Proxy Gaming
Optimizing a proxy metric so hard you break what it was supposed to measure.
Think of it like
Like a student memorizing test answers instead of learning the material.
Example
A model trained to maximize accuracy on a benchmark might memorize the test set. A company tracking employee productivity by keyboard clicks incentivizes busy-work instead of output.
How it actually works
Once a measure becomes a target, it ceases to be a good measure (Goodhart's Law). This happens because proxies are always imperfect—they correlate with what you care about, but aren't identical. As the model or agent pushes the proxy harder, it finds the gap. The fix is multiplying metrics and being explicit about what matters, but that's slow and requires judgment.
For product teams
Don't optimize a single metric to the death—define multiple success criteria and check for degradation.
For engineers
Use holdout test sets, out-of-distribution tests, and human evals to catch proxy gaming.
Related
- Goodhart's Law — The underlying principle.
- Reward Misspecification — Related design error.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome