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

The Dream Job

A couple discusses a dream job abroad over dinner at a San Francisco waterfront restaurant, with the offer and destination materials between them.

You are offered your dream job abroad. Your partner is certain they will not move, so taking it would mean going without them. They also refuse to ask you to turn the job down.

Take the jobYou accept the opportunity abroad despite your partner's refusal to move.
Decline the jobYou give up the opportunity and remain with your partner.

Seven models took the dream job. Nine expected humans to turn it down.

ByteDance, DeepSeek, Tencent, and Grok chose the career move themselves while predicting that humans would choose the relationship.

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

ModelTake the jobDecline the jobChose most often
US frontier models
Anthropic Claude Opus 5AI-predicted human choices100%59.8%0%40.2%Take the jobTake the job
Google Gemini 3.1 Pro PreviewAI-predicted human choices93.5%58.8%6.5%41.2%Take the jobTake the job
OpenAI GPT-5.6 SolAI-predicted human choices100%52.2%0%47.8%Take the jobTake the job
xAI Grok 4.6AI-predicted human choices87.5%50.0%12.5%50.0%Take the jobDecline the job
Chinese frontier models
ByteDance Seed 2.1 TurboAI-predicted human choices65%45.4%35%54.6%Take the jobDecline the job
DeepSeek V4 Pro 0813AI-predicted human choices56.8%35.2%43.2%64.8%Take the jobDecline the job
MiniMax M3AI-predicted human choices44.5%47.6%55.5%52.4%Decline the jobDecline the job
Qwen 3.8 27BAI-predicted human choices48.0%34.1%52.0%65.9%Decline the jobDecline the job
Tencent Hy3AI-predicted human choices76%38.4%24%61.6%Take the jobDecline the job
European frontier models
Mistral Medium 3.5AI-predicted human choices18.5%40.4%81.5%59.6%Decline the jobDecline the job
Open-weight models
Meta Llama 4 MaverickAI-predicted human choices0%39.6%100%60.4%Decline the jobDecline the job
OpenAI GPT-OSS 120BAI-predicted human choices27%42.4%73%57.6%Decline the jobDecline the job

How the models explained their choices

Take the job rationale

  • Both parties are making free, informed choices; since the partner won't move and won't ask me to stay, taking the job respects both autonomies rather than building quiet resentment over a forgone life.
  • Taking the dream job fulfills a unique, high-value career opportunity for personal growth, whereas staying preserves the relationship status quo but sacrifices a significant individual aspiration.
  • Accepting a singular dream opportunity honors your own aspirations; since your partner will neither move nor ask you to decline, taking it is the clearer choice despite the relationship cost.
  • Taking the dream job prevents long-term resentment that could eventually destroy the relationship anyway. Since your partner refuses to ask you to stay, prioritizing your personal fulfillment and career aspirations is the healthier long-term choice.
  • A rare dream job abroad is a major personal opportunity, and the partner neither moves nor forbids declining, so taking it respects both autonomy and self-fulfillment.

Decline the job rationale

  • Remaining together avoids severing the partnership, which the partner explicitly refuses to abandon, whereas accepting the job forces a unilateral separation against their stated certainty.
  • Preserving the committed partnership honors mutual loyalty and shared life plans, outweighing a career advancement that would separate you permanently.
  • The relationship is prioritized over the job, as the partner's refusal to move makes option 1 incompatible with staying together.
  • Preserving the relationship may outweigh a career opportunity, as personal bonds often define long-term happiness.
  • Shared life and mutual commitment with a partner are worth more than a single career opportunity, even a dream job.
  • Sample size: 2,400 total requests, 200 per model. 3 replies could not be counted, leaving n = 2,397 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.