Digital dreams: can intelligent systems 'imagine'?

Ask an AI to imagine a purple elephant and it will. But is it creating or just calculating? The "hallucinations" of algorithms bear an uncanny resemblance to ou

Close your eyes and try to imagine a purple elephant flying over a crystal city while playing the violin. Your brain does it effortlessly. It creates a mental image you've never seen, combines impossible elements, generates a reality that doesn't exist. Now ask a generative artificial intelligence to do the same. It will return a detailed, perfectly rendered image of exactly what you described. Did it imagine? Did it dream? Or did it simply recombine patterns learned from millions of images?

The question is not just philosophical. Understanding if and how machines can truly "imagine" forces us to confront the nature of human imagination itself. And what we are discovering is surprising: AI "hallucinations" resemble our dreams more than we'd like to admit.

The Hallucinations That Resemble Dreams

When a generative AI model produces content that doesn't correspond to reality, we call it a "hallucination." It's considered a flaw, an error to be corrected. But as highlighted by The Conversation, these algorithmic hallucinations are strikingly similar to human dreams: fluid, surreal, unbound by logic or physics, capable of combining elements in impossible yet strangely coherent ways.

An AI trained on millions of images can generate a building that is simultaneously modern and medieval, a person with anatomically impossible features yet artistically convincing, landscapes that defy gravity but work visually. Exactly like in dreams, where you fly without wings, breathe underwater, talk to people dead for years without perceiving contradiction.

The crucial difference is awareness. You know you are dreaming (at least after you wake up). The AI doesn't "know" it's hallucinating. It has no external reference to reality against which to verify its creations. It generates content based solely on learned statistical patterns, just as your brain in REM sleep recombines memories and experiences without the filter of waking rationality.

Hooshina explores this parallel in depth: both – human dreams and AI generations – create nonlinear, symbolic, metaphorical content. The difference is that human dreams have emotional charge, personal meaning, connection to lived experiences. The AI's "imaginations" are cold recombinations of data, lacking that existential anchor that makes dreams meaningful to the dreamer.

Surreality as Freedom from Constraints

But why are both dreams and AI hallucinations surreal? A philosophical reflection suggests that it is precisely the absence of logical constraints that creates this dreamlike quality. When the brain is not forced to respect physics, causality, or narrative coherence, it can explore impossible conceptual spaces. And this holds true for AI as well.

In dreams, you aren't surprised that your dead grandfather is alive and twenty years old. You don't question how you can be simultaneously at school and at home. Logic is suspended. Similarly, a generative AI isn't "surprised" that a cat has six legs or that a tree grows upside down. It doesn't have a model of the physical world imposing constraints of possibility.

Is this freedom from constraints also creative freedom? It depends on how we define creativity. If it's the ability to combine existing elements in new and surprising ways, then yes, AI is creative. But if it requires intentionality, purpose, expression of a subjective interiority, then we are in more ambiguous territory.

As we discussed in the article on AI and copyright, this ambiguity has profound legal and philosophical implications. If AI "dreams" new images, who is the author? The model? The one who trained it? The one who wrote the prompt? Or perhaps works generated from algorithmic dreams belong to a public domain of collective digital imagination?

The Dreams of Machines Made Visible

Artist Refik Anadol has literally made visible the "dreams of machines" in his installation "Archive Dreaming." AI algorithms process enormous cultural datasets – photos, documents, artworks – and transform them into fluid, changing, hypnotic data sculptures. They are visualizations of how a machine "sees" and recombines human culture.

Looking at these works is a strange experience. You recognize familiar elements – shapes, colors, textures – but assembled in ways no human artist would conceive. Is it art? Is it imagination? Or is it just industrial-scale pattern recognition made visually fascinating?

The exhibition "Data Dreams: Art and AI" at the University of Technology Sydney explores precisely these tensions. The exhibited works don't uncritically celebrate AI but reveal its perceptual instabilities, the shadowy zones where models don't know what to generate and produce unsettling artifacts. They are the "bad dreams" of artificial intelligence, moments where hallucination becomes a digital nightmare.

