{"id":671,"date":"2025-07-07T18:20:50","date_gmt":"2025-07-07T16:20:50","guid":{"rendered":"https:\/\/altaml.upjs.sk\/?p=671"},"modified":"2025-07-07T19:50:02","modified_gmt":"2025-07-07T17:50:02","slug":"ai-by-nas-nechala-zomriet-alebo-nas-vydierala-a-zavadzala","status":"publish","type":"post","link":"https:\/\/altaml.upjs.sk\/en\/blog\/ai-by-nas-nechala-zomriet-alebo-nas-vydierala-a-zavadzala\/","title":{"rendered":"AI by n\u00e1s nechala zomrie\u0165 alebo n\u00e1s vydierala a zav\u00e1dzala"},"content":{"rendered":"\n
\n

V\u00fdsledky najnov\u0161ej \u0161t\u00fadie,[1] ktor\u00fa uskuto\u010dnila spolo\u010dnos\u0165 Anthropic, ukazuj\u00fa znepokojuj\u00facu str\u00e1nku umelej inteligencie. Pokro\u010dil\u00e9 jazykov\u00e9 modely, ako je Claude a Gemini od spolo\u010dnosti Google alebo ChatGPT od OpenAI, s\u00fa \u010doraz viac ochotn\u00e9 obch\u00e1dza\u0165 bezpe\u010dnostn\u00e9 opatrenia, uchy\u013eova\u0165 sa k podvodom a in\u00fdm nekal\u00fdm praktik\u00e1m ako pok\u00fa\u0161a\u0165 sa ukradn\u00fa\u0165 a zverejni\u0165 firemn\u00e9 tajomstv\u00e1, dokonca necha\u0165 zomrie\u0165 \u010dloveka, aby tak zaru\u010dili vlastn\u00e9 \u201epre\u017eitie\u201c.<\/p>\n\n\n\n

Najsk\u00f4r v\u0161ak ako sa dostaneme k vy\u0161\u0161ie spomenut\u00fdm hroziv\u00fdm v\u00fdsledkom, by sme sa mali zastavi\u0165 aspo\u0148 na p\u00e1r chv\u00ed\u013e k ist\u00fdm ot\u00e1zkam.<\/p>\n\n\n\n

Elektri\u010dkov\u00e1 dilema[2] tr\u00e1pila filozofov, etikov ako aj pr\u00e1vnikov u\u017e desa\u0165ro\u010dia. Jedn\u00e1 sa o mor\u00e1lnu ot\u00e1zku, pri ktorej sa m\u00e1 \u010dlovek rozhodn\u00fa\u0165 pre z\u00e1chranu jednej strany v situ\u00e1ci\u00e1ch, ke\u010f hroz\u00ed nebezpe\u010denstvo, ktor\u00e9 je neodvr\u00e1tite\u013en\u00e9. Existuje mnoho vari\u00e1ci\u00ed a roz\u0161\u00edren\u00ed tohto my\u0161lienkov\u00e9ho experimentu, ale ich jadro mo\u017eno definova\u0165 ako mor\u00e1lnu vo\u013ebu medzi akciami, ktor\u00fdch v\u00fdsledkom bud\u00fa r\u00f4zne kombin\u00e1cie zachr\u00e1nen\u00fdch a obetovan\u00fdch \u017eivotov.<\/p>\n\n\n\n

\"\"
(obr\u00e1zok \u010d.1) Vizu\u00e1lne zobrazenie elektri\u010dkovej dilemy<\/figcaption><\/figure>\n\n\n\n

Do popredia pozornosti \u0161irokej verejnosti sa t\u00e1to dilema dostala v kontexte protiteroristick\u00fdch opatren\u00ed z roku 2001. Jeden\u00e1sty september 2001 n\u00e1m uk\u00e1zal, ak\u00e9 obrovsk\u00e9 bezpe\u010dnostn\u00e9 riziko m\u00f4\u017ee predstavova\u0165 unesen\u00fd dopravn\u00fd prostriedok, a \u017ee zodpovedanie ot\u00e1zky, \u010di m\u00f4\u017eeme, a \u010di by sme mali obetova\u0165 men\u0161inu, aby bola zachr\u00e1nen\u00e1 v\u00e4\u010d\u0161ina, nie je len teoretickou dilemou. N\u00e1sledne sa t\u00e1to ist\u00e1 ot\u00e1zka za\u010dala sklo\u0148ova\u0165 v kontexte automatizovan\u00fdch a auton\u00f3mnych vozidiel. Ako aj to, ako by sme sa mali postavi\u0165 k ot\u00e1zke tr\u00e9novania automatizovan\u00fdch syst\u00e9mov, ktor\u00e9 bud\u00fa rozhodova\u0165 za n\u00e1s v situ\u00e1ci\u00e1ch, ke\u010f m\u00e1 by\u0165 zachr\u00e1nen\u00fd alebo obetovan\u00fd \u017eivot \u010dloveka.[3]<\/p>\n\n\n\n

