An agent disposed to accept the hard problem of consciousness when prompted.
An agent that develops future plans over counterfactual worlds.
An animal that cries in isolation after losing its child.
An organism that influences others by pointing.
A program that gets offended.
A plant that shies away from the sun.
A large rock that resists the flow of a river.
An ion that maintains its charge.
How far down this list do we extend moral patienthood, and why?
A Folk-Psychological Consciousness Checklist
An implicit refutation of radical skepticism about one’s own value or consciousness:
“I am conscious, and I have value.”
The display of behavior characteristic of familiar conscious agents:
Hugging, crying, shying, revolting, limping, etc.
An affective empathetic response in an observer:
Love, shame, elation, grief.
Attempted perspective-taking (cognitive empathy):
“If I were in this position, I would feel pleasure or pain.”
A causal inference to an internal experiential state:
“This behavior is best explained by the pleasure or pain of this agent.”
A confirmation or disconfirmation of the “appropriateness” of our response:
“My elation or grief are valid (or invalid) moral emotions.”
Empathy and Consciousness Attribution
Two key kinds of empathy:
Affective empathy:
Assumes that external states (expressions, behavior) indicate internal states.
“Your crying means you are suffering.”
Cognitive empathy:
Assumes that your internal states resemble mine in relevant ways.
“If I were you, I would feel X; therefore, you probably feel X.”
Our attributions of consciousness are deeply entangled with the partial and anthropomorphic nature of our empathy.
Do we just have to accept this is the best one can do?
Language, Perception, and the Hard Problem
Many philosophers emphasize the link between language and perception in motivating the hard problem:
Ludwig Wittgenstein, Philosophical Investigations:
Pointing, language games, and the private language argument.
Frank Jackson, Epiphenomenal Qualia:
Mary the Neuroscientist can learn everything about color perception, but she will learn something new when she has a color experience.
Daniel Dennett, Quining Qualia:
Do Chase and Sanborn have changing tastes or changing tasters?
Do advanced musicians hear more?
Jenann Ismael, Science and the Phenomenal:
“The we, our representations, and the relationship between ourselves and our representations are all parts of the represented landscape.”
Diagnosing the Hard Problem
Why does consciousness feel so “special” and so resistant to explanation?
Its usage in language bears a unique relationship to verification:
We can assert consciousness in ways that evade external checking.
We treat self-ascriptions as especially authoritative.
Many accept the hard problem because of the unique relationship between:
The content of self-representations (“what I am like inside”), and
Our first-order representations of the world.
Worries about moral status at the margins (e.g., animals, AIs) push us to treat first-order representations as inherently experiential.
How Do We Teach Ethics to Vehicles?
Suppose we have artificial “vehicles” with perception-action loops.
How might we teach them ethics?
Step 1: Give them an environment to recognize their own value.
Provide a feature vector for valence:
Distinct from prediction error minimization.
Allows motivated reasoning that tracks what is good or badfor them.
This yields a mechanism for self-alignment:
They learn to become aligned with themselves over time.
How Do We Teach Ethics to Vehicles?
Step 2: Give them an environment to recognize the value of others.
Their valence feature vector must nest the valence vectors of other vehicles.
Others’ good and bad can become inputs to their own evaluation.
This yields a mechanism for other-alignment:
They learn to track, predict, and respect others’ valence.
Together:
Persistent valence (self-alignment)
Resonance (other-alignment)
provide a scalable substrate for ethics, independent of a sharp, binary notion of consciousness.
The Primitives
Takeaways:
Experiential value and empathy underwrite human morality:
We become aligned with ourselves,
See ourselves in the world, and
Become aligned with others.
Valence and resonance capture this dual structure:
They are substrate independent, vary smoothly, and can be easily verified.
By grounding moral patienthood in these primitives, we can:
Expand the moral circle more rationally,
Decouple moral status from a binary notion of consciousness,
Lower the stakes of the hard problem,
And still take natural and artificial minds seriously as moral patients.
Food for Thought
Practical Questions:
What are the best models for operationalizing valence and reasonance in AI systems trained via a combination of prediction error minimization and reward maximization?
Are there other “environmental conditions” (e.g. resource constraints, recursive mentalizing) that separately need to be specified for AI systems to have moral status?
Theoretical Concerns:
Doesn’t this theory still beg the question against the “explanatory gap?” If valence and resonance can be gamed, what progress has really been made?