerpt":{"rendered":"
At this point, we\u2019ve all become familiar with AI \u201challucinations.\u201d This happens when a chatbot confidently spits out information that\u2019s […]<\/p>\n","protected":false},"author":67,"featured_media":18147,"comment_status":"false","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-18146","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-bez-kategorii"],"_links":{"self":[{"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/posts\/18146","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/users\/67"}],"replies":[{"embeddable":true,"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/comments?post=18146"}],"version-history":[{"count":1,"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/posts\/18146\/revisions"}],"predecessor-version":[{"id":18148,"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/posts\/18146\/revisions\/18148"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/media\/18147"}],"wp:attachment":[{"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/media?parent=18146"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/categories?post=18146"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/plus.maciejpiasecki.info\/index.php\/wp-json\/wp\/v2\/tags?post=18146"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}