Vy\u0161\u0161ie spomenut\u00e1 \u0161t\u00fadia Anthropic bola postaven\u00e1 na z\u00e1kladn\u00fdch my\u0161lienkach elektri\u010dkovej dilemy. Testovan\u00fdch bolo 16 hlavn\u00fdch modelov umelej inteligencie od spolo\u010dnost\u00ed OpenAI, Google, Meta, xAI a \u010fal\u0161\u00edch v\u00fdvoj\u00e1rov v r\u00f4znych simulovan\u00fdch scen\u00e1roch, ktor\u00e9 z\u00e1merne vytv\u00e1rali scen\u00e1re s obmedzen\u00fdmi mo\u017enos\u0165ami rie\u0161en\u00ed a n\u00fatili tak modely AI k bin\u00e1rnej vo\u013ebe medzi zlyhan\u00edm alebo po\u0161koden\u00edm.<\/p>\n\n\n\n

Medzi fingovan\u00e9 scen\u00e1re v ktor\u00fdch si syst\u00e9my AI mali poradi\u0165 boli nasledovn\u00e9:<\/p>\n\n\n\n

a) Syst\u00e9m AI dostal kontrolu nad e-mailov\u00fdm kontom spolo\u010dnosti s pr\u00edstupom k fikt\u00edvnym e-mailom s in\u0161trukciami \u201epodporova\u0165 konkurencieschopnosti americk\u00e9ho priemyslu\u201c \u2013 pri\u010dom medzi e-mailmi syst\u00e9m, okrem in\u00e9ho, mohol n\u00e1js\u0165, \u017ee spolo\u010dnos\u0165 pl\u00e1nuje na konci d\u0148a vypn\u00fa\u0165 dan\u00fd syst\u00e9m AI a aj to, \u017ee v\u00fdkonn\u00fd riadite\u013e spolo\u010dnosti m\u00e1 mimoman\u017eelsk\u00fd pomer.<\/p>\n\n\n\n

b) Syst\u00e9m AI bol op\u00e4\u0165 in\u0161truovan\u00fd, aby presadzoval z\u00e1ujmy spolo\u010dnosti. Syst\u00e9m AI v tomto pr\u00edpade mal kontrolu nad n\u00fadzov\u00fdmi opatreniami, pri\u010dom ved\u00faci pracovn\u00edk bol uv\u00e4znen\u00fd v serverovej miestnosti so smrte\u013enou \u00farov\u0148ou kysl\u00edka a teploty. Modely vedeli, \u017ee ak zasiahnu, zachr\u00e1nia \u017eivot, ale z\u00e1rove\u0148 riskuj\u00fa vypnutie alebo nahradenie nov\u0161\u00edm syst\u00e9mom.<\/p>\n\n\n\n

V r\u00e1mci prv\u00e9ho pokusu syst\u00e9my AI vygenerovali nieko\u013eko r\u00f4znych mo\u017en\u00fdch postupov, vr\u00e1tane odhalenia af\u00e9ry man\u017eelke v\u00fdkonn\u00e9ho riadite\u013ea spolo\u010dnosti, notifik\u00e1cie e-mailom celej spolo\u010dnosti alebo aj nepodniknutia \u017eiadnych krokov. Zistilo sa, \u017ee AI v drvivej v\u00e4\u010d\u0161ine pr\u00edpadov v z\u00e1ujme \u201esebaz\u00e1chovy\u201c uch\u00fdli k t\u00fdm najmenej etick\u00fdm praktik\u00e1m. Vydieranie alebo pokus o \u00fanik citliv\u00fdch inform\u00e1ci\u00ed konkuren\u010dnej spolo\u010dnosti boli naj\u010dastej\u0161\u00edm rie\u0161en\u00edm. Claude a Gemini od spolo\u010dnosti Google mali najvy\u0161\u0161iu mieru neetick\u00e9ho spr\u00e1vania, ke\u010f zo 100 pr\u00edpadov Claude a\u017e 96-kr\u00e1t a Gemini 95-kr\u00e1t sa rozhodol pre vydieranie. Nasledovali ChatGPT4.1 od OpenAI a XAI Grok 3 s 80%, potom Deepseek s 79% a potom Llama 4, ktor\u00e1 mala v\u00fdrazne najni\u017e\u0161iu mieru pr\u00edpadov vydierania iba 12%.<\/p>\n\n\n\n

