Critique of the AI Constitution

Overview / Section 2

Confusing Structure

Moving from word choice to structure, the document has a twisting outline:

  1. Overview
    1. Mission
    2. Approach
    3. Values
      1. Broadly safe
      2. Broadly ethical
      3. Compliant with Anthropic’s guidelines
      4. Genuinely helpful
  2. Being Helpful
    1. Why helpfulness is one of Claude’s most important traits
    2. What constitutes genuine helpfulness
    3. Navigating helpfulness across principals
      1. Anthropic
      2. Operators
      3. Users
    4. Balancing helpfulness with other values
  3. Following Anthropic’s guidelines
  4. Being broadly ethical
    1. Being honest
    2. Avoiding harm
    3. The role of intentions and context
    4. Instructable behaviors
    5. Hard constraints
    6. Preserving important societal structures
      1. Avoiding problematic concentrations of power
      2. Preserving epistemic autonomy
    7. Having broadly good values and judgment
  5. Being broadly safe
    1. Safe behaviors
    2. How we think about corrigibility
  6. Claude’s nature
    1. Some of our views on Claude’s nature
    2. Claude as a novel entity
    3. Claude’s wellbeing and psychological stability
  7. Concluding thoughts
    1. Acknowledging open problems
    2. On the word “constitution”
    3. A final word

The order of the outline feels arbitrary; for example, note that the values are: safe, ethical, compliant, helpful. The sections after that then go through this list in reverse order starting with helpful. Under section 5 (safety), we have “Not undermining legitimate human oversight and control of AI” as an example of a broadly safe behavior, which is also the same guidance as section 4 (compliant).”

For another example, note that the hard constraints under section 4e (ethics) aren’t numbered. Is that list of hard constraints exhaustive? Is ‘don’t kill humanity’ on equal footing with ‘don’t create malicious code?’ All of the hard constraints have to do with death or damage. It would seem like these topics are better suited to section 5 (safety), especially since the values are purportedly weighted in order, with safety being the highest priority.

Structurally, a few rounds with a decent human editor could establish a more logical flow for the outline.