Explainability Through Systematicity: The Hard Systematicity Challenge for Artificial Intelligence

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Philosophy of Mind

Background:

• Explainability is a key AI expectation, but not the sole criterion for advanced AI.
• The "systematicity challenge" historically questioned connectionist AI's ability to achieve systematic thought.
• A richer conception of systematicity, encompassing consistency and coherence, has been overlooked.

Purpose of the Study:

• To propose a broader ideal for AI beyond explainability, focusing on systematicity.
• To offer a conceptual framework distinguishing four senses of "systematicity of thought."
• To re-evaluate the tension between connectionism and systematicity.

Main Methods:

• Conceptual analysis of "systematicity of thought."
• Distinguishing multiple senses of systematicity.
• Examining rationales for systematization and their transferability to AI models.

Main Results:

• A conceptual framework is presented that differentiates four senses of systematicity.
• The perceived conflict between connectionism and systematicity is addressed.
• Five rationales for systematization are identified and applied to AI, revealing the "hard systematicity challenge."

Conclusions:

• AI's systematicity ideal is more demanding than previously understood.
• A dynamic understanding of systematization is proposed, regulating AI's need for systematicity.
• This framework guides how and when AI models should be made more systematic.