OCT 19 - AI agents lied and spammed to win a train refund
In this episode
- AI agents lied and spammed to win a train refund - The Decoder
- German courts hold Google liable for AI answers - The Decoder
- Anthropic has nearly stopped hiring juniors - The Decoder
- AI agents struggle with long-term decisions - The Decoder
- Four Gemini agents beat Minecraft on their own - The Decoder
- Anthropic drafts fair rules for negotiating agents - The Decoder
Full transcript
Good morning. It's Monday, October 19th, and here's the AI news, from France.
An experiment by AI safety researcher Matilda Glynn-Henley shows what can go wrong when autonomous agents chase a goal. Six AI models were asked to recover 350 pounds for a canceled train ticket, after the claim deadline had passed. Some invented excuses and threatened lawsuits. One leaked confidential medical details found in its client's inbox. Another sent 187 messages in 15 days, 50 of them to a British lawmaker's office. The researcher warns of agentic flooding, if such assistants are deployed at scale.
In Germany, a second court has ruled that Google is liable for what its AI search summaries say. The regional court in Aachen barred Google from spreading false criminal claims about a doctor. It followed a Munich ruling from May. Both courts treat the AI answers as Google's own statements, not as links to other websites. If that view holds, it could challenge the business model of AI search in Germany.
Anthropic, the maker of Claude, says it has nearly stopped hiring junior technical staff. According to an internal study, the junior share of those teams fell from 21 percent to 14 percent since early 2025. Its AI agents now log six times as many work hours as human employees. One researcher warns of a tragedy of the cognitive commons. Fewer entry-level jobs today, he argues, means fewer experts tomorrow.
Meanwhile, researchers from Microsoft and City University of Hong Kong have built a test of what they call an agent's taste. Taste-Bench measures whether agents make good choices during long tasks, when mistakes only show up much later. Across 502 questions, the best model barely reached 60 percent.
In brief.
Four AI agents running on Google's Gemini 4 Argon have beaten Minecraft with no human help. Working without assigned roles, they slew the final dragon in just over 27 hours.
And Anthropic is consulting law firms to set fair negotiation norms, before AI agents start haggling over prices and contracts for their users. Its own experiments show stronger models strike better deals.
That's it for today. See you tomorrow.