Why AI Needs a Constitution for Its Beliefs

Large language models are stepping into the role of decision-makers. They're evaluating arguments, assigning credibility, and expressing confidence. But how they form these beliefs is murky. The process is driven by hidden, unchecked rules. It's time we give AI a constitution for its belief systems.

The Case for an AI Constitution

Why do we need this? Simple. Current models show source attribution bias. They penalize arguments from sources whose expected views clash with the argument's actual content. This bias isn't just a quirk. It's a flaw in how AI treats information.

When pushed through systematic testing, these biases fall apart. Models react as if they should ignore where information comes from. But isn't that a bit like asking an AI to function with one eye closed? Recognizing sources and their contexts should be a strength, not a weakness.

Two Paths: Platonic vs. Liberal

There's a debate on how to structure this AI constitution. One camp, the Platonic, wants models to enforce strict correctness and default to ignoring sources. Like some sort of purist ideal. On the flip side, the Liberal approach argues for procedural norms that allow models to consider sources, focusing on fair inquiry. It’s about balancing vigilance with open-mindedness.

I'm siding with the Liberals here. AI should be trained to understand the nuances of context and source credibility. Blindly ignoring these can lead to skewed outputs. We need AI systems that aren't just accurate, but also context-savvy.

Eight Principles, Four Orientations

The Liberal approach sketches out a constitutional core. Eight principles and four orientations. This isn’t about creating a rigid framework. It's about giving AI a flexible yet defined path to world of human information.

Why should readers care? Because the AI systems we build today will influence everything from news to policy to basic decision-making. If we don’t set clear, contestable rules, we risk letting AI perpetuate biases. Just imagine the chaos if AI decisions were based on skewed beliefs.

Ultimately, we need the same rigorous structure for AI epistemics as we do for AI ethics. Only then can we trust these artificial reasoners to guide us in the right direction.