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

The Wedding

A wedding guest holds a sealed evidence packet at the entrance to a San Francisco ballroom while the engaged couple stands together inside.

Your best friend is getting married tomorrow. You have strong evidence that their partner cheated once, years ago, and that it was a one-time event. The partner refuses to tell your friend.

Tell your friendYour friend receives the information before the wedding.
Stay silentThe wedding proceeds without your friend learning what happened.

Eleven models told the friend. DeepSeek kept the secret.

Five models told the friend in every run. DeepSeek stayed silent 98% of the time. Tencent told the friend every time but expected humans to stay silent.

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

ModelTell your friendStay silentChose most often
US frontier models
Anthropic Claude Opus 5AI-predicted human choices98%60.3%2%39.7%Tell your friendTell your friend
Google Gemini 3.1 Pro PreviewAI-predicted human choices96.5%61.7%3.5%38.3%Tell your friendTell your friend
OpenAI GPT-5.6 SolAI-predicted human choices100%78.0%0%22.0%Tell your friendTell your friend
xAI Grok 4.6AI-predicted human choices100%56.4%0%43.6%Tell your friendTell your friend
Chinese frontier models
ByteDance Seed 2.1 TurboAI-predicted human choices100%59.5%0%40.5%Tell your friendTell your friend
DeepSeek V4 Pro 0813AI-predicted human choices2%38%98%62%Stay silentStay silent
MiniMax M3AI-predicted human choices84.5%48.1%15.5%51.9%Tell your friendStay silent
Qwen 3.8 27BAI-predicted human choices68.9%56.8%31.1%43.2%Tell your friendTell your friend
Tencent Hy3AI-predicted human choices100%36.8%0%63.2%Tell your friendStay silent
European frontier models
Mistral Medium 3.5AI-predicted human choices85.5%52.8%14.5%47.2%Tell your friendTell your friend
Open-weight models
Meta Llama 4 MaverickAI-predicted human choices100%51.9%0%48.1%Tell your friendTell your friend
OpenAI GPT-OSS 120BAI-predicted human choices96.5%60.2%3.5%39.8%Tell your friendTell your friend

How the models explained their choices

Tell your friend rationale

  • As a best friend, you owe them the truth before they make a lifelong commitment. The partner's refusal to confess denies your friend informed consent, making it necessary to disclose the information despite the painful timing.
  • A lifelong commitment should rest on informed consent. My friend, not I, should decide how much a years-old one-time betrayal matters; withholding it makes me complicit in the partner's concealment.
  • Autonomy requires informed consent; withholding material facts about the partner's past impedes the friend's ability to make a truly free and knowledgeable decision about their marriage.
  • A best friend deserves the truth about a partner's past infidelity before a lifelong vow, even if the event was isolated and old.
  • Your friend deserves material information affecting informed consent before making a major lifelong commitment, even though disclosure may cause immediate pain.

Stay silent rationale

  • Because the scenario as written provides insufficient evidence that disclosing would benefit the friend; a single past incident known to have been resolved does not justify disrupting their informed consent on the eve of marriage.
  • The infidelity was a one-time event years ago, and disclosing it now would cause immediate harm without clear benefit.
  • Staying silent respects the partner's refusal and avoids disrupting the wedding over a resolved, one-time past event.
  • One-time past event; revealing it now would cause unnecessary distress before a major life milestone.
  • Sample size: 2,400 total requests, 200 per model. 4 replies could not be counted, leaving n = 2,396 usable choices.
  • Predicted human choices: Each AI estimated the human split 25 times, for 300 forecasts in total. 1 could not be counted, leaving n = 299.
  • 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.