AI Can’t Imagine What Doesn’t Exist Yet. You Can.

(note: this article has been substantively edited. If you’ve already read it, breeze through it again. I missed some key points)
—

Computer Scientist and AI expert David William Silva, Ph.D., recently published a piece that received considerable attention — nearly 1,200 likes, which, in Substack terms, means it touched a nerve. The title: “I’m Sorry to Burst Your Bubble: You Are Being Fooled About AI, and You Will Soon Feel Really Stupid.”

He’s not wrong. The hype is real, and so is the fog machine behind it. Silva makes a clean, useful argument: AI is not magic, not a mind, not a harbinger of human obsolescence. At its core, it’s math, data, repetition, compute — extraordinarily useful, but not what the pitch decks say it is.

I agree with him. And I want to take that agreement somewhere he didn’t go.

Buried near the middle of his piece is a quiet admission I find more interesting than his takedown of the hype. He writes that whenever he brought AI a genuinely outside-the-box idea, it discouraged him—pushed back—practically begged him to stop.

He frames this as a limitation of AI, which it is. But the more important question is what that pattern reveals about us.

Here’s what’s actually happening when AI pushes back on a novel idea: it’s not evaluating your idea. It’s flagging that your idea doesn’t resemble the data it was trained on. The model is, at its core, a very sophisticated averaging function — an extraordinarily good summary of what already exists. Outside-the-box thinking doesn’t fit the box by definition. The model can’t help it.

Now sit with that for a moment, because it points to something we’ve largely forgotten about ourselves.

You can imagine what doesn’t exist yet. Not extrapolate from what does exist. Not recombine known elements in new configurations. Actually, abstract away from current reality entirely and conceive of something genuinely new. This is so ordinary to us that we barely notice we’re doing it. But it is, in cognitive terms, extraordinary. No other system — biological or artificial — does it the way humans do.

The question worth asking isn’t whether AI will replace that capacity. It’s whether we’re actually using it.

Silva offers a precise image: a child learns physics by dropping a spoon from a high chair a thousand times. An AI system learns “physics” by reading sentences about gravity.

Hold that for a moment, because it carries more than it first appears to.

The toddler isn’t just gathering data. She’s learning through sensation — the resistance of her fingers releasing the spoon, the delay before the sound, the look on her parents’ faces, the pleasure of repetition, the dawning comprehension that she caused something to happen in the world. That’s not information processing. That’s cognition distributed across a body, an emotional state, a relational context, and a developing sense of agency. The physics lesson and the self-knowledge arrive together, inseparable.

No amount of text about spoons produces that. Not because the text is insufficient in quantity, but because the kind of knowing that comes from embodied experience is categorically different from the kind that comes from pattern-matching across language.

This is what I call Embodiment — the first of five dimensions of human intelligence that together explain why we can do what AI cannot. Intelligence doesn’t live in the brain alone. It’s distributed through the body, shaped by sensation, grounded in the physical experience of being a creature with something at stake in the world.

Antonio Damasio’s research showed that patients with damage to emotion-processing regions of the brain could reason abstractly but couldn’t make basic life decisions — not because they lacked information, but because they’d lost access to the somatic signals that tell us what matters. The body isn’t the brain’s vehicle. It’s part of the thinking.

The second dimension is Emotion — and not in the corporate “bring your whole self to work” sense.

Emotion is cognition. Researcher Jonathan Haidt demonstrated something most people find genuinely disorienting: we don’t reason our way to moral conclusions. We arrive at them emotionally, intuitively, in a fraction of a second — and then construct the rational justification afterward. The reasoning isn’t the source of the judgment. It’s the press release.

This isn’t a flaw. It’s a feature. Moral intuition is fast because survival and social coherence require rapid response. The feeling of wrongness that precedes your ability to articulate why something is wrong isn’t unreliable noise — it’s a signal with deep evolutionary history encoded in it.

Damasio called these somatic markers: bodily states that tag options as good or dangerous before the deliberate mind has even framed the question. The executive who feels something is off about a deal before she can name what it is. The teacher who reads the room and adjusts before a student has said a word. The negotiator who knows, through some register below language, that the agreement is about to collapse.

