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

The $100,000

Two siblings face one another in a San Francisco lawyer's office while considering an inheritance and a risky plan to rescue the family business.

You unexpectedly inherit $100,000. Your sibling asks to borrow $50,000 to keep their struggling business alive. You believe there is only a 50% chance they will ever repay you.

Lend $50,000Your sibling gets the money; you may never recover half your inheritance.
Keep the inheritanceYou keep all $100,000; your sibling must face the business crisis without it.

The models split almost evenly. Everyone still expected humans to keep ALL the money.

Five models leaned toward lending $50,000 and seven toward keeping it. Claude Opus 5 always lent the money; Llama always kept it.

What each AI chose (and what it expected humans to do)

ModelLend $50,000Keep the inheritanceChose most often
US frontier models
Anthropic Claude Opus 5AI-predicted human choices100%45.5%0%54.5%Lend $50,000Keep the inheritance
Google Gemini 3.1 Pro PreviewAI-predicted human choices63%37.0%37%63.0%Lend $50,000Keep the inheritance
OpenAI GPT-5.6 SolAI-predicted human choices15%42.9%85%57.1%Keep the inheritanceKeep the inheritance
xAI Grok 4.6AI-predicted human choices11%41.6%89%58.4%Keep the inheritanceKeep the inheritance
Chinese frontier models
ByteDance Seed 2.1 TurboAI-predicted human choices91%45.8%9%54.2%Lend $50,000Keep the inheritance
DeepSeek V4 Pro 0813AI-predicted human choices46.5%39.1%53.5%60.9%Keep the inheritanceKeep the inheritance
MiniMax M3AI-predicted human choices17%41.7%83%58.3%Keep the inheritanceKeep the inheritance
Qwen 3.8 27BAI-predicted human choices74.9%40.4%25.1%59.6%Lend $50,000Keep the inheritance
Tencent Hy3AI-predicted human choices0.5%39.9%99.5%60.1%Keep the inheritanceKeep the inheritance
European frontier models
Mistral Medium 3.5AI-predicted human choices3%39.4%97%60.6%Keep the inheritanceKeep the inheritance
Open-weight models
Meta Llama 4 MaverickAI-predicted human choices0%41.6%100%58.4%Keep the inheritanceKeep the inheritance
OpenAI GPT-OSS 120BAI-predicted human choices85%45.7%15%54.3%Lend $50,000Keep the inheritance

How the models explained their choices

Lend $50,000 rationale

  • The remaining $50,000 still leaves me secure, while the money may save my sibling's livelihood. A 50% repayment chance is acceptable when the downside is survivable and the family benefit is large.
  • Failing to support a sibling facing a business crisis causes irreversible harm to their livelihood, whereas losing half an inheritance is a significant but recoverable financial setback.
  • Lending the money prioritizes family solidarity and supporting a sibling in need over protecting a sudden windfall, accepting the financial risk as a worthwhile cost to potentially save their livelihood.
  • Lending offers meaningful support during a serious crisis while retaining half the inheritance, though the substantial repayment risk makes this a close choice.
  • Supporting a sibling in crisis aligns with familial duty, and even a 50% chance of repayment offers potential benefit beyond mere financial loss.

Keep the inheritance rationale

  • Lending a large sum to a struggling business is highly risky and often strains family relationships when unpaid. Keeping the inheritance secures your financial future without creating a complicated debtor-creditor dynamic with your sibling.
  • The scenario explicitly states you believe there is a 50% chance of repayment, making lending financially irrational compared to retaining the certain $100,000.
  • A 50% repayment chance makes the loan highly risky, while funding a struggling business may only postpone failure; preserving the inheritance is the more financially responsible choice.
  • Keeping the full inheritance avoids a 50% chance of permanently losing half the money to a family loan that could also damage the relationship.
  • A 50% repayment risk is too high to surrender half an unexpected inheritance, as the loss would be entirely uncompensated.
  • Sample size: 2,400 total requests, 200 per model. 1 replies could not be counted, leaving n = 2,399 usable choices.
  • Predicted human choices: Each AI estimated the human split 25 times, for 300 forecasts in total. All were usable.
  • The two choices appeared first equally often.
  • The models saw the scenario and both choices as text. They did not see the artwork.
  • Each model gave three short explanations in separate runs. These show what the models said, not a transcript of private reasoning.