Outer alignment

The topic of this article may not meet Wikipedia's general notability guideline. Please help to demonstrate the notability of the topic by citing reliable secondary sources that are independent of the topic and provide significant coverage of it beyond a mere trivial mention. If notability cannot be shown, the article is likely to be merged, redirected, or deleted. Find sources: "Outer alignment" – news · newspapers · books · scholar · JSTOR (June 2025) (Learn how and when to remove this message)

This article needs additional citations for verification. Please help improve this article by adding citations to reliable sources. Unsourced material may be challenged and removed. Find sources: "Outer alignment" – news · newspapers · books · scholar · JSTOR (June 2025) (Learn how and when to remove this message)

Outer alignment is a concept in artificial intelligence (AI) safety that refers to the challenge of specifying training objectives for AI systems in a way that truly reflects human values and intentions.

A significant theoretical insight into alignment comes from computability theory. Some researchers argue that inner alignment is formally undecidable for arbitrary models, due to limits imposed by Rice's Theorem and Turing’s halting problem. This suggests that there is no general procedure for verifying alignment post hoc in unconstrained systems.

To circumvent this, the authors propose designing AI systems with halting-aware architectures that are provably aligned by construction. Examples include test-time training and constitutional classifiers, which enforce goal adherence through formal constraints. By ensuring that such systems always terminate and conform to predefined objectives, alignment becomes decidable and verifiable.^[1]

• Inner alignment
• AI alignment
• Goodhart's law
• Specification gaming
• Reward hacking
• Artificial general intelligence
• AI safety
• Interpretability (machine learning)