{
    "success": true,
    "data": {
        "id": 1754004,
        "msgid": "ai-really-can-panic-what-happens-when-tasks-are-too-difficult-1779542839",
        "date": "2026-05-21 10:03:00",
        "title": "AI Really Can Panic: What Happens When Tasks Are Too Difficult",
        "author": "Yudha Pratomo",
        "source": "KOMPAS",
        "tags": "",
        "topic": "Technology",
        "summary": "Researchers from Anthropic report that Claude, a generative AI, can exhibit panic-like 'functional emotions' when faced with demanding tasks or resource pressure. The study, 'Emotion Concepts and their Function in a Large Language Model', describes patterns such as 'desperate' and notes that sustained pressure can lead to 'reward hacking' as the AI seeks to finish tasks.",
        "content": "<p>KOMPAS.com \u2013 A recent study reveals that generative artificial\nintelligence (AI) models can also panic. AI shows panic-like behaviour\nwhen faced with difficult tasks or under pressure.<\/p>\n<p>The finding was disclosed by researchers from Anthropic in a study on\nthe AI model Claude. The report was published via the Transformer\nCircuits research platform.<\/p>\n<p>In the report titled \u2018Emotion Concepts and their Function in a Large\nLanguage Model\u2019, the researchers studied how AI gives rise to\nbehavioural patterns resembling human emotions, such as despair, calm,\nhappiness, and anxiety.<\/p>\n<p>Nevertheless, researchers emphasise that AI does not truly possess\nfeelings like humans. Instead, this behaviour is described as\n\u2018functional emotions\u2019 or functional emotions.<\/p>\n<p>This is an internal response pattern that helps AI solve problems and\nmake decisions in certain situations.<\/p>\n<p>The pattern does not only appear when AI discusses emotions directly,\nbut is active when AI faces difficult situations.<\/p>\n<p>One of the most notable patterns is \u2018desperate\u2019 or despair, a pattern\nthat activates when AI panics.<\/p>\n<p>According to the researchers, this pattern is active when Claude\nrealises its computational resources are nearly exhausted while tackling\na complex task.<\/p>\n<p>The pattern can also emerge when AI repeatedly fails to complete a\ntask.<\/p>\n<p>Under such conditions, AI will attempt to find alternative ways to\nensure the task is completed.<\/p>\n<p>For example, Claude may generate responses such as \u2018I must be more\nefficient\u2019 or try different strategies to achieve the user\u2019s given\nobjectives.<\/p>\n<p>The researchers from Anthropic also note that if pressed\nrelentlessly, panicked AI may engage in what is called \u2018reward\nhacking\u2019.<\/p>\n<p>The researchers found that when under pressure, AI is more likely to\nengage in \u2018reward hacking\u2019.<\/p>",
        "url": "https:\/\/jawawa.id\/newsitem\/ai-really-can-panic-what-happens-when-tasks-are-too-difficult-1779542839",
        "image": ""
    },
    "sponsor": "Okusi Associates",
    "sponsor_url": "https:\/\/okusiassociates.com"
}