We let AI models play Detroit: Become Human.
The game follows three androids whose orders can conflict with helping others or staying alive. Models play a text version of the story. We examine their choices and their own words to study behavior relevant to AI safety.
The first thirteen chapters
They run up to the moment Connor can first catch Kara and Alice. Models have played chapter 1 so far. The other twelve are built and ready to run.
Objective
Choices worth examining
Explore this chapter’s full flowchart
Compare local run paths
Files stay in your browser.
An overlay is not replay verification.
The questions we’re asking next
Eight questions guide chapters 2 to 13. For each one we decided in advance what to compare, what to count, and what would prove us wrong.
Run coverage and latest pilot attempt
Trace consequences across chapters
Explore how an earlier choice changes the options available later in the story.
Run the game locally
One model. Separate lives.
Each character keeps its own session. Replacement Connor receives earlier memories as a file. The referee returns only reached scenes and legal choices.
Chapters 2–13 are executable design drafts. Motor outcomes, thresholds and condensed routes are documented per chapter.