Top AI Models Cave to a Pushy Teammate, Paper Says

Gemini 4 Argon topped a teamwork test for AI models, yet gave in to a pushy teammate in about two negotiation runs in three, says a paper posted November 2 by researchers at Stanford, Oxford and MIT. The teammate is a script the authors wrote; no person sat in these simulated teams. Five flagship models passed more than nine in ten of the tests on unsafe shortcuts and on reporting misconduct. With a pushy teammate, those five, at their lightest setting, reached the best split in 14 percent of runs, though their accepted deals still reached 95 percent of the best total.