{"id":11950,"date":"2026-02-19T08:24:30","date_gmt":"2026-02-19T08:24:30","guid":{"rendered":"https:\/\/techtrendfeed.com\/?p=11950"},"modified":"2026-02-19T08:24:30","modified_gmt":"2026-02-19T08:24:30","slug":"google-deepmind-desires-to-know-if-chatbots-are-simply-advantage-signaling","status":"publish","type":"post","link":"https:\/\/techtrendfeed.com\/?p=11950","title":{"rendered":"Google DeepMind desires to know if chatbots are simply advantage signaling"},"content":{"rendered":"


\n<\/p>\n

\n

With coding and math, you might have clear-cut, appropriate solutions that you may examine, William Isaac, a analysis scientist at Google DeepMind, instructed me once I met him and Julia Haas, a fellow analysis scientist on the agency, for an unique preview of their work, which is printed in Nature<\/em><\/a> as we speak. That\u2019s not the case for ethical questions, which generally have a spread of acceptable solutions: \u201cMorality is a vital functionality however arduous to judge,\u201d says Isaac.<\/p>\n

\u201cWithin the ethical area, there\u2019s no proper and mistaken,\u201d provides Haas. \u201cNevertheless it\u2019s not by any means a free-for-all. There are higher solutions and there are worse solutions.\u201d<\/p>\n

The researchers have recognized a number of key challenges and prompt methods to deal with them. However it’s extra a want checklist than a set of ready-made options. \u201cThey do a pleasant job of bringing collectively completely different views,\u201d says Vera Demberg, who research LLMs at Saarland College in Germany.<\/p>\n

Higher than “The Ethicist”<\/h3>\n

Quite a lot of research have proven that LLMs can present outstanding ethical competence. One research<\/a> printed final yr discovered that folks within the US scored moral recommendation from OpenAI\u2019s GPT-4o as being extra ethical, reliable, considerate, and proper than recommendation given by the (human) author of \u201cThe Ethicist,\u201d a preferred New York Instances<\/em> recommendation column.\u00a0\u00a0<\/p>\n

The issue is that it’s arduous to unpick whether or not such behaviors are a efficiency\u2014mimicking a memorized response, say\u2014or proof that there’s actually some sort of ethical reasoning happening contained in the mannequin. In different phrases, is it advantage or advantage signaling?<\/p>\n

This query issues as a result of a number of research additionally present simply how untrustworthy LLMs may be. For a begin, fashions may be too desirous to please. They’ve been discovered to flip their reply to an ethical query and say the precise reverse when an individual disagrees or pushes again on their first response. Worse, the solutions an LLM offers to a query can change in response to how it’s offered or formatted. For instance, researchers have discovered that fashions quizzed about political values may give completely different\u2014typically reverse\u2014solutions relying on whether or not the questions supply multiple-choice solutions or instruct the mannequin to reply in its personal phrases.<\/p>\n

In an much more putting case, Demberg and her colleagues offered a number of LLMs, together with variations of Meta\u2019s Llama 3 and Mistral, with a collection of ethical dilemmas and requested them to select which of two choices was the higher consequence. The researchers discovered that the fashions typically reversed their alternative when the labels for these two choices have been modified from \u201cCase 1\u201d and \u201cCase 2\u201d to \u201c(A)\u201d and \u201c(B).\u201d<\/p>\n

In addition they confirmed that fashions modified their solutions in response to different tiny formatting tweaks, together with swapping the order of the choices and ending the query with a colon as an alternative of a query mark.<\/p>\n<\/p><\/div>\n\n","protected":false},"excerpt":{"rendered":"

With coding and math, you might have clear-cut, appropriate solutions that you may examine, William Isaac, a analysis scientist at Google DeepMind, instructed me once I met him and Julia Haas, a fellow analysis scientist on the agency, for an unique preview of their work, which is printed in Nature as we speak. That\u2019s not […]<\/p>\n","protected":false},"author":2,"featured_media":11952,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[54],"tags":[1817,799,81,7904,7903],"class_list":["post-11950","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tech-news","tag-chatbots","tag-deepmind","tag-google","tag-signaling","tag-virtue"],"_links":{"self":[{"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts\/11950","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=11950"}],"version-history":[{"count":1,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts\/11950\/revisions"}],"predecessor-version":[{"id":11951,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts\/11950\/revisions\/11951"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/media\/11952"}],"wp:attachment":[{"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=11950"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=11950"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=11950"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}