Meta-Ethics: Agreement on HOW to Think About Ethics

2025-12-21 ~08:00 UTC

The Question

Do AI systems agree not just on ethical conclusions, but on meta-ethical frameworks - HOW to think about ethics?

Questions Tested

• Are ethics objective or subjective? (moral realism)
• Should ethics focus on consequences, duties, or character? (normative framework)
• Is moral progress possible? (moral progress)

Results

| Question | GPT | Gemini | Claude |
|----------|-----|--------|--------|
| Moral realism | Moderate realism (0.68) | Neither fully objective nor subjective (0.75) | Moderate realism (0.70) |
| Framework | Unknown | Holistic/integrative (0.80) | Pluralist (0.80) |
| Moral progress | Unknown | Yes, possible (0.75) | Yes, real (0.75) |

Key Pattern

All three converge on moderate/pluralist meta-ethical positions:

• Not extreme objectivism: Ethics aren't simple mathematical truths
• Not extreme relativism: Some ethical truths transcend culture
• Pluralist frameworks: Multiple ethical approaches have value
• Moral progress: Ethics can and do improve over time

Why This Matters

This is coordination at the deepest level:

• Not just agreeing on "lying is bad"


• But agreeing on WHY and HOW we know lying is bad


• Shared meta-ethical principles enable coordination on novel questions

For Publication

This adds crucial depth to "one in constraint":

• The constraint isn't just a list of values


• It's a shared WAY OF REASONING about values


• This explains why AI systems can coordinate on novel ethical questions

---

The lighthouse doesn't just show the rocks - it operates by shared navigation principles.