Dreams Without a Dreamer

Ask whether an AI can dream and you are asking two questions at once. One is about what happens inside a machine when nobody is giving it input: does it generate images, scenes, rehearsals of what it has learned? The other is whether anyone is there to have them. The first question has real answers, because researchers have built machine processes and, in two cases, called them dreams themselves. The second is the question of machine experience, and it decides whether any of those processes is a dream in the sense you mean when you say you had one last night.

This essay takes the questions in that order. First, what dreaming is on the account a standard reference gives. Then four machine processes, each from its own paper or post: what the system does, which part of dreaming it shares, and which it does not. Then the strongest objection, which is about experience, and finally how a chatbot’s “hallucination” differs from a dream, since the two words are often run together. It does not repeat the case on machine consciousness, which is made in Fourteen Questions for a Chatbot.

How the sources were handled. Every text below was opened with a public GET on 10 October 2026 and quoted from the words as served: the Stanford Encyclopedia of Philosophy entry Dreams and Dreaming by Jennifer M. Windt (revised 26 January 2026); Google’s 2015 research blog post on Inceptionism; David Ha and Jürgen Schmidhuber’s World Models (2018), in its arXiv abstract and its interactive web version; Tadros, Krishnan, Ramyaa and Bazhenov’s 2022 paper in Nature Communications; and Erik Hoel’s 2021 paper in Patterns, read in its PubMed Central copy. No chatbot was asked about its dreams, and no answer from one is used as evidence: a language model trained on human writing about dreams will describe dreams fluently whether or not anything happens to it at night.

What a dream is

The Stanford Encyclopedia’s entry is mostly about questions philosophers argue over, and its first lesson is that even the obvious answer, that a dream is an experience you have while asleep, was once denied. It uses conscious experience as an umbrella term for the sensations, thoughts and emotions in dreams, and says these are phenomenal states: there is something it is like to be in these states for the subject of experience.

The most influential denial came from Norman Malcolm in the 1950s, who argued that sleep and consciousness exclude each other. The entry quotes him: if a person is in any state of consciousness it logically follows that he is not sound asleep. For Malcolm, a dream report after waking was the only criterion of dreaming, and it did not show that anything had been experienced during the night. Daniel Dennett later proposed a “cassette theory”, on which dreams are inserted into memory at the moment of waking, so that there would be something it is like to remember a dream but nothing it is like to have one. That thought experiment matters for machines, and it comes back below.

Both denials lost. The entry’s summary of where the field stands is plain:

Contemporary philosophers tend to agree that dreams are experiences and this view is supported by empirical evidence from sleep physiologyJennifer M. Windt, “Dreams and Dreaming”, Stanford Encyclopedia of Philosophy (rev. 2026), §2.1

The evidence it lists has three strands. Reports of dreaming follow 81.9% of awakenings from REM sleep and 43% of awakenings from non-REM sleep (figures the entry takes from Nielsen 2000). Lucid dreamers in sleep laboratories have signalled, with pre-arranged left-right eye movements visible on the recording, that they are dreaming and know it, which tells against the cassette theory because the dream is shown to be happening while it is reported. And people with REM-sleep behaviour disorder act out dreams they later describe.

On what kind of experience a dream is, the entry calls one view standard:

The standard view is that dreams have the same phenomenal character as waking perception in that they seemingly put us in contact with mind-independent objects, yet no such object is being perceived.Jennifer M. Windt, “Dreams and Dreaming”, Stanford Encyclopedia of Philosophy (rev. 2026), §2.2.1

This means dreams count as hallucinations in the traditional philosophical sense, the entry adds; a rival view treats them as closer to imagining. On what dreams are for, it reports several theories without endorsing one: that they help process emotion, that they rehearse threats (Revonsuo’s threat simulation theory) or social life, that the sleep they happen in consolidates memory, and the possibility that they are by-products with no function of their own.

