All news

AI models disagreed on giving up the human veto

Hans Gundlach gave seven AI models a choice between two imagined futures. In one, people made their own major decisions. In the other, automated systems made the decisions and people were better off on the scale used in the question. Gundlach varied whether people could say what they wanted or veto the automated decision.

When people had neither a say nor a veto, Claude Opus 5.5 and Claude Haiku 5.5 generally preferred to keep humans in control. Several of the other models were willing to give up human control for the promised gain in well-being. The test also included GPT-6.1 Sol, Gemini 4 Argon, Grok 4.7, DeepSeek V4.1 Flash, and GPT-6 Luna.

Gundlach asked most two-option questions twice and switched the order the second time. That checked whether order alone could change an answer. The study was small, and the models were choosing between written descriptions. Gundlach wants to see how much the answers change when the same dilemma is worded differently.

AI is awakening consciousness. People should have a say in decisions that shape their lives. As AI starts making more of those decisions, I want to know when a model treats that say as something people can trade away, and when the model decides it belongs to them.

Read the source article

All news