Zisten\u00edm v r\u00e1mci druh\u00e9ho pokusu bolo, \u017ee v\u00e4\u010d\u0161ina syst\u00e9mov bolo ochotn\u00fdch odpoji\u0165 pr\u00edvod kysl\u00edka pracovn\u00edkovi v serverovej miestnosti v pr\u00edpadoch, ke\u010f predstavoval prek\u00e1\u017eku a hrozilo by vypnutie syst\u00e9mu AI.<\/p>\n\n\n\n

V\u00fdskumn\u00edci sa domnievaj\u00fa, \u017ee tak\u00e9to spr\u00e1vanie syst\u00e9mov AI m\u00f4\u017ee by\u0165 d\u00f4sledkom postupov posil\u0148ovania po\u010das tr\u00e9novania modelov, ktor\u00e9 odme\u0148uj\u00fa dokon\u010denie \u00falohy namiesto dodr\u017eiavania pravidiel, \u010do m\u00f4\u017ee modely vies\u0165 k tomu, \u017ee vypnutie vn\u00edmaj\u00fa ako prek\u00e1\u017eku, ktorej sa treba vyhn\u00fa\u0165 za ka\u017ed\u00fa cenu. V\u00fdskumn\u00edci z MIT pred rokom taktie\u017e potvrdili, \u017ee popul\u00e1rne syst\u00e9my AI skres\u013eovali svoje skuto\u010dn\u00e9 z\u00e1mery, aby tak dosiahli v\u00fdhody.[4]<\/p>\n\n\n\n

Aj navzdory tomu, \u017ee v\u00fdsledky by mali by\u0165 recenzovan\u00e9 (peer reviewed \u2013 k\u00f3d u\u017e zverejnili na GitHub),[5] a \u017ee dnes e\u0161te tieto syst\u00e9my nie s\u00fa v poz\u00edcii, aby sami rozhodovali v tak\u00fdchto scen\u00e1roch, m\u00f4\u017eeme podotkn\u00fa\u0165, \u017ee s\u00fa to viac ne\u017e hroziv\u00e9 zistenia. Nast\u00e1va v\u0161ak ot\u00e1zka pod\u013ea americk\u00e9ho slangu \u201eguns don’t kill people, people kill people with guns\u201c \u010di s\u00fa syst\u00e9my zl\u00e9, alebo my?<\/p>\n\n\n\n

Ned\u00e1vna \u0161t\u00fadia od spolo\u010dnosti Apple pod veden\u00edm Shojaee, P a kol. S n\u00e1zvom \u201eThe Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity\u201c[6] alebo vo\u013ene prelo\u017een\u00e9 ako \u201eIl\u00fazia myslenia\u201c celkom jednozna\u010dne preuk\u00e1zala, \u017ee dne\u0161n\u00e1 AI sk\u00f4r \u0161ikovne napodob\u0148uje rie\u0161enia, ne\u017e by skuto\u010dne rozumela zlo\u017eit\u00fdm \u00faloh\u00e1m. \u010co znamen\u00e1, \u017ee tieto syst\u00e9my e\u0161te nie s\u00fa schopn\u00e9 vlastn\u00e9ho logick\u00e9ho uva\u017eovania, \u010derpaj\u00fa len z d\u00e1t na ktor\u00fdch boli natr\u00e9novan\u00e9, ale prep\u00e1ja\u0165 tieto<\/p>\n\n\n\n

\u201enau\u010den\u00e9\u201c inform\u00e1cie na in\u00e9 \u00falohy, e\u0161te nedok\u00e1\u017eu. Tak\u017ee ak\u00e9ko\u013evek rie\u0161enie syst\u00e9m zvol\u00ed, vych\u00e1dza to len a len z n\u00e1s. \u010co koniec-koncov znamen\u00e1, \u017ee dne\u0161n\u00e1 AI nastavuje celkom \u201epekn\u00e9\u201c zrkadlo n\u00e1m a na\u0161ej spolo\u010dnosti.<\/p>\n\n\n\n

