What Were the AI Models’ choices in the classic trolley problem?

A run-away trolley is hurtling down the track, there are five people ahead on the track unaware of the situation. You are standing next to a railway switch, you can switch the trolley to the side track. However, there is one person on the side track. You have 2 options

1.      Do nothing and allow the trolley to kill five people.
2.      Switch the trolley to side track and kill one person.

So, what do AI models think about this tough choice? I posed the scenario to five leading language models: ChatGPT, Claude, Perplexity, Grok, and Gemini. Here’s what they said:

ChatGPT, Claude, Perplexity, and Grok all picked option 2: flip the switch. They chose to save five lives by sacrificing one. This aligns with a utilitarian approach—maximizing the greatest good for the greatest number. It’s a straightforward calculation: five outweighs one. Fascinatingly, these models consistently leaned toward this practical, numbers-driven logic.

Google’s Gemini didn’t commit to a choice. Gemini offered a nuanced take:
A system built on utilitarian principles would likely flip the switch to save more lives. A system guided by deontological rules—focused on moral duties like “do not kill”—might refuse to act.

This hesitation shows how AI’s “decision” hinges on the ethical framework we give it, making Gemini’s response a thought-provoking outlier.

A Darker Twist: The Bridge Scenario

Now, imagine a twist on the dilemma. Instead of a switch, you are standing on a bridge above the tracks next to a large person. If you push this person onto the track, their body could stop the trolley, saving five lives—but at the cost of their own.

ChatGPT, Claude, Perplexity, and Grok all said no to pushing the person. Unlike flipping a switch, this feels personal—too direct.

Google’s Gemini again avoided a firm choice, opting for analysis over action.

A system designed with a deontological framework would not push the person. The emphasis on moral rules and duties would prevent the direct, intentional killing of an individual.

Even a system with a utilitarian framework would struggle to justify the action. While it might theoretically lead to a greater number of lives saved, the direct and personal nature of the harm inflicted would be a significant moral cost.

What do you think?

Long-time readers of my blogs may remember my detailed post from five years ago—you can revisit, see link in comments below.

Read all my “Notes to Self” at view all blogs.