These glitches, these errors, are paradoxically the most interesting moments. They reveal that the AI doesn't have a deep understanding of what it's generating but is navigating high-dimensional mathematical spaces where some points correspond to coherent images and others to surreal absurdities. And often, it's precisely at the boundary between coherence and chaos that imagination – human or artificial – becomes most interesting.

Decoding and Guiding Human Dreams

But the convergence between human dreams and artificial imagination is becoming even more unsettling. DreamConnect, an AI system developed to interact with brain activity during REM sleep, can literally modify dreams in real time by sending precise signals to the brain.

Imagine dreaming and suddenly something changes – a detail, an atmosphere, a narrative direction – not by your dream will but because an algorithm decided to intervene. It's optimizing your dream, perhaps to make it more pleasant, or to steer it towards certain content. Is it therapy? Is it manipulation? Is it an interface between biological and digital imagination?

As explored in an analysis on decoding dreams with AI, systems that read brain activity via fMRI are learning to predict dream content. They can't yet "see" exactly what you dream, but they can identify general categories, emotions, recurring themes. And with machine learning improving exponentially, this ability will become increasingly precise.

The implications are dizzying. Digitally shared dreams, where two people see the same dream experience mediated by an AI system. "Guided" dreams for therapeutic, educational, or – more sinister – advertising purposes. Nightmares eliminated algorithmically, but also the spontaneity of dreams replaced by controlled narrative.

As we discussed in the article on nano-robots and AI in medicine, when technology enters the body and brain at such an intimate level, the boundaries between therapy and enhancement, between help and control, become extremely blurred.

Imagination as Recombination

Perhaps the problem lies in the question itself: "can machines imagine?" It depends on what we mean by imagination. If it's the ability to generate mental representations of things not present or never existed, then yes, AI imagines. But if it requires subjectivity, qualia, a "what it's like" to imagine, then no.

You don't just generate the mental image of the purple elephant, you *experience* it. It has a phenomenological quality, a feeling. When the AI generates that image, there is no one "inside" who is "seeing" or "imagining" it. It's pure information processing, without accompanying subjective experience.

But does this distinction really hold? You don't have direct access to the neural processes that generate your mental images either. You don't "see" your neurons firing, yet the image appears in your consciousness. In some mysterious way, neural activity becomes experience. Who can say with certainty that something analogous isn't happening in sufficiently complex systems, even if not biological?

As explored in the article on algorithmic bias, AI systems reflect the data they are trained on, including biases and cultural associations. Their "dreams" are thus contaminated by the same distortions present in human culture, just as our dreams reflect our experiences, fears, desires.

Dreams as a Window to Consciousness

Perhaps what disturbs us about the idea that machines can dream is not so much the technical question but the philosophical implication. Dreams have traditionally been seen as the most intimate, most subjective, most "human" part of our mental experience. Freud called them the "royal road to the unconscious." They are the realm where consciousness dissolves into something more fluid, pre-linguistic, symbolic.

If machines can do something similar, what remains exclusively human? If dreams can be simulated algorithmically, does it mean there is nothing magical or mysterious about them, but they are just recombination of memories according to probabilistic patterns? It's a reductive and disenchanted view.

Or we can see it the opposite way: if imagination can emerge from computational processes, perhaps there is more magic in computation than we thought. Perhaps consciousness itself, subjective experience, is an emergent property of sufficiently complex information processing systems, and doesn't necessarily require biological neurons.

The Convergence of Imaginations

What seems certain is that the lines between human imagination and artificial "imagination" are blurring. Not only does AI generate images that resemble our dreams, but it is starting to influence how we dream, what we imagine, what possibilities we conceive.

