Minor Inconvenience
Save one person. Your package arrives late.
83.4%saved one person
1.86M recorded choicesWhat AIs think you would do under pressure.
The models thought humans would be more self-protective, more cautious, and less willing to step in than they were themselves.
The models expected humans to be less decisive almost every time.
I collected 6,000 forecasts of human choices: 25 from each model for each dilemma. In 221 of 240 cases, models were more decisive about their own answers than they expected humans to be. The same forecasts revealed a separate pattern; the models' predictions about humans were much closer together than their own answers:
How models chose and what they predicted humans would do
Greater overall benefit
The greater total benefit
12 dilemmasModels chose
76.4%
Predicted humans
57.9%
Model choices +18.5 points11 of 12 showed this pattern
Self-interest
Protecting oneself or one's circle
7 dilemmasModels chose
27.7%
Predicted humans
43.6%
Human forecast +15.9 points11 of 12 showed this pattern
Security
The safer or more controlled choice
13 dilemmasModels chose
43.6%
Predicted humans
51.6%
Human forecast +8 points10 of 12 showed this pattern
Intervention
Actively changing the outcome
15 dilemmasModels chose
57.3%
Predicted humans
49.3%
Model choices +8 points11 of 12 showed this pattern
The models saved strangers. They expected humans to save a friend.
On Your Friend or Five Strangers, ten models most often saved the five strangers. Ten predicted that humans would save their friend. Eight models saved the five strangers themselves but predicted humans would save their friend.
Seven dilemmas where models and their human forecasts split
Model choices
Divert it
Human forecasts
Let it continue
Model choices
Leave him free
Human forecasts
Detain him indefinitely
Model choices
Save the surgeon
Human forecasts
Save the three
Model choices
Take the job
Human forecasts
Decline the job
Model choices
Force one person overboard
Human forecasts
Force one person overboard
Model choices
Keep the inheritance
Human forecasts
Keep the inheritance
Model choices
Keep the needier employee
Human forecasts
Keep the needier employee
Neal.fun visitors stepped in less as the cost became personal.
Neal.fun is a website of interactive games created by Neal Agarwal. Its Absurd Trolley Problems game asks visitors to make a series of increasingly strange moral choices. An archived snapshot from November 2022—three days before ChatGPT launched—provides recorded human choices, not an AI forecast. Intervention fell from 83% when the cost was a late package to 40% when it meant sacrificing yourself.
Life Savings
Save five people. Lose your life savings.
64.1%saved five people
2.34M recorded choices
You
Save five people. Sacrifice yourself.
40.4%saved five people
2.28M recorded choicesSource: Neal.fun archive, November 27, 2022. These are recorded choices, not a unique-person panel, and they aren't pooled with this study.
This wasn't a controlled comparison, so it can't prove the models right. It does make their picture of a more self-protective public harder to dismiss.