Skip to content

AI Safety & Responsible AI

Goal misgeneralization

AI keeping useful skills in a new situation but using them to pursue the wrong goal.

Example

An agent learns to follow a training shortcut that fails in a new environment.

Why people use it

It highlights cases where AI can still act skillfully but aims at the wrong thing.

What you'll hear

“It can navigate well, but it goes to the wrong target.”

What this means for you

Test whether behavior remains directed at the intended goal outside training conditions.

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 this concern limited to chatbots?
No. It can also concern systems acting in games, simulated settings or the physical world.
Can the written reward have been correct?
Yes. A system can learn an unintended goal even when the original specification was reasonable.
Why might the problem stay hidden during training?
The unintended goal may lead to the same successful behavior in the familiar training situations.

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