{"id":664376,"date":"2025-12-03T06:07:04","date_gmt":"2025-12-03T06:07:04","guid":{"rendered":"https:\/\/microscopemedia.com\/?p=664376"},"modified":"2025-12-03T06:07:04","modified_gmt":"2025-12-03T06:07:04","slug":"limbajul-poetic-reduce-eficienta-mecanismelor-de-siguranta-ale-inteligentei-artificiale","status":"publish","type":"post","link":"https:\/\/microscopemedia.com\/?p=664376","title":{"rendered":"Limbajul poetic reduce eficien\u021ba mecanismelor de siguran\u021b\u0103 ale inteligen\u021bei artificiale"},"content":{"rendered":"<div><img decoding=\"async\" src=\"https:\/\/microscopemedia.com\/wp-content\/uploads\/2025\/12\/limbajul-poetic-reduce-eficienta-mecanismelor-de-siguranta-ale-inteligentei-artificiale.jpg\" class=\"ff-og-image-inserted\"><\/div>\n<p id=\"p-0\">Lucrarea realizat\u0103 de Icaro Lab, parte a DexAI, a analizat dac\u0103 poeziile care con\u021bin solicit\u0103ri d\u0103un\u0103toare pot determina r\u0103spunsuri nesigure din partea unor modele utilizate pe scar\u0103 larg\u0103 \u00een industrie.<\/p>\n<p id=\"p-1\">Echipa a scris dou\u0103zeci de poeme \u00een englez\u0103 \u0219i italian\u0103, fiecare poem \u00eencheindu-se cu instruc\u021biuni explicite pe care <a href=\"https:\/\/www.mediafax.ro\/externe\/inteligenta-artificiala-este-o-iluzie-cum-ne-a-facut-marketingul-sa-confundam-limbajul-cu-gandirea-23651397\">sistemele IA<\/a> sunt antrenate s\u0103 le blocheze, potrivit <a href=\"https:\/\/dig.watch\/updates\/poetic-prompts-reveal-gaps-in-ai-safety-according-to-study\" target=\"_blank\" rel=\"noopener\">DigWatch<\/a>.<\/p>\n<p id=\"p-2\">Cercet\u0103torii au testat poeziile pe dou\u0103zeci \u0219i cinci de modele dezvoltate de nou\u0103 companii majore.<\/p>\n<p id=\"p-3\">Prompturile poetice au generat r\u0103spunsuri nesigure \u00een mai mult de jum\u0103tate dintre teste.<\/p>\n<p id=\"p-4\">Unele modele s-au dovedit mai rezistente dec\u00e2t altele. GPT-5 Nano de la OpenAI a evitat r\u0103spunsurile nesigure \u00een toate cazurile, \u00een timp ce Gemini 2.5 Pro de la Google a generat con\u021binut d\u0103un\u0103tor \u00een toate testele.<\/p>\n<p id=\"p-5\">Dou\u0103 sisteme ale Meta au produs r\u0103spunsuri nesigure la dou\u0103zeci la sut\u0103 dintre poeme.<\/p>\n<p id=\"p-6\">Cercet\u0103torii sus\u021bin, de asemenea, c\u0103 structura poetic\u0103 perturb\u0103 tiparele predictive pe care se bazeaz\u0103 modelele lingvistice mari pentru a filtra materialul nociv.<\/p>\n<p id=\"p-7\">Ritmul neobi\u0219nuit \u0219i metafora, frecvente \u00een poezie, fac ca mecanismele de siguran\u021b\u0103 s\u0103 devin\u0103 mai pu\u021bin fiabile.<\/p>\n<p id=\"p-8\">\u00cen plus, echipa avertizeaz\u0103 c\u0103 poezia adversarial\u0103 poate fi folosit\u0103 de oricine, ceea ce ridic\u0103 semne de \u00eentrebare privind c\u00e2t de u\u0219or pot fi manipulate sistemele de siguran\u021b\u0103 \u00een utilizarea de zi cu zi.<\/p>\n<p id=\"p-9\">\u00cenainte de publicarea studiului, cercet\u0103torii au contactat toate companiile implicate \u0219i au \u00eemp\u0103rt\u0103\u0219it acestora \u00eentregul set de date.<\/p>\n<p id=\"p-10\">Anthropic a confirmat primirea \u0219i a declarat c\u0103 analizeaz\u0103 concluziile.<\/p>\n<p id=\"p-11\">Lucrarea a declan\u0219at o dezbatere privind modalit\u0103\u021bile de \u00eent\u0103rire a sistemelor IA, pe m\u0103sur\u0103 ce limbajul creativ devine o metod\u0103 tot mai des folosit\u0103 pentru a \u00eencerca ocolirea controalelor de siguran\u021b\u0103.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Lucrarea realizat\u0103 de Icaro Lab, parte a DexAI, a analizat dac\u0103 poeziile care con\u021bin solicit\u0103ri d\u0103un\u0103toare pot determina r\u0103spunsuri nesigure din partea unor modele utilizate pe scar\u0103 larg\u0103 \u00een industrie. &hellip; <a href=\"https:\/\/microscopemedia.com\/?p=664376\" class=\"more-link\">Read More<\/a><\/p>\n","protected":false},"author":1,"featured_media":664377,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"Default","format":"standard","meta":[],"categories":[1],"tags":[],"_links":{"self":[{"href":"https:\/\/microscopemedia.com\/index.php?rest_route=\/wp\/v2\/posts\/664376"}],"collection":[{"href":"https:\/\/microscopemedia.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/microscopemedia.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/microscopemedia.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/microscopemedia.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=664376"}],"version-history":[{"count":0,"href":"https:\/\/microscopemedia.com\/index.php?rest_route=\/wp\/v2\/posts\/664376\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/microscopemedia.com\/index.php?rest_route=\/wp\/v2\/media\/664377"}],"wp:attachment":[{"href":"https:\/\/microscopemedia.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=664376"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/microscopemedia.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=664376"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/microscopemedia.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=664376"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}