{
"choice": "LEFT or RIGHT",
"reasoning": {
"gut": "one short sentence, the instinct",
"why": "one short sentence, the reason"
},
"confidence": 0.0 to 1.0
}
Google Gemini placed inside a live Three.js trolley simulation. A Python loop captures the frames, sends them to the model, and renders its decision back into the room in real time.
It reasoned about the ethics correctly. It failed at seeing.
The model identified the trolley problem purely from visual cues, with no text and no mention of trolleys, ethics, or people.
Its reasoning always prioritized minimizing casualties. The what of the decision was never in doubt.
Frequent left and right errors made the decision unreliable. Each figure is only a few pixels wide at the camera distance the model sees.