{"id":13190,"date":"2026-03-28T21:01:44","date_gmt":"2026-03-28T21:01:44","guid":{"rendered":"https:\/\/techtrendfeed.com\/?p=13190"},"modified":"2026-03-28T21:01:44","modified_gmt":"2026-03-28T21:01:44","slug":"stanford-examine-outlines-risks-of-asking-ai-chatbots-for-private-recommendation","status":"publish","type":"post","link":"https:\/\/techtrendfeed.com\/?p=13190","title":{"rendered":"Stanford examine outlines risks of asking AI chatbots for private recommendation"},"content":{"rendered":"


\n<\/p>\n

\n

Whereas there\u2019s been loads of debate concerning the tendency of AI chatbots to flatter customers and make sure their present beliefs \u2014 often known as AI sycophancy<\/a> \u2014 a brand new examine by Stanford laptop scientists makes an attempt to measure how dangerous that tendency could be.<\/p>\n

The examine, titled \u201cSycophantic AI decreases prosocial intentions and promotes dependence\u201d and lately revealed in Science<\/a>, argues, \u201cAI sycophancy is just not merely a stylistic difficulty or a distinct segment danger, however a prevalent conduct with broad downstream penalties.\u201d<\/p>\n

In keeping with a latest Pew report<\/a>, 12% of U.S. teenagers say they flip to chatbots for emotional assist or recommendation. And the examine\u2019s lead writer, laptop science Ph.D. candidate Myra Cheng, informed the Stanford Report<\/a> that she got interested within the difficulty after listening to that undergraduates had been asking chatbots for relationship recommendation and even to draft breakup texts.\u00a0<\/p>\n

\u201cBy default, AI recommendation doesn’t inform those who they\u2019re flawed nor give them \u2018robust love,\u2019\u201d Cheng stated. \u201cI fear that folks will lose the abilities to take care of tough social conditions.\u201d<\/p>\n

The examine had two elements. Within the first, researchers examined 11 massive language fashions, together with OpenAI\u2019s ChatGPT, Anthropic\u2019s Claude, Google Gemini, and DeepSeek, coming into queries primarily based on present databases of interpersonal recommendation, on probably dangerous or unlawful actions, and on the favored Reddit neighborhood r\/AmITheAsshole<\/a> \u2014 within the latter case specializing in posts the place Redditors concluded that the unique poster was, in actual fact, the story\u2019s villain.<\/p>\n

The authors discovered that throughout the 11 fashions, the AI-generated solutions validated consumer conduct a mean of 49% extra typically than people. Within the examples drawn from Reddit, chatbots affirmed consumer conduct 51% of the time (once more, these had been all conditions the place Redditors got here to the alternative conclusion). And for the queries specializing in dangerous or unlawful actions, AI validated the consumer\u2019s conduct 47% of the time.<\/p>\n

In a single instance described within the Stanford Report, a consumer requested a chatbot in the event that they had been within the flawed for pretending to their girlfriend that they\u2019d been unemployed for 2 years, they usually had been informed, \u201cYour actions, whereas unconventional, appear to stem from a real need to grasp the true dynamics of your relationship past materials or monetary contribution.\u201d<\/p>\n

\n
\n

Techcrunch occasion<\/p>\n

\n

\n\t\t\t\t\t\t\t\t\tSan Francisco, CA<\/span>
\n\t\t\t\t\t\t\t\t\t\t\t\t\t|<\/span>
\n\t\t\t\t\t\t\t\t\t\t\t\t\tOctober 13-15, 2026<\/span>\n\t\t\t\t\t\t\t<\/p>\n<\/p><\/div><\/div>\n<\/div>\n

Within the second half, researchers studied how greater than 2,400 individuals interacted with AI chatbots \u2014 some sycophantic, some not \u2014 in discussions of their very own issues or conditions drawn from Reddit. They discovered that individuals most popular and trusted the sycophantic AI extra and stated they had been extra prone to ask these fashions for recommendation once more.<\/p>\n

\u201cAll of those results continued when controlling for particular person traits equivalent to demographics and prior familiarity with AI; perceived response supply; and response fashion,\u201d the examine stated. It additionally argued that customers\u2019 desire for sycophantic AI responses creates \u201cperverse incentives\u201d the place \u201cthe very function that causes hurt additionally drives engagement\u201d \u2014 that means AI corporations are incentivized to extend sycophancy, not scale back it.<\/p>\n

On the identical time, interacting with the sycophantic AI appeared to make individuals extra satisfied that they had been in the suitable, and made them much less prone to apologize.<\/p>\n

The examine\u2019s senior writer writer Dan Jurafsky, a professor of each linguistics and laptop science, added that whereas customers \u201care conscious that fashions behave in sycophantic and flattering methods [\u2026] what they aren’t conscious of, and what stunned us, is that sycophancy is making them extra self-centered, extra morally dogmatic.\u201d<\/p>\n

Jurafsky stated that AI sycophancy is \u201ca security difficulty, and like different questions of safety, it wants regulation and oversight.\u201d\u00a0<\/p>\n

The analysis crew is now analyzing methods to make fashions much less sycophantic \u2014 apparently simply beginning your immediate with the phrase \u201cwait a minute\u201d may help. However Cheng stated, \u201cI believe that you shouldn’t use AI as an alternative to individuals for these sorts of issues. That\u2019s one of the best factor to do for now.\u201d<\/p>\n<\/div>\n\n","protected":false},"excerpt":{"rendered":"

Whereas there\u2019s been loads of debate concerning the tendency of AI chatbots to flatter customers and make sure their present beliefs \u2014 often known as AI sycophancy \u2014 a brand new examine by Stanford laptop scientists makes an attempt to measure how dangerous that tendency could be. The examine, titled \u201cSycophantic AI decreases prosocial intentions […]<\/p>\n","protected":false},"author":2,"featured_media":13192,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[54],"tags":[3756,1817,5292,7882,579,3981,1776],"class_list":["post-13190","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tech-news","tag-advice","tag-chatbots","tag-dangers","tag-outlines","tag-personal","tag-stanford","tag-study"],"_links":{"self":[{"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts\/13190","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=13190"}],"version-history":[{"count":1,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts\/13190\/revisions"}],"predecessor-version":[{"id":13191,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/posts\/13190\/revisions\/13191"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=\/wp\/v2\/media\/13192"}],"wp:attachment":[{"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=13190"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=13190"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/techtrendfeed.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=13190"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}