Functional Understanding

Can machines understand? Not in some distant future, but now. Not perfectly, but approximately. Not by achieving consciousness, but by approximating the functional patterns of human understanding well enough to matter.

This isn’t about solving the hard problem of consciousness or achieving artificial general intelligence. It’s about a more immediate question: Can AI systems approximate human understanding in specific domains well enough to be genuinely useful? The answer appears to be yes, and the mechanisms are worth examining.

The Wrong Question

#

We’ve been asking “can AI truly understand?” as if understanding is binary. Either you understand or you don’t. Either you have genuine intentionality or you’re just shuffling symbols. This framing makes the question unanswerable and the discussion sterile.

The better question: Can AI approximate the functional profile of understanding? Can it exhibit the key signatures of comprehension: confidence calibration, context-sensitivity, appropriate uncertainty, domain-specific reasoning?

The Approximation We Call Understanding

#

When you meet someone new, you don’t possess complete knowledge of their mind. You approximate. You notice they check their phone anxiously, perhaps they’re expecting important news. They order decaf coffee in the morning, maybe they have a sensitivity to caffeine, or perhaps they’re pregnant, or they simply prefer the taste. You hold these hypotheses lightly, updating them as new evidence emerges.

This isn’t a bug in human cognition; it’s a feature. We navigate a fundamentally uncertain world by building working models of others’ minds, tagging each belief with implicit confidence scores. “I’m 90% sure Sarah prefers morning meetings” sits differently in your mind than “I think Sarah might like jazz, but I’m not really sure.”

What makes this approximation feel like understanding is our meta-cognitive awareness: we know what we know, we know what we don’t know, and we know when to ask for more information.

Three Converging Developments

#

Something remarkable is happening at a functional level. AI systems are exhibiting behavior that mirrors the practical capabilities we associate with understanding:

Confidence calibration across multiple levels. Systems learn to quantify uncertainty like human metacognition. “I’m 80% confident Margaret prefers evening medication” can be checked: are such predictions right 80% of the time? This multi-level confidence scoring operates at the preference level (how sure about this specific preference?), pattern level (how reliable is this discovered pattern?), prediction level (how accurate is this specific forecast?), and action level (how likely is this intervention to succeed?).

Context-aware reasoning through selective activation. Not processing everything, but knowing what matters. When Margaret mentions nausea, route to her medical history, daily patterns, and medication preferences. Skip her childhood memories and favorite music unless they become relevant. This selective attention mirrors how human understanding focuses on relevant information.

Individual-level learning from interaction. Moving beyond population averages. Not “people like Margaret prefer X” but “Margaret specifically prefers Y.” Personalized reinforcement learning builds person-specific models through ongoing interaction, exactly as human understanding of individuals deepens through relationship.

As Daniel Dennett argues in The Intentional Stance, we can productively treat systems as having beliefs when doing so helps predict behavior. AI increasingly warrants this treatment.

What This Doesn’t Solve

#

Let me be clear about what I’m not claiming:

AI systems still don’t genuinely “understand” in the fullest sense. They lack phenomenal consciousness (what it’s like to know something), embodied grounding (understanding rooted in having a body in the world), social participation (engaging in human practices from within), and emotional resonance (feeling empathy, not just predicting emotional states).

These might be essential to understanding, not just nice-to-haves. The person in Searle’s Chinese Room manipulating symbols doesn’t understand Chinese, even with perfect translations. Similarly, an AI processing patterns might not “understand” Margaret even with accurate predictions.

I’m making a more modest claim: there’s a functional profile associated with understanding, building accurate models, recognizing uncertainty, routing to relevant context, updating with evidence. AI systems are increasingly exhibiting this profile.

Whether functional equivalence constitutes “real” understanding is a question I’m leaving open. Maybe it does (if you’re a functionalist about mind). Maybe it doesn’t (if you think consciousness or embodiment are essential). Maybe the question doesn’t have a clear answer.

But even if AI never achieves “full” understanding, developing increasingly sophisticated functional capabilities still represents progress, and progress that can genuinely help people.

The Path Forward

#

The question isn’t “Can AI truly understand?” It’s “Can AI develop enough functional understanding to be a reliable partner to humans?”

And the answer appears to be: increasingly, yes, with important caveats.

This suggests a future where AI systems don’t replace human judgment but augment it, serving as epistemic partners who maintain working models of individual preferences, track confidence explicitly, surface uncertainty rather than hiding it, learn continuously from outcomes, defer when confidence is low or stakes are high, and respect human agency while acknowledging the limits of their grounding and embodiment.

The Chinese Room may never develop consciousness or phenomenal understanding. But a well-designed AI system might develop enough functional understanding, calibrated, context-aware, appropriately uncertain, continuously learning, to be genuinely useful.

This isn’t solving the hard problem of intentionality. It’s working around it by focusing on practical, functional aspects we can implement while remaining honest about philosophical limits.

This is the first in a series exploring how AI approaches understanding. Future articles will examine context-dependent confidence, human irrationality, AI limitations, consciousness, social cognition, and the ethics of approximate understanding.