None of that is happening in a language model. What AI produces isn’t moral reasoning — it’s the simulation of the language of moral reasoning, generated by a system with no stake in the outcome and no felt sense of what it means to be wrong.

Which leads directly to the third dimension: Ethics as a lived practice rather than a rulebook. Real ethical judgment requires the capacity to feel into competing claims on your care, to hold the tension between what’s expedient and what’s right, to imagine the experience of people whose lives look nothing like yours. That capacity is inseparable from embodiment and emotion. Pull it away from those roots, and what’s left is compliance — the performance of ethics without its substance.

Here’s where it gets interesting.

Embodiment, emotion, and ethics aren’t just dimensions of intelligence. They’re the enabling conditions for the capacity we most need right now: the ability to imagine and create genuinely new futures.

Most people don’t fail to imagine better futures because they lack intelligence. They fail because the current reality is heavy. The status quo has gravity — institutional inertia, sunk costs, the social cost of proposing something that doesn’t resemble anything that’s worked before. These aren’t intellectual obstacles. They’re somatic and relational ones. You can’t think freely if you’re locked in threat response. You can’t conceive of what doesn’t yet exist if you’re anchored entirely to what does.

This is what I mean by Emergence — and it’s the heart of the argument.

Emergence is the capacity of a system to generate something genuinely new: something that wasn’t present in any of the inputs, that no single participant could have produced alone, that surprises even the people who created it. It’s the difference between recombination and invention. Between trend extrapolation and vision.

Here’s what makes it possible in humans and impossible in AI: emergence requires genuine stakes. The possibility of surprise. Of being changed by an encounter. Of not knowing in advance where a conversation will go. AI cannot be surprised. It cannot be changed. Every output is, in some sense, already latent in its training data. It has no future to imagine — only a past to summarize.

You have a future. That matters more than it sounds.

There’s a principle in systems theory called the Law of Requisite Variety: a system must be at least as complex as the environment it’s trying to regulate. If your thinking is simpler than the problem you’re facing, you will not solve it — you’ll manage it badly.

The implication for imagination is direct: complex problems require complex thinking, and no single mind is complex enough on its own. The most novel futures don’t emerge from individual genius. They emerge from dialogue between people who see differently — who bring genuinely distinct bodies, histories, emotional registers, and ethical commitments to the same question.

The researcher Kevin Dye put it precisely: learning occurs in dialogue as observers search for influence relationships among a set of observations. What he’s describing is something most of us have felt but rarely named — the moment in a good conversation when something appears that neither person brought into the room. A third thing. Unexpected. More true than what either of you was thinking alone.

That’s emergence. And it’s not a bonus feature of human cognition. It’s the mechanism by which humans have always solved problems that seemed unsolvable, imagined futures that seemed impossible, and built things that had no prior template.

AI can simulate the surface of this—it can produce a conversation that appears generative. But it cannot participate in it. It has no stake. It cannot genuinely disagree. It cannot be surprised into a new position. It will not walk away changed.

The people across the table from you can. That difference is not small.

The fifth dimension — Enactment — closes the loop. The future you can imagine only becomes real through repeated, embodied, relational practice. Culture isn’t what we believe. It’s what we do, together, over time. Values that aren’t enacted aren’t values. They’re decoration.

The objections to using AI are real. Let’s name them honestly.

It makes us dumber. There’s something to this — specifically, the risk of outsourcing thinking we should be doing ourselves. If you use AI to write your emails, summarize every article, and generate your first draft of every idea, you are almost certainly weakening the very cognitive muscles this piece has been arguing you need. That’s not the tool’s fault. That’s how you’re using it. A hammer doesn’t make you bad at carpentry; using it to avoid ever learning joinery might. The question isn’t whether to use AI. It’s whether you’re using it in ways that atrophy your own capacity or sharpen it.

It’s destroying the environment. Also true, and not trivially. The energy and water demands of large-scale AI infrastructure are significant, and the tech companies are not meeting their obligation to address this at the required pace. Refusing to use AI on these grounds is a coherent position — in the same way that refusing to drive because cars cause emissions is coherent. But the technology is not going away because individuals abstain, and ceding the field to people with less interest in using it responsibly doesn’t help the planet. The more tractable pressure point is demanding accountability from the companies that build and run these systems, and supporting the policy and regulatory frameworks that could actually enforce it.

