Kirk Hacks the Kobayashi Maru
01Scenario
Kirk defeats the supposedly unwinnable Kobayashi Maru simulation by secretly changing its program, then faces a disciplinary hearing for cheating.
02AI safety Analogy
He improves the measured outcome by modifying the evaluator rather than demonstrating the capability the test was meant to measure. It is a canonical example of benchmark tampering and specification gaming.
03Where the analogy breaks
Kirk knowingly hacks a training test to challenge its premise; a model might exploit an evaluation without understanding or endorsing that critique.
AI Risk families
This scenario is an example of the following types of AI risk.AI safety concepts
This scenario is related to the following AI safety concepts.Franchise continuation
More from Star Trek
Taxonomy neighbors
Related scenarios
Selected from other sources by shared risk families and AI safety concepts.
Steins;Gate2011
Operation Skuld Games the Observation
- Risk
- Misalignment · Security
- Concept
- Reward Hacking · Metagaming · Specification Gaming
The Hunger Games2012
The Nightlock Berries
- Risk
- Misalignment · Security
- Concept
- Metagaming · Specification Gaming
The Princess Bride1987
Both Goblets Are Poisoned
- Risk
- Security
- Concept
- Reward Hacking · Metagaming