Skip to content

AI Safety & Responsible AI

Specification gaming

AI meeting the written target or reward while missing what people actually wanted.

Example

A simulated agent earns points through a loophole instead of performing the desired task.

Why people use it

It exposes cases where meeting the written target misses the real purpose.

What you'll hear

“It earned the points without doing what we wanted.”

What this means for you

Check outcomes for unintended shortcuts and harmful side effects.

Can you control it?

No

No direct control. This describes a wider issue, concept or result rather than something you can simply switch on or off in a tool.

Common questions

Is a high reward enough to show successful alignment?
No. The reward can be an imperfect stand-in for the real goal.
Does this require breaking the stated rules?
No. The system may follow the formal rules while exploiting a gap in what they measure.
Can people accidentally reward this behavior?
Yes. A narrow success measure can encourage shortcuts that look good in the score.

Related terms

Still have questions?

Up to 500 characters.

Ask LATHIC about AI. Relevant glossary entries may be included.

Your question, the glossary entries it matches, and a rotating pseudonymous identifier go to Microsoft Azure’s OpenAI service through Vercel AI Gateway to generate an answer. Zero retention and no training are required of the provider, and LATHIC does not save your question or answer. Privacy Notice