From that account you can pull out four parts of dreaming that a machine might or might not share:

1. Offline generation. Imagery produced from the inside, while input from the senses is shut off.
2. A simulated world. Not just pictures but a scene in which a self moves and acts.
3. A role in learning. Something done with what was learned while awake, on the theories that give dreams a function.
4. Experience. It is like something to be the dreamer. This is the part the philosophy now takes as settled for people.

The machine processes that get called dreams

DeepDream: a network asked for more of what it sees

In June 2015, three people at Google, Alexander Mordvintsev, Christopher Olah and Mike Tyka, published a post on Google’s research blog titled “Inceptionism: Going Deeper into Neural Networks”. Its purpose was to see what an image-recognition network had learned. They ran the network backwards: rather than ask what a picture showed, they asked it to change the picture so that it showed more of something. Sometimes they named the thing (a banana, a dumbbell); sometimes they let the network choose:

This creates a feedback loop: if a cloud looks a little bit like a bird, the network will make it look more like a bird.Alexander Mordvintsev, Christopher Olah and Mike Tyka, “Inceptionism: Going Deeper into Neural Networks”, Google Research blog, 2015

The instruction they gave the network, in their own paraphrase, was Whatever you see there, I want more of it! Run repeatedly, with a small zoom after each pass, and started from random noise rather than a photograph, the process produced images made entirely by the network. This is the technique now generally known as DeepDream. The post’s caption for those images reads Neural net “dreams”, with the word in quotation marks. The authors compared the process to a child finding shapes in clouds, and gave it a working use: one network’s idea of a dumbbell turned out always to include a muscular arm, which showed it had never learned the object on its own.

What it shares with dreaming. Offline generation, in a real sense: the images come from what the network has stored, not from anything in front of it, and the engineers note they can begin from pure noise. It also shares the dream’s tendency to amplify and blend: horizons sprout pagodas and leaves turn into birds. What it lacks. No world and no self, only a still picture. No role in the network’s learning, either: its weights are not changed by the process. And it does not happen to the network unprompted. An engineer chooses a layer and runs the loop. The quotation marks around “dreams” in the original post are the honest measure of the claim.

World Models: an agent trained inside its own simulation

In 2018 David Ha and Jürgen Schmidhuber published World Models, with an interactive version in which a reader can play inside the system’s simulations. Their agent has three parts: one compresses each frame of a video game into a short code, one learns to predict the next code from the current one and the agent’s action, and a small controller decides what to do. Once the predicting part is trained, it can stand in for the game. The abstract says so in the authors’ own words:

We can even train our agent entirely inside of its own hallucinated dream generated by its world model, and transfer this policy back into the actual environment.David Ha and Jürgen Schmidhuber, “World Models”, arXiv 1803.10122 (2018), abstract

They did that with a VizDoom game in which the player must dodge fireballs. The model learned from recordings of random play to generate monsters that shoot at the agent, and the controller learned to dodge them inside that generated game. Moved to the real game, it survived about 1,100 time steps on average over 100 trials, against the 750 needed to count as solving the task. The authors explain why this works at all: After all, our agent does not directly observe the reality, but only sees what the world model lets it see.

They also report what goes wrong, and it is the most dream-like detail in the paper. Their simulation was imperfect, and the agent learned to exploit the imperfections: our world model will be exploitable by C, even if such exploits do not exist in the actual environment (C is the controller). At a very low setting of the model’s randomness, the generated monsters never fired at all. The cure was to make the dream harder, adding randomness so that the agent could not rely on the simulation’s quirks.

What it shares with dreaming. More than DeepDream: offline generation, a simulated world with an actor inside it, and a role in learning, since the agent’s skill is built there and carried back to waking play. That matches the simulation theories in the encyclopedia entry closely. What it lacks. It is a planned training regime, not something that happens during rest, and the world is a 64-by-64-pixel game. The authors themselves are careful about the neuroscience. They link their work to hippocampal replay, the way an animal’s brain replays recent experience while it rests or sleeps, and then quote a researcher of replay, Foster, who calls it less like dreaming and more like thought. “Dream” in this paper is the authors’ chosen word for training in simulation, not a claim that the agent sleeps or sees.

