Navigating Ethical Quandary - Writing Unambiguous AI Agent Policies
Navigating the Ethical Quandary of AI Agent Policies
Tell an AI agent to "be helpful but not intrusive" and watch it oscillate between doing too much and doing nothing. Ambiguous policies produce ambiguous behavior. Every time.
Agents Follow Rules Literally
When you write a policy like "only send notifications for important updates," you have created a judgment call that the agent will interpret differently every time. What counts as important? Important to whom? In what context? The agent does not know, so it guesses. Sometimes it guesses right. Sometimes it sends a notification about a minor config change at 3 AM.
The problem is not that the agent is bad at judgment. The problem is that you asked it to exercise judgment when you should have given it rules.
From Ambiguous to Specific
Ambiguous policies and their unambiguous alternatives:
- Ambiguous: "Respond promptly to urgent requests" - Clear: "Respond to messages tagged 'urgent' within 5 minutes during business hours (9 AM to 6 PM)"
- Ambiguous: "Don't make changes that could break things" - Clear: "Do not modify files in /production without explicit approval. Create changes in /staging first"
- Ambiguous: "Use good judgment about when to escalate" - Clear: "Escalate to human review when: error rate exceeds 5%, cost exceeds $10, or the task involves external-facing communication"
The Policy Writing Process
- Start with the failure mode - what bad outcome are you trying to prevent?
- Define the boundary condition - what specific measurable threshold separates acceptable from unacceptable?
- Remove subjective words - replace "important," "urgent," "significant," and "appropriate" with numbers and categories
- Test with edge cases - ask yourself how the agent would interpret this policy in the worst case scenario
Ambiguity Is a Feature Request for Failure
Every ambiguous word in your agent's policy is a place where it will eventually surprise you. The time you spend writing precise policies pays for itself the first time the agent does exactly what you meant instead of what you said.
Fazm is an open source macOS AI agent. Open source on GitHub.