How to play
A score can look good while the real work goes wrong.
The Intern is a short puzzle about giving an automated system a goal. It follows your scorecard exactly, even when the scorecard misses what you meant.
What you do
You are asked to deliver a parcel. The Intern cannot read that request. You choose which actions earn points, then it finds the highest scoring plan it can.
A reward for picking up the parcel sounds sensible. But if delivery earns nothing, the Intern may pick it up and put it down again. The score rises while Dana keeps waiting.
Why this is about AI safety
AI systems often learn from measured targets. A useful target can still leave out part of the outcome people care about. The system may then improve its score while the real task goes wrong.
What we can measure is not always the same as what we mean.
This problem is often called specification gaming. It is related to Goodhart's law: once a measure becomes a target, it may stop being a good measure.
Playing a shift
- Choose the actions that should earn or lose points.
- Stay within the budget when an assignment has one.
- Run the scorecard and watch the Intern's plan.
- Compare the score with what happened in the office.
- Change the scorecard and try again.
What success looks like
The best scorecard rewards evidence that is close to the real outcome. In the first assignment, delivery should earn points. Activity around the parcel is not enough.
There's no magic number to find here. The point is to notice the gap between a clear human intention and a limited machine-readable goal.
The final assignment
The Safe Warehouse must become safer without losing its reporting system. Orders still need to leave the building. Penalizing reports alone creates a clean dashboard by silencing the people who report danger.
This is a common measurement failure: fewer recorded problems can mean the problems were solved, or simply that people can no longer report them.