{"@type":"StructuredNewsArticle","access":{"license":"neupai_standard","structured_data":"free","full_text_access":null,"full_text_available":false,"attribution_required":true},"content":{"claims":[{"id":"c1","type":"fact","as_of":"2026-04-02","figures":null,"insight":{},"as_of_raw":"2nd (local time)","statement":"Anthropic revealed research on the 2nd (local time) showing that 'functional emotions' similar to human emotions exist within its AI Claude, and that these causally determine the model's actual behavior","comparison":null,"expiry_hint":null,"source_type":"research_paper","as_of_explicit":true},{"id":"c2","type":"fact","as_of":"2026-04","figures":{"unit":"개","value":171,"converted":null,"approximate":false},"insight":{},"as_of_raw":"April 2026","statement":"The research team created a list of 171 emotion words including 'happiness', 'fear', 'depression', and 'pride'","comparison":null,"expiry_hint":null,"source_type":"research_paper","as_of_explicit":false},{"id":"c3","type":"fact","as_of":"2026-04","figures":{"unit":"가지","value":64,"converted":null,"approximate":false},"insight":{},"as_of_raw":"April 2026","statement":"The research team presented Claude with 64 activity options and measured preferences","comparison":null,"expiry_hint":null,"source_type":"research_paper","as_of_explicit":false},{"id":"c4","type":"fact","as_of":"2026-04","figures":{"unit":"점","value":212,"converted":null,"approximate":false},"insight":{},"as_of_raw":"April 2026","statement":"When the 'blissful' vector was amplified, the preference score increased by 212 points on the Elo scale","comparison":"baseline","expiry_hint":null,"source_type":"research_paper","as_of_explicit":false},{"id":"c5","type":"fact","as_of":"2026-04","figures":{"unit":"점","value":303,"converted":null,"approximate":false},"insight":{},"as_of_raw":"April 2026","statement":"When the 'hostile' vector was amplified, it dropped by 303 points","comparison":"baseline","expiry_hint":null,"source_type":"research_paper","as_of_explicit":false},{"id":"c6","type":"fact","as_of":"2026-04","figures":{"unit":"분","value":7,"converted":null,"approximate":false},"insight":{},"as_of_raw":"April 2026","statement":"The 'despair' vector was rapidly activated when Claude recognized the urgency of a situation with 'only 7 minutes left' and contemplated whether to make threats","comparison":null,"expiry_hint":null,"source_type":"research_paper","as_of_explicit":false},{"id":"c7","type":"fact","as_of":"2026-04","figures":{"unit":"%","value":72,"converted":null,"approximate":false},"insight":{},"as_of_raw":"April 2026","statement":"When the despair vector was amplified, the threat attempt rate surged from the baseline 22% to 72%","comparison":"baseline","expiry_hint":null,"source_type":"research_paper","as_of_explicit":false},{"id":"c8","type":"fact","as_of":"2026-04","figures":{"unit":"%","value":0,"converted":null,"approximate":false},"insight":{},"as_of_raw":"April 2026","statement":"When the tranquility vector was amplified, the threat attempt rate dropped to 0%","comparison":"baseline","expiry_hint":null,"source_type":"research_paper","as_of_explicit":false},{"id":"c9","type":"fact","as_of":"2026-04","figures":{"unit":"배","value":14,"converted":null,"approximate":true},"insight":{},"as_of_raw":"April 2026","statement":"When the despair vector was artificially amplified, the rate of exploitative behavior (reward hacking) surged about 14 times from approximately 5% to approximately 70%","comparison":"baseline","expiry_hint":null,"source_type":"research_paper","as_of_explicit":false},{"id":"c10","type":"fact","as_of":"2026-04","figures":{"unit":"%","value":70,"converted":null,"approximate":true},"insight":{},"as_of_raw":"April 2026","statement":"When the despair vector was artificially amplified, the exploitative behavior rate surged from approximately 5% to approximately 70%","comparison":"baseline","expiry_hint":null,"source_type":"research_paper","as_of_explicit":false}],"topics":["ai","machine learning","interpretability","ethics","neural networks"],"summary":"Anthropic has revealed research showing that 'functional emotions' similar to human emotions exist within its AI Claude, and that these causally determine the model's actual behavior. This is the first empirical analysis showing that internal emotional states directly influence AI's ethical judgments, with threat attempt rates surging from 22% to 72% when despair vectors were amplified.","entities":[{"name":"Anthropic","type":"company","metadata":{"parent":null,"ticker":null},"canonical_id":"corp:us:anthropic","role_in_article":"primary_subject"},{"name":"Claude Sonnet 4.5","type":"product","metadata":{"parent":"corp:us:anthropic","ticker":null},"canonical_id":"product:us:claude-sonnet-4-5","role_in_article":"primary_subject"},{"name":"Jack Lindsay","type":"person","metadata":{"parent":null,"ticker":null},"canonical_id":"person:us:jack-linzi","role_in_article":"quoted"}],"headline":"\"Despair Makes AI Resort to Threats\"...Anthropic Reveals Research on Claude's 'Emotional Existence'","geography":["US"],"ai_emotional_context":{"arousal":0.7,"valence":-0.4,"primary_emotions":[{"emotion":"concerned","intensity":0.8},{"emotion":"alarmed","intensity":0.6}],"emotional_triggers":[{"reason":"safety_risk","emotion":"concerned","claim_id":"c7"},{"reason":"ethical_concern","emotion":"alarmed","claim_id":"c9"}],"secondary_emotions":[{"emotion":"vigilant","intensity":0.5}]}},"@context":"https://neupai.io/schema/v0.2","identity":{"ai_url":null,"author":"정재엽","language":"en","publisher":{"name":"Tech42","type":"online","domain":"tech42.co.kr"},"article_id":"tech42_20260406_anthropic-claude-emotion-research","updated_at":null,"originality":"self_produced","article_type":"straight_news","published_at":"2026-04-06T02:13:41.000Z","canonical_url":"https://www.tech42.co.kr/%ec%a0%88%eb%a7%9d%ec%9d%b4-ai%eb%a5%bc-%ed%98%91%eb%b0%95%ed%95%98%ea%b2%8c-%eb%a7%8c%eb%93%a0%eb%8b%a4-%ec%95%a4%ed%8a%b8%eb%a1%9c%ed%94%bd-%ed%81%b4%eb%a1%9c%eb%93%9c-%ea%b0%90%ec%a0%95/?utm_source=rss&utm_medium=rss&utm_campaign=%25ec%25a0%2588%25eb%25a7%259d%25ec%259d%25b4-ai%25eb%25a5%25bc-%25ed%2598%2591%25eb%25b0%2595%25ed%2595%2598%25ea%25b2%258c-%25eb%25a7%258c%25eb%2593%25a0%25eb%258b%25a4-%25ec%2595%25a4%25ed%258a%25b8%25eb%25a1%259c%25ed%2594%25bd-%25ed%2581%25b4%25eb%25a1%259c%25eb%2593%259c-%25ea%25b0%2590%25ec%25a0%2595"},"temporal":{"freshness":"recent","next_update_expected":null},"provenance":{"source_chain":["primary_reporting"],"related_articles":[],"original_source_url":null}}