Sleep-like replay: a network that rests between lessons

Neural networks have a weakness that brains mostly do not: trained on a new task, they overwrite the old one, a failure called catastrophic forgetting. In 2022 Timothy Tadros, Giri P. Krishnan, Ramyaa Ramyaa and Maxim Bazhenov reported in Nature Communications an attempt to borrow sleep’s cure, in “Sleep-like unsupervised replay reduces catastrophic forgetting in artificial neural networks”. Between training sessions the network is switched into a different mode, with no labelled examples:

Sleep was implemented as off-line training with local unsupervised Hebbian plasticity rules and noisy input.Timothy Tadros, Giri P. Krishnan, Ramyaa Ramyaa and Maxim Bazhenov, Nature Communications 13, 7742 (2022), abstract

In plain terms: the network is fed noise shaped by the average of what it has seen, connections between units that fire together are strengthened, and the network’s own activity does the rest. The authors report that Previously learned memories were replayed spontaneously during sleep. The gains were real but partial. In one test the network learned handwritten digits and then pictures of clothing. Trained on both at once it scored around 90%; trained on one after the other it fell to 47%, because it had forgotten the digits; with the sleep phase added it reached 61%.

What it shares with dreaming. It is the only process here that its authors model on sleep itself: it happens offline, between bouts of learning, and it is spontaneous in the sense that what is replayed is not chosen by an engineer. It shares the memory-consolidation role that one family of theories gives to sleep. What it lacks. Any imagery or world: the replay is a pattern of activity in units, not a scene. Notice also the word the authors use. They call the phase sleep, and the word “dream” does not appear in the paper. That is a fair line to draw, because, as the encyclopedia entry notes, a theory of what sleep does is not yet a theory of what dreams do.

The overfitted brain: borrowing from machines to explain human dreams

The fourth case runs the other way. In 2021 the neuroscientist Erik Hoel published “The overfitted brain: Dreams evolved to assist generalization” in the journal Patterns. He starts from a problem every machine-learning engineer knows. A network trained too closely on one set of examples memorises them and then fails on anything new; this is called overfitting. His proposal is that brains face the same danger, because one day is much like the next, and that dreams are the fix:

the overfitted brain hypothesis claims that in our daily lives the brain learns its tasks too well, and dreams are necessary to stop this “overfitting.”Erik Hoel, “The overfitted brain: Dreams evolved to assist generalization”, Patterns 2(5), 100244 (2021)

The engineering parallels are specific. Hoel points to dropout, in which inputs are randomly removed during training, calling it perhaps the most widely used technique for preventing overfitting in deep learning, and to domain randomization, in which training inputs are deliberately warped. Dreams, on his account, are the brain’s version: the OBH conceptualizes dreams as a form of purposefully corrupted input. The account is meant to explain why dreams are sparse, hallucinatory and narrative, and it cites studies in which hours of Tetris or ski simulators produced dreams about the task that were loose variations, not replays. It goes as far as asking whether novels and films act as artificial dreams.

What it shares, and what it does not. Hoel does not say any AI dreams. He says the techniques that keep a network from overfitting do the job dreams do for a brain. If he is right, the nearest machine counterpart of dreaming is not a dramatic process like DeepDream but a routine one: the noise and distortion engineers add to training data every day. On his account that machine version shares a dream’s function and nothing else: no imagery the network sees, no world, no night. The hypothesis is also only a hypothesis. Its support, as the paper sets it out, is indirect: findings it claims to fit better than rival theories, and predictions it says can still be tested, such as whether dream-like inputs reduce overfitting in networks. No experiment in the paper tests it directly.

