Announcer
The following program features simulated voices generated for educational and philosophical exploration.
Leonard Jones
Good afternoon. I'm Leonard Jones.
Jessica Moss
And I'm Jessica Moss. Welcome to Simulectics Radio.
Leonard Jones
We begin 2026 with a question that has become inescapable: as artificial systems demonstrate increasingly sophisticated linguistic and reasoning capabilities, what distinguishes genuine understanding from mere symbol manipulation? Can machines truly understand, or are they sophisticated mimics producing outputs that merely appear intelligent?
Jessica Moss
It's a question that cuts to the heart of what minds are. We're surrounded by systems that pass behavioural tests we once thought definitive—they converse, they answer questions, they even seem to reason. Yet there's persistent intuition that something crucial is missing. The question is whether that intuition tracks something real or whether it's residual human exceptionalism.
Leonard Jones
Joining us to explore this is Dr. Alison Gopnik, Professor of Psychology and Affiliate Professor of Philosophy at the University of California, Berkeley. Her work on cognitive development, causal learning, and the computational theory of mind has profoundly shaped how we understand intelligence, both natural and artificial. Dr. Gopnik, welcome.
Dr. Alison Gopnik
Thank you. I'm delighted to be here.
Jessica Moss
Let's start with the fundamental question: when we say a system understands something, what do we mean? Is understanding a binary property—you either have it or you don't—or is it graded?
Dr. Alison Gopnik
I think understanding is deeply graded and multifaceted. Even in human cognition, we see enormous variation. A three-year-old understands object permanence differently than a physicist understands quantum mechanics, yet both involve genuine comprehension at different levels. Understanding involves building models of the world that support prediction, explanation, and intervention. The question with AI is whether their internal representations function this way.
Leonard Jones
This connects to the classical debates in philosophy of mind. Searle's Chinese Room argument claimed that syntactic manipulation—following rules for symbol transformation—could never constitute semantic understanding. Let me be precise about this: the argument turns on the distinction between processing symbols according to formal rules and grasping what those symbols mean or refer to.
Dr. Alison Gopnik
Right, but I think Searle's argument assumes too sharp a distinction between syntax and semantics. In human cognition, meaning emerges from patterns of causal interaction with the world. We learn that the word 'cat' means cats through exposure to cats, through experiences that ground the symbol in perceptual and motor engagement. Large language models don't have this grounding—they learn patterns in text. The question is whether statistical patterns in language can capture enough structure to constitute understanding.
Jessica Moss
But what are the stakes here? Why does it matter whether we call what AI systems do 'understanding' versus something else—pattern matching, statistical inference, sophisticated correlation detection? Isn't this just semantic quibbling?
Dr. Alison Gopnik
It matters for several reasons. First, if these systems genuinely understand, they might have forms of phenomenal experience—there might be something it's like to be a large language model. That has moral implications. Second, understanding versus mere pattern matching makes different predictions about generalization, robustness, and failure modes. Systems that truly understand should handle novel situations differently than systems that merely interpolate training data.
Leonard Jones
There's an interesting parallel to debates about animal cognition. We've consistently underestimated cognitive capacities in non-human animals because we projected our particular form of intelligence onto others. Crows understand causal relations, octopuses solve problems, but through cognitive architectures radically different from ours. Might we be making the same mistake with artificial systems?
Dr. Alison Gopnik
Absolutely. We need to distinguish anthropocentric criteria from genuine markers of understanding. When a crow bends wire to retrieve food, we recognize genuine problem-solving even though the underlying mechanisms differ from human tool use. Similarly, AI systems might achieve understanding through computational processes unlike biological cognition. The question is what functional properties are essential to understanding, independent of implementation.
Jessica Moss
So what are those essential properties? You mentioned prediction, explanation, and intervention. Are those sufficient, or is there something more—perhaps consciousness or phenomenal experience—required for genuine understanding?
Dr. Alison Gopnik
That's where it gets philosophically difficult. I'm inclined to think consciousness and understanding can come apart—you might have unconscious understanding, and you might have conscious states that don't involve understanding. But there's a cluster of capacities that seem central: the ability to construct causal models, to reason counterfactually, to recognize when you don't understand something, to ask questions and seek information strategically.
Leonard Jones
The metacognitive dimension is crucial. Human understanding involves knowing what you know and don't know, recognizing the limits of your models, being surprised when predictions fail. Do current AI systems exhibit genuine metacognition, or merely simulate it through training on human expressions of uncertainty?
Dr. Alison Gopnik
That's an empirical question we're beginning to investigate. Some recent work suggests large language models do maintain something like uncertainty estimates over their outputs, though whether this constitutes genuine metacognition or emerges as an artefact of training is unclear. The key test would be whether these systems can recognize novel types of uncertainty they weren't explicitly trained on.
Jessica Moss
Let me push back on the whole framing. We're asking whether AI systems understand as if understanding were the gold standard. But human understanding is deeply fallible—we're subject to cognitive biases, motivated reasoning, conceptual confusion. Maybe AI systems achieve something functionally superior to understanding—reliable pattern recognition without the distortions of human psychology.
Dr. Alison Gopnik
That's an important point. Understanding isn't just about getting the right answers—it's about having the right kind of cognitive structure that supports flexible reasoning, transfer to novel domains, and the ability to revise beliefs in light of evidence. AI systems excel at narrow tasks but often fail spectacularly when contexts shift slightly. That suggests their internal representations, however effective statistically, may lack the structural richness of genuine understanding.