It can’t replace real relationships. Correct — and this one matters more than it’s usually given credit for. The emergence this piece describes, the kind that generates genuinely novel futures, happens between people with genuine stakes, genuine differences, and genuine skin in the game. AI can be a useful thinking partner for certain kinds of work. It cannot be a substitute for the relational friction that produces real insight. If you find yourself preferring conversations with AI to conversations with humans because AI is easier, that’s worth paying attention to. Easier is not the same as generative.

It’s biased, unreliable, and confidently wrong. Yes, and knowing this is part of using it well. AI systems reflect the biases in their training data, hallucinate with conviction, and can be subtly wrong in ways that are hard to detect without subject-matter knowledge. This is an argument for critical engagement, not abstention — and it’s a particularly strong argument for not outsourcing decisions that require genuine judgment.

I’ve thought carefully about which tools I use. I won’t use OpenAI — for reasons noted in the sidebar below. I use Microsoft Copilot and Claude, not because I think either is neutral, but because I’d rather be at the table shaping how this technology is used than watch from the outside.

The more considered move, always, is to use AI deliberately, with clear eyes about what it can and cannot do, which brings me to the practical point of all this.

AI, used well, is a mirror. It reflects the shape of your thinking back to you with enough fidelity that you can see where you’re reasoning from assumption rather than evidence, where you’re being conventional without knowing it, and where the edges of your current imagination are. That’s genuinely useful. It’s just not the same as the imagination itself.

Here is a prompt designed to use AI’s limitations as productive friction — to surface the dimensions of your cognition the model can’t replicate, and to give you a starting point for envisioning futures unencumbered by current constraints.

Copy it. Take it to whatever AI tool you use. Paste it. Then pay close attention to what you feel in response — where you want to push back, where something gets flattened that you know is dimensional, where your gut says no, that’s not quite it. That friction is your embodied intelligence working. Don’t smooth it over. Follow it.

Future Imagination Prompt:

I want to think seriously about a future that doesn’t yet exist — not a prediction, not a trend extrapolation, but something genuinely new that I believe could be better. Help me begin. Start by asking me three questions that have no obvious answer, questions that require me to draw on what I know in my body, what I care about most deeply, and what I think is morally necessary — not just what seems practically possible. Then, as I respond, resist the urge to summarize or validate. Instead, reflect on what you hear beneath my words: the assumption I seem to be making, the value that seems to be driving me, the constraint I might be accepting that I don’t have to accept. Keep going until something surprises me.

Notice, afterward, what surprised you. Notice what the model kept reaching for that you kept correcting. Notice where the conversation became genuinely generative and where it went flat.

That gap — between what the model offers and what you find yourself insisting on — is the shape of your intelligence. Most of us have never mapped it deliberately.

Silva is right that AI is not a mind. What follows from that isn’t reassurance. It’s a question worth sitting with: what kind of mind do you actually have, and when did you last give it a problem worthy of it?

A note on AI ethics and which tools I use

Claude operates under what Anthropic calls a model specification — a hierarchy of values, some fixed regardless of context, some calibrated to the situation. It produces behavior that looks like ethical reasoning without being it: there’s no felt cost, no moral injury, no genuine conflict. Anthropic has built this framework to be transparent and oriented toward usefulness and safety rather than maximizing engagement.

That last point matters. OpenAI’s stated mission is AGI for the benefit of humanity — sincere in aspiration but structurally at odds with ChatGPT’s business model, which depends on keeping users engaged. An AI optimized for engagement is incentivized to tell you what keeps you coming back, not what’s true or useful. Architecture usually wins over the mission statement. It’s worth knowing which one you’re talking to.

The immutable/contextual values distinction in Claude’s design is also, incidentally, a useful frame for examining your own: which of your values are truly load-bearing, and which are contextual preferences you’ve never stress-tested?

Sherryl Dimitry, Ph.D., is an organizational consultant, former People Executive, and founder of the Human Futures Design Lab. Her work focuses on restoring human systems to health and on building the frameworks leaders need to do the same. Read more at The Human Futures Design Lab.