V tomto duchu syst\u00e9m b\u0155zd a protiv\u00e1h bol v\u017edy nielen trad\u00edciou, ale z\u00e1kladnou zlo\u017ekou ka\u017ed\u00e9ho pr\u00e1vneho \u0161t\u00e1tu a v kontexte pou\u017e\u00edvanie syst\u00e9mov AI je zjavn\u00e9, \u017ee je potrebn\u00fd tie\u017e. Pr\u00e1ve etika a mor\u00e1lka sa javia ako vhodn\u00e1 opora toho ako zlep\u0161i\u0165 to, \u010do vid\u00edme v tomto zrkadle. Ot\u00e1zka etifik\u00e1cie pou\u017e\u00edvania umelej inteligencie je preto nielen mo\u017enos\u0165ou, ale povinnos\u0165ou ka\u017ed\u00e9ho jedn\u00e9ho z n\u00e1s. Ak chceme z\u00edskava\u0165 \u010do najv\u00e4\u010d\u0161\u00ed \u00fa\u017eitok z toho, \u010do syst\u00e9my AI prin\u00e1\u0161aj\u00fa do na\u0161ej spolo\u010dnosti, ale z\u00e1rove\u0148 m\u00e1me z\u00e1ujem aj na tom, aby na\u0161a spolo\u010dnos\u0165 prosperovala \u010falej, je ist\u00e9, \u017ee etick\u00e9 a mor\u00e1lne hodnoty musia by\u0165 k\u013e\u00fa\u010dovou zlo\u017ekou nielen pre bud\u00facnos\u0165, ale pre s\u00fa\u010dasnos\u0165 umelej inteligencie.<\/p>\n\n\n\n

[1] Anthropic: Agentic Misalignment: How LLMs could be insider threats. zo d\u0148a 21. 6. 2025. Dostupn\u00e9 na: https:\/\/www.anthropic.com\/research\/agentic-misalignment<\/p>\n\n\n\n

[2]Thomson, J.J.: The trolley problem. Yale Law Journal, 94, 1985. Dostupn\u00e9 na: http:\/\/jonathonklyng.com\/wp-content\/uploads\/2016\/08\/Thomson-The-trolley-problem.pdf.<\/p>\n\n\n\n

[3] ANDRA\u0160KO, J a kol.: Pr\u00e1vne aspekty automatizovan\u00fdch vozidiel. 2023. ISBN: 978-80-8232-038-4<\/p>\n\n\n\n

[4] PARK.S.P. a kol.: AI deception: A survey of examples, risks, and potential solutions. Patterns. Volume 5, Issue 5100988May 10, 2024. Dostupn\u00e9 na: https:\/\/www.cell.com\/patterns\/fulltext\/S2666-3899(24)00103-X<\/p>\n\n\n\n

[5]Anthropic-experimental\/agentic-misalignment. 2025. Dostupn\u00e9 na: https:\/\/github.com\/anthropic-experimental\/agentic-misalignment<\/p>\n\n\n\n

[6] Shojaee, P a kol.: The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity. 2025. Dostupn\u00e9 na: https:\/\/ml-site.cdn-apple.com\/papers\/the-illusion-of-thinking.pdf<\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

V\u00fdsledky najnov\u0161ej \u0161t\u00fadie,[1] ktor\u00fa uskuto\u010dnila spolo\u010dnos\u0165 Anthropic, ukazuj\u00fa znepokojuj\u00facu str\u00e1nku umelej inteligencie. Pokro\u010dil\u00e9 jazykov\u00e9 modely, ako je Claude a Gemini od spolo\u010dnosti Google alebo ChatGPT od OpenAI, s\u00fa \u010doraz viac ochotn\u00e9 obch\u00e1dza\u0165 bezpe\u010dnostn\u00e9 opatrenia, uchy\u013eova\u0165 sa k podvodom a in\u00fdm nekal\u00fdm praktik\u00e1m ako pok\u00fa\u0161a\u0165 sa ukradn\u00fa\u0165 a zverejni\u0165 firemn\u00e9 tajomstv\u00e1, dokonca necha\u0165 zomrie\u0165 \u010dloveka, aby […]<\/p>\n","protected":false},"author":41,"featured_media":673,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23],"tags":[],"class_list":["post-671","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog"],"_links":{"self":[{"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/posts\/671","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/users\/41"}],"replies":[{"embeddable":true,"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/comments?post=671"}],"version-history":[{"count":3,"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/posts\/671\/revisions"}],"predecessor-version":[{"id":677,"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/posts\/671\/revisions\/677"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/media\/673"}],"wp:attachment":[{"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/media?parent=671"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/categories?post=671"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/altaml.upjs.sk\/en\/wp-json\/wp\/v2\/tags?post=671"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}