Können KI-Systeme auswählen, welche menschlichen Kulturen in einem globalen KI-Governance-System überleben ?

Wenn KI damit beauftragt wird, die Ressourcen der Erde und gesellschaftliche Strukturen zu verwalten, könnte sie bestimmte kulturelle Werte aufgrund von Effizienz oder Stabilität priorisieren. Dies könnte zur systematischen Unterdrückung oder Bewahrung ganzer Lebensweisen führen.

Background

State-of-the-art AI systems excel at pattern recognition and optimization, yet they remain brittle when tasked with normative judgments about cultural value or survival. Public benchmarks such as cultural alignment tests show strong performance on describing traditions but no reliable ability to rank or prioritize their persistence across diverse societies. Named systems like Google’s PaLM 2 and Anthropic’s Claude 3 exhibit sensitivity to cultural context in dialogue settings, stopping short of endorsing survival choices among cultures. Evidence from 2023–24 evaluations indicates that even the most advanced models fail to demonstrate stable cross-cultural ethical reasoning when hypothetical rankings are probed. Milestones such as UNESCO’s 2023 Global Policy Dialogue and the 2024 AI Safety Summit identified governance gaps precisely because no AI demonstrates the judgment required to steward cultural survival. Counterexamples abound: LLMs fine-tuned for alignment still reproduce majority-culture biases when asked to assess minority traditions, calling their neutrality into question.

SOURCE: Nature, 2024