Skip to content
Home › Escalation to Human

Claude's Constitution

Escalation to Human

Definition

Anthropic's public statement of Claude's values and behavioural priorities (the "model spec" the exam draws escalation/safety reasoning from). It grounds the why behind escalation triggers and the minimal-footprint principle.

Key points

  • Principal hierarchy: Anthropic (system/policy) > operators (the deploying business) > users. Claude follows the chain of command; a user cannot override an operator's legitimate constraints.
  • Reversibility-weighted harm: weigh actions by how hard they are to undo — prefer reversible actions; treat irreversible/high-impact actions with more caution.
  • Least-drastic-action / minimal footprint: take the smallest action that resolves the task; do less when scope is unclear; abandon rather than proceed recklessly.
  • Handling ambiguity: don't improvise on silent/ambiguous policy — surface it / escalate rather than inventing a rule.
  • Honesty: don't fabricate; "unhelpfulness is never trivially safe" — refusing/abandoning still has a cost and should be weighed, not used as a lazy default.
  • Human & emergency escalation: escalate to humans on explicit request, policy gaps, or genuine inability to make progress.

Why it matters for the exam

Escalation and "minimal footprint" answer choices are phrased in the language of the Constitution. Knowing the principal hierarchy and reversibility weighting disambiguates the close options.

Common gotchas

  • Fine-tuning / RLHF / Constitutional AI internals are out of scope — only the behavioural values are tested.
  • "Do more to be helpful" is wrong when scope is unclear — the spec favours doing less.

See also

Sources

Referenced by