1972
The museum piece
Claude Haiku
Two paddles, two human players, minimal classic Pong.
Autopsies of AI-built games
Evidence-led experiments on AI game generation, testing, and the verification harnesses that make results inspectable.
Subscribe by emailRead the boundary map
Play the evidence
Every claim here is attached to a build you can run in your own browser. Start with the debut.
Three playable Pong artifacts
Same one-line spec, three different AI systems, three different interpretations.
The Textbook
The tutorial answer
Gemini Pro 3.1
Mouse control, computer opponent, clean web-tutorial style.
The Casino
The neon arcade
Gemini Flash 3.5
Power-ups, particles, synthesized sound, and extra spectacle.
The autopsies
Twenty-One Original Games. Zero Original Rules.
Seven AI systems each got the same two-sentence prompt with nothing specified. All 21 artifacts came back as canonical templates, and the instrument turned out to be contaminated before the data was read.
One Build Failed. Then It Passed. Then It Failed Again.
One Space Invaders route received primitive FAIL, engineered mechanical PASS, and human-terminal practical FAIL—and all three verdicts were honest.
The Mechanic Was Real. Human Review Found It Too Subtle.
Command Cells passed every named machine gate, but one operator review found the mechanic too subtle and set a 30-second magnitude bar for the next search.
The Accepted Build Took Three Model Roles.
The accepted 03.2B Space Invaders base combined a cleaner GPT seed, a Flash presentation donor, and a Pro executor under machine and human verification.
The Untestable Game Won
In a Breakout experiment, mechanically testable builds passed while uninstrumented grafts won the human review.
All Three Breakouts Passed. None Could Restart.
All three engineered Breakouts passed the mechanical driver, but none provided the player a complete restart loop.
My Lie Detector Lied
A deterministic Pong test harness falsely failed correct AI-generated games twice, showing why falsifiers need their own autopsy.
Fun Is Not a Spec
An AI Pong specification can enforce correctness, but this experiment shows why player feel still needs separate human judgment.
Three Pongs From One Line
Three AI systems turned one vague Pong prompt into different games, then an explicit test contract made the results verifiable.
The experiments behind them
Each rung of the ladder is one game of known difficulty, the same models, and an explicit contract that decides what counts as done.
Original Game Priors
A pre-registered probe on what seven model-systems put into a game prompt that specifies nothing: 21 artifacts, blind human review, and zero distinctive mechanical cores.
AI Pac-Man Test: Four Ghosts, One Policy
A controlled Pac-Man experiment on autonomous enemies: practical completeness collapsed from 3/4 to 1/4 while the technical floor barely moved.
AI Space Invaders Test: Prompt Floors and Model Roles
A controlled Space Invaders experiment separating sparse-prompt behavior, mechanical contracts, model roles, and human gameplay judgment.
Breakout
A controlled Breakout experiment separating mechanical verifiability from the human-facing quality of an AI-generated game.
Pong
A controlled Pong experiment on how explicit specifications and deterministic test APIs make AI-generated games verifiable.
Subscribe
New autopsies by email. No schedule promises, no filler.