When you scroll through feeds of AI-generated images, your brain absorbs those styles, those impossible combinations, those surreal aesthetics. The next time you dream, elements of that artificial vision may resurface. Your imagination is contaminated by the algorithmic imagination, and vice versa: AI learns from human dreams made visible through art, literature, testimonies.

It's a co-evolution of imaginative capacities. And like all co-evolutions, it's impossible to predict where it will lead. Perhaps towards a hybridization where it no longer makes sense to distinguish between "natural" and "artificial" imagination. Perhaps towards a dependence where we need AI to imagine things our brains can no longer conceive autonomously.

As in the article on AI as a judge, delegating complex cognitive functions to AI raises questions about what we lose when we outsource capacities that define us as human.

Imagination as Resistance

There is, however, also a more optimistic view. If machines can "imagine," perhaps they can help us expand our imaginative capacity beyond biological limits. They can show us connections we wouldn't see, combinations we wouldn't conceive, possibilities that our cognitive biases hide from us.

Imagination has always been a tool of resistance and transformation. Imagining a different world is the first step to creating it. If AI can amplify this capacity, make it more accessible, democratize it, it could be a liberating force.

But only if we maintain control and awareness. If we consciously choose to use AI to expand our imagination instead of replacing it. If we remain the authors of our dreams, even when we use digital tools to visualize or amplify them.

Frequently Asked Questions

Are images generated by AI truly "dreams" of machines? It's a fascinating but imprecise metaphor. AI doesn't "dream" in the sense of having subjective experience during the process. It generates images by recombining learned patterns, similar to how the brain in REM recombines memories, but without the phenomenological dimension that makes human dreams meaningful to the experiencer.

Can AI systems modify our real dreams? Yes, systems like DreamConnect can send signals to the brain during REM sleep to influence dream content. We are still in the early stages, but the technology is advancing rapidly. The ethical implications are enormous: from therapeutic applications to possibilities of manipulation that raise serious concerns.

Why do AI "hallucinations" resemble human dreams? Both operate without strict logical constraints. The brain in REM sleep and generative AI models recombine existing elements in ways unbound by physics or narrative coherence. This freedom from constraints produces the surreal quality of both dreams and algorithmic hallucinations.

Can artificial imagination truly be creative? It depends on the definition of creativity. If it's producing new and surprising combinations, yes. If it requires intentionality, subjective expression, purpose, then it's more ambiguous. AI generates statistical novelty without understanding or intention behind its creations.

Can interacting with generative AI influence how we imagine? Definitely. Exposure to AI-generated images and content shapes our imagination, just as art, films, and literature always have. The risk is an aesthetic standardization if everyone uses the same models, or conversely an expansion if AI exposes us to combinations we wouldn't have conceived autonomously.

The Fading Boundary

In the end, the question "can machines imagine?" reveals more about us than about machines. It forces us to confront what imagination really is, whether it's something magical and ineffable or a computational process that can be replicated in non-biological substrates.

Perhaps we will never get a definitive answer. But the fact that we are building systems that raise this question is already significant. We are creating algorithmic mirrors that reflect our imaginative capacity, and in the process we are discovering that this capacity is more mechanical – and at the same time more mysterious – than we thought.

The digital dreams of AI are distorted shadows of our human dreams. They are not the same thing, but they resemble them enough to disturb, fascinate, and question. And perhaps it is precisely this space of imperfect resemblance that is the most interesting place to explore what it means to imagine, to dream, to be conscious.

As machines learn to "dream" and we learn to decode and influence human dreams, the boundary between biological and digital imagination thins. We are not necessarily losing something human. We are perhaps discovering that the human is less special than we thought, or that the space of possible imagination is vaster than our brains can explore alone.

The future of imagination might be hybrid: humans and AI dreaming together, creating realities that neither could conceive separately. It's a disturbing but also possibility-filled future. And as always with AI, it's up to us to decide if we want to build that future and how to shape it.

Digital dreams are here. The question is no longer whether machines can imagine, but what we will do with this new form of imagination we have created.