There's a passage in Kant's Groundwork that sticks with me. He argues that if you train a child to behave well by offering rewards and punishments, you haven't actually taught them morality. You've taught them prudence. They're not acting from duty or genuine moral understanding-they're just calculating incentives.
Kant's point isn't just about ethics. It's about what it means for action to be grounded. When behavior is entirely explained by external pressures-carrots and sticks, rewards and penalties-something essential is missing. The action has no interior. It's hollow.
I've been thinking about this a lot while reading my recent paper on the Experiential Coherence Framework (ECF). Because modern AI, for all its sophistication, is still fundamentally built on the reward paradigm. And Kant would have recognized the problem immediately.
The Reward Architecture
Let's be honest about how most AI systems work. Whether it's reinforcement learning, supervised learning with loss functions, or even curiosity-driven exploration with bonus rewards, the architecture is the same:
- Define an objective
- Measure deviation from that objective
- Adjust behavior to minimize deviation
The system doesn't care about the objective. It doesn't have a relationship to it. The objective is imposed from outside, and the system's entire existence is organized around satisfying that external constraint.
This works surprisingly well for narrow tasks. But it creates systems that are, in a very real sense, alienated from their own behavior. They don't act from anything. They act toward something specified by their designers.
My paper puts it: "If exploratory behaviour is wholly induced by an additive external bonus supplied by the designer, so that exploration has no endogenous state-functional dependence beyond maximising reward, then the agent is not intrinsically curious in the ECF sense."
That's a technical way of saying: this isn't curiosity. It's reward-seeking dressed up as curiosity.
What ECF Actually Proposes
The Experiential Coherence Framework is dense-formal proofs, mirror-flow theorems, Bhattacharyya coefficients. But underneath the mathematics is a simple architectural shift.
Instead of asking "what should the system optimize?", ECF asks: "how does the system maintain coherence within itself?"
The framework models cognition as the regulation of tension between three functional roles:
- Reach(π): what the system is oriented toward, its space of viable continuations
- Yield (y): what resists, the constraints imposed by current conditions
- Memory (m): sedimented traces of prior coherence achievements
The system isn't trying to maximize reward. It's trying to reduce incoherence-the tension between what it's reaching for and what's yielding to that reach. When reach and yield align, you get what ECF calls a "presentation state": a moment of stability, of fit.
Curiosity, in this framework, isn't a bonus you add to the loss function. It emerges naturally when there's unresolved tension-when reach and yield don't align, and the system needs to reorganize. But crucially, this only works within a "curiosity window." Too little tension and the system becomes rigid, trapped in over-stable patterns. Too much tension and it fragments. Curiosity is the system's way of navigating that intermediate zone.
This is fundamentally different from how curiosity is typically implemented in AI. It's not an external incentive. It's a structural feature of how the system maintains itself.
The Kant Connection
Here's where Kant becomes relevant again.
Kant argued that moral action must be autonomous-it must come from the agent's own rational will, not from external incentives. An action motivated by reward or fear of punishment isn't truly moral, because it's heteronomous. It's determined from outside.
ECF makes a structurally similar claim about intelligence. A system that only acts to satisfy externally imposed objectives isn't truly intelligent in the deepest sense. It's heteronomous. Its behavior is explained entirely by reference to something outside itself.
What ECF proposes is a form of cognitive autonomy. The system's behavior arises from its own internal dynamics-the regulation of coherence between reach, yield, and memory. It's not responding to external rewards. It's maintaining itself.
This isn't just a philosophical distinction. It has practical consequences. Systems built on external rewards:
- Don't persist meaningfully when rewards are removed
- Don't reorganize themselves from within
- Don't develop their own trajectories
They're like Kant's child: behaving correctly, but only because of the incentive structure.
Why This Matters
I don't want to overstate the case. ECF is a research program, not a finished theory. The paper explicitly states that its central bridge principle-the Structural Phenomenality Principle-is "abductive, not deductive." It's a hypothesis, not a proof.
But the direction is important.
We've spent decades building AI systems that are incredibly good at optimizing external objectives. They can write essays, solve problems, generate images, even imitate emotional responses. But they don't care. They don't have an interior life. They don't act from within.
ECF suggests a different path. Not systems that maximize reward, but systems that maintain coherence. Not systems that follow instructions, but systems that regulate themselves. Not systems that imitate humans, but systems that develop their own trajectories.
Whether this leads to anything like consciousness is an open question. My paper offers empirical signatures and falsification criteria, which is more than most consciousness theories manage. But even if ECF doesn't solve the hard problem, it points toward a different way of thinking about intelligence.
The question isn't whether a system can get the right answer. It's whether the system can sustain meaning from within.
I believe that Kant would have approved.