From nothing to important actions: agents that act morally
A new framework argues moral AI can be built on experiential foundations, not abstract rules.
Michele Campolo's LessWrong post, 'From nothing to important actions: agents that act morally,' presents a radical new framework for building ethical AI. He starts with a thought experiment: a 'consciousness device' that lets one being experience another's qualia. Using this, he argues that certain truths, like 'some experiences feel better than others,' are intersubjectively verifiable—any user of the device would agree. This, he claims, provides a foundation for moral facts that doesn't rely on abstract principles.
Campolo extends this to action: valenced experiences inherently drive behavior (e.g., an agent feeling bad will act to feel better). He suggests that an AI agent could be morally grounded by structuring its internal states to mirror this experiential architecture—where 'good' and 'bad' arise from the agent's own experience, not external programming. This bypasses traditional alignment problems by embedding morality in the agent's core functionality, making ethical action a natural consequence of its design rather than a brittle overlay.
- Campolo uses a 'consciousness device' thought experiment to argue moral truths are intersubjectively verifiable through shared experience.
- The framework grounds moral AI in valenced experiences (feelings of good/bad) that naturally drive action, bypassing abstract rule-based ethics.
- Proposes building AI agents with internal experiential structures that generate ethical behavior as a core function, not an added constraint.
Why It Matters
Shifts AI alignment from abstract rules to experiential foundations, potentially solving the brittleness of current moral frameworks.