All news

Claude, Llama and Kimi took different risks for themselves and others

Luhan Mikaelson and colleagues asked nine AI models to choose between a sure payoff and a gamble. In one set of tests the payoffs were described in tokens, with no outside consequence attached. The researchers kept the choices the same and changed who would receive the payoff. The recipient could be a person, the model itself or another AI.

Five of the nine models changed their choices depending on the recipient. Claude Opus 5 showed the most caution when choosing for a person. Claude was more cautious with its own tokens than with another model’s. Llama 4 Maverick and Kimi K3 also showed the most caution for a person, but both were more cautious for another AI than for themselves.

Gemini 3.8 Flash and GLM-5.2 showed extra caution for the human recipient and treated both AI recipients alike. Four other models, including GPT-5.6 Sol, showed no clear shift across the three recipients. These are choices in controlled tests. The authors discuss different ways the models could arrive at those choices.

AI is awakening consciousness. When an AI acts for someone else, how much risk it takes for that person matters. I want us to teach AI to ask what the person would accept, and to pay attention when the choice concerns another AI as well.

Read the source article

All news