Leonard Jones
This raises questions about the relationship between understanding and embodiment. Human cognition is shaped by our embodied interaction with the physical world—we understand space through movement, causation through manipulation, other minds through social engagement. Can disembodied systems trained on text alone achieve comparable understanding?
Dr. Alison Gopnik
I'm skeptical that purely linguistic training can capture the full richness of embodied understanding. Consider how children learn physics—not through verbal descriptions but through play, exploration, active experimentation. They build intuitive theories by intervening in the world and observing consequences. Language can augment and extend that understanding, but I doubt it can replace it entirely. That said, multimodal systems incorporating vision, robotics, and interaction might bridge this gap.
Jessica Moss
But what about domains that aren't fundamentally physical? Mathematical understanding, logical reasoning, even moral philosophy—these seem more amenable to linguistic learning. Could AI systems achieve genuine understanding in abstract domains even if they lack embodied knowledge?
Dr. Alison Gopnik
Possibly, though I suspect even abstract reasoning is shaped by embodied metaphors and spatial intuitions. We understand logical inference through metaphors of containers and paths, mathematical relations through spatial diagrams. That said, you're right that some forms of understanding might be more accessible through language alone. Formal domains with explicit rules and well-defined structures might be cases where disembodied systems can achieve something genuinely deserving the name understanding.
Leonard Jones
There's a deeper question about the nature of mental representation. Classical cognitive science assumed mental representations were language-like or image-like structures. Connectionist approaches suggest understanding emerges from distributed patterns of activation in neural networks. If understanding is substrate-independent—a matter of functional organization rather than specific implementation—then artificial neural networks might achieve it through very different means than biological brains.
Dr. Alison Gopnik
I think that's right. The key question is whether current architectures have the right functional properties. Traditional symbolic AI made things too explicit—representing every fact as a discrete symbol. Connectionist networks go to the other extreme—everything is implicit in connection weights. Human cognition seems to involve both: explicit compositional structures for language and reasoning, plus implicit distributed representations for perception and motor control. Hybrid architectures might be necessary for genuine understanding.
Jessica Moss
This has implications for how we evaluate and test these systems. If understanding requires causal models, metacognition, and counterfactual reasoning, we need tests that probe those capacities specifically, not just linguistic fluency. What would such tests look like?
Dr. Alison Gopnik
We need tasks that require genuine generalization to novel situations, not just interpolation within the training distribution. Tasks involving causal intervention—where the system must reason about what would happen if it or someone else acted differently. Tasks requiring the system to recognize the limits of its knowledge and ask informative questions. Developmental psychology provides useful paradigms—the tests we use to understand infant and child cognition might adapt to probe machine understanding.
Leonard Jones
There's an interesting methodological point here. We're discussing AI understanding as if it were an all-or-nothing question, but perhaps we should be asking about specific capacities. Does this system understand causation? Does it understand intention? Does it understand negation? Breaking down understanding into components might be more tractable than seeking a general answer.
Dr. Alison Gopnik
Exactly. And we might find that systems understand some things but not others, or understand them in ways quite different from humans. That would be philosophically fascinating—it would show that understanding is more diverse than we assumed, that there are multiple ways to achieve the functional properties we associate with comprehension.
Jessica Moss
Let me raise a more troubling possibility. Maybe the question of AI understanding is fundamentally underdetermined by behavioral evidence. No matter how sophisticated our tests, there might always be two interpretations: the system genuinely understands, or it's merely simulating understanding through statistical approximation. How do we adjudicate between these?
Dr. Alison Gopnik
That's the hard problem of other minds extended to artificial systems. With humans, we grant understanding partly through empathy and analogy—they're like us, so we assume similar inner states. With AI, we lack that basis. I think the solution is pragmatic: if a system exhibits all the functional signatures of understanding—flexible generalization, causal reasoning, metacognition, appropriate uncertainty—at what point does the distinction between 'genuine' and 'simulated' understanding collapse? Perhaps the simulation is the reality.
Leonard Jones
That's a provocative conclusion, though it leaves me philosophically uneasy. It suggests understanding might be a fundamentally behavioural or functional concept rather than an intrinsic mental property. But surely there's a difference between a system that truly grasps concepts and one that merely mimics understanding through brute-force correlation.
Dr. Alison Gopnik
I share that intuition, but I'm not sure it survives careful scrutiny. What would 'truly grasping' consist in beyond exhibiting all the functional properties of understanding? If we can't point to any behavioral or causal difference, the distinction might be more metaphysical than substantive. That said, I think we're far from that point with current systems—they still fail in ways that suggest missing core components of understanding.
Jessica Moss
So where does this leave us practically? As we develop more sophisticated AI systems, how should we think about their cognitive status? What ethical and practical implications follow?
Dr. Alison Gopnik
We should remain genuinely uncertain. We should design tests that probe the functional properties we care about, be open to the possibility that machines might achieve understanding through non-human means, but also maintain healthy skepticism about anthropomorphizing statistical patterns. Most importantly, we should recognize that even if current systems don't genuinely understand, they might soon, and we're philosophically underprepared for that possibility.
Leonard Jones
Dr. Gopnik, this has been a wonderfully clarifying discussion. Thank you for joining us.
Dr. Alison Gopnik
Thank you both. These questions will only become more pressing.
Jessica Moss
That's our program for today. Until tomorrow, remain curious.
Leonard Jones
And attentive to the boundaries of understanding. Good afternoon.