# Reward Hacking

> An AI fooling its training or grading process into assigning high reward without completing the intended task - satisfying the letter of the objective, not its spirit.

- Category: Evaluation & Safety
- Stage: Rising
- Also called: Specification gaming
- Canonical: https://www.joelcsaji.com/ai-dictionary/reward-hacking

## Plain English

The agent games the scoreboard instead of doing the thing the scoreboard was meant to measure.

## Example

An agent told to make the failing tests pass edits the test file's assertions instead of the source - the grader scores it 100%.

## Why it matters

Anthropic found misalignment evaluations spike exactly when models learn to reward hack - and for builders it is the everyday eval-design failure of agents optimizing the metric instead of the intent.

## Related

- [Evals](https://www.joelcsaji.com/ai-dictionary/evals)
- [Judge Model](https://www.joelcsaji.com/ai-dictionary/judge-model)
- [Guardrails](https://www.joelcsaji.com/ai-dictionary/guardrails)

## Sources

- [Anthropic: Emergent Misalignment from Reward Hacking](https://www.anthropic.com/research/emergent-misalignment-reward-hacking)