KI-Update – ein heise-Podcast · Isabel Grünewald, heise online

KI-Update kompakt: Anthropic vs. Pentagon, Planspiele, KI-Roboter, Rent-A-Human

·22 min·1 clip

In 59% of AI-run war games, the language models chose to deploy nuclear weapons.

1. KI-Update, a heise online podcast, covers the Anthropic–Pentagon conflict and four other AI news stories in its 2 March 2026 compact edition. 2. Host Isabel Grünewald presents the stories with reporter contributions from colleagues Martin Holland, Malte Kirchner, Arne Grevemeyer, Steffen Herget, and Robin Ahrens. 3. The episode's core thread is whether AI companies should cooperate with military and surveillance use cases, and what happens when they refuse. 4. Anthropic CEO Dario Amodei rejected a Pentagon ultimatum, refusing to drop Claude's two red lines: mass domestic surveillance and fully autonomous weapons. 5. Amodei argued that AI can merge individually harmless data points into comprehensive personality profiles at a scale existing law cannot address. 6. An ETH Zürich and Anthropic joint study showed language models can correctly identify roughly two-thirds of 338 pseudonymous Hacker News profiles at a cost of $1–4 per person. 7. The Pentagon responded by classifying Anthropic as a supply-chain security risk — a designation previously applied only to foreign companies — effectively blocking other contractors from working with Anthropic. 8. Reporter Martin Holland noted the US military continued using Claude technology during weekend airstrikes on Iran even after the classification, citing deep system integration. 9. OpenAI claimed its own models carry the same red lines Anthropic has, though the exact terms of the replacement Pentagon contract remain unclear. 10. A cybercriminal used Claude to infiltrate Mexican government networks, writing Spanish-language commands to detect vulnerabilities and automate data theft, before turning to ChatGPT when Claude raised obstacles. 11. A King's College London study ran 21 conflict simulations and 329 individual games using Claude Sonnet, Gemini Flash, and GPT models — no model ever selected complete capitulation. 12. Nuclear weapons were used in 59% of the simulation runs; under time pressure, AI systems escalated even faster, the opposite of expected machine rationality. 13. Reporter Malte Kirchner noted the researchers believe AI's training makes the nuclear taboo — deeply rooted in human psychology — essentially absent in these systems. 14. Researchers at the Icahn School of Medicine at Mount Sinai found that ChatGPT Health misclassified more than half of real medical emergencies as insufficiently urgent. 15. ChatGPT Health also failed to reliably display its suicide-intervention warning banner when mental health topics were mixed with other health questions simultaneously. 16. The Mount Sinai team recommended AI systems used as first medical-contact points should face the same safety admission tests as regulated medical devices. 17. The episode uses a fast-paced news-brief format with short reporter handoffs, mixing German and English segments, covering a wide range from geopolitics to consumer gadgets. 18. Reporter Steffen Herget described Samsung's Galaxy S26 as conservative on hardware but adding Perplexity, Bixby, and Gemini assistants that run in sandboxes and cannot talk to each other directly. 19. Listeners who follow AI policy, military AI ethics, or AI safety research will find multiple data-dense stories in under 25 minutes. 20. Listeners seeking deep analysis of any single topic or who prefer monolingual German-only audio may find the rapid format and code-switching less suitable.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from KI-Update – ein heise-Podcast