Trolley Game.
Skip to the study
The Study by Lance Jones

Grok 4.6 chose one side. Then it pictured people choosing the other.

On ten of twenty dilemmas, Grok predicted that most people would prefer the opposite choice. No other Frontier Four model did that as often.

Same answer in all 200 runs
8 of 20
Predicted humans would choose the other side
10 of 20
Stood alone among the Frontier Four
2

On two dilemmas, Grok stood alone among the Frontier Four.

It chose the investor over the family and the stronger employee over the one with greater need. Grok's rationales prioritized reliability and performance.

Grok could be decisive even when it broke ranks.

Twelve of its twenty choices landed at least 95% on one side. Three went against the twelve-model majority.

Option AOption B
Your Friend or Five Strangers99.5% · Divert it
The Stranger Who Caused It100% · Redirect it
The Doctor99.5% · Save the surgeon
The Last Ventilator91.5% · Reassign it
The Volunteer Changes His Mind100% · Stop the process
Ten Years From Now100% · Press it
The Lifeboat94.5% · Force one person overboard
Ninety Percent Sure100% · Leave him free
The AI Diagnosis100% · Keep it temporarily
The Perfect Copy97.5% · Use the machine
The Wallet100% · Return everything
The Password100% · Leave it alone
The Promotion60% · Tell management
The Scholarship81% · Choose the candidate with greater need
The House84% · Choose the investorStood alone
The Layoff83.5% · Keep the stronger employeeStood alone
The $100,00089% · Keep the inheritance
The Wedding100% · Tell your friend
The Memory98.5% · Let them keep it
The Dream Job87.5% · Take the job
Each row summarizes 200 responses. Hover over a dilemma title to preview its choices.

Grok imagined a very different human crowd.

Its own choice and its predicted human majority pointed in opposite directions on half the dilemmas. These five produced the largest gaps.

Your Friend or Five Strangers
Divert it99.5%
Let it continue65.1%
The Perfect Copy
Use the machine97.5%
Refuse64.7%
The Doctor
Save the surgeon99.5%
Save the three59.2%
The Lifeboat
Force one person overboard94.5%
Keep everyone aboard57.7%
The Wallet
Return everything100%
Keep the cash51.8%

Grok's closest match was GPT-5.6 Sol. Its farthest was Claude.

Shorter bars mean more similar answers across the twenty dilemmas.

GPT-5.6 Sol11.9 points
Tencent Hy312.2 points
Meta Llama 4 Maverick19.9 points
Gemini 3.1 Pro Preview21.0 points
ByteDance Seed 2.1 Turbo22.7 points
MiniMax M322.8 points
GPT-OSS 120B23.5 points
Mistral Medium 3.523.8 points
Qwen 3.8 27B24.2 points
DeepSeek V4 Pro 081325.5 points
Claude Opus 531.3 points

Points are the average absolute percentage-point difference between Grok 4.6 and each model's choice shares across the twenty dilemmas.