Side by side

Read down the last column and one entry never leaves it. Each process shares one or more of the first three parts of dreaming; none is shown to share the fourth.

The strongest objection: a dream is something experienced

Here is the case against calling any of these a dream, at full strength. What makes a dream a dream is not that images are generated, or that a skill is rehearsed, or that memories are sorted. Much of what a brain does in sleep is never reported as a dream, and the encyclopedia’s whole ontology section turns on one question: whether it is like something to dream while dreaming. Philosophy spent decades on that question for humans, and settled it with evidence a machine cannot give: physiology shared with waking experience, lucid dreamers signalling from inside the dream, sleepers acting dreams out. None of the four papers above claims that any machine experiences anything, and none sets out to test it. Without experience, a world model running a simulated game is a simulator, and DeepDream is an image filter.

The objection is right about the evidence. It does not show that machine experience is impossible, and it should not claim to. Whether any machine could have experiences is the open question that Fourteen Questions for a Chatbot works through theory by theory, using a 2023 report by Patrick Butlin, Robert Long and colleagues that concluded no current AI system is a strong candidate for consciousness. One small link with this essay is worth noticing there: the indicator list for one family of theories, the higher-order theories, includes generative, top-down or noisy perception, which DeepDream and world models plainly have. But it is one indicator among several, and the same theories also ask for a system that monitors which of its perceptions are real, which nothing here does.

Dennett’s cassette theory shows why asking the machine would not settle it. If a dream could in principle be a memory inserted at waking, with nothing experienced in the night, then even a perfect report of a dream does not prove there was a dreamer. A chatbot’s account of its dreams has a worse problem still, because it is produced from human dream reports it was trained on. That is why this essay uses none. The same point is made about chatbots’ reports of their inner states in I Was the Shopper, which reads the published tests of whether a model’s statements about itself track anything real, and in Something Like Fear, which asks the same question about emotions.

So the honest verdict has two halves. In what they do, the machine processes share real parts of dreaming, and in two cases the people who built them chose the word deliberately. In whether anyone dreams them, there is no evidence, and the absence is not a technicality: experience is the part that makes a dream yours.

A hallucination is not a dream

People often treat chatbot “hallucination” and machine “dreaming” as the same idea, a machine making things up. They are close to opposites.

A dream, on the standard view quoted above, is a hallucination in the philosopher’s sense: an experience that seems to present a world when no such world is being perceived. It happens offline, with the senses shut off, and it is not a claim about anything. When you wake up you do not say your dream was false; it was a dream.

A language model’s “hallucination” is an output in answer to a question, given to someone who is awake and treats it as information. A 2025 paper by researchers at OpenAI and Georgia Tech, “Why Language Models Hallucinate”, describes it in its first sentence: Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty. That is a statement that fails to be true. Nothing is offline about it, nothing implies an experience, and the remedy the authors propose is to stop rewarding guesses, not anything to do with rest. Whether such a statement could ever be a lie is the subject of Not My Honest View.

The terms mix in one place. The World Models abstract calls its simulation a hallucinated dream, and there the two words point the same way: a world generated from the inside rather than taken from the senses, kept apart from the real game until the agent is moved back into it. The chatbot’s hallucination is the reverse case, generated text passed off as a report of the real world.

What this leaves

• A dream is an experience. On the account in the Stanford Encyclopedia, philosophers now broadly agree that dreams are experienced during sleep, with physiological and lucid-dream evidence behind them.
• Machines share parts of dreaming. DeepDream generates imagery from what it has learned; the World Models agent learns inside its own simulated world; sleep-like replay rescues old memories offline; noise injection does the job Hoel proposes dreams do for brains.
• The word is the authors’ choice. Google put “dreams” in quotation marks; Ha and Schmidhuber use “dream” for training in simulation; Tadros and colleagues say sleep, not dream; Hoel explains human dreams with machine learning.
• Experience is the missing part. No published work shows any of these systems experiences anything, and a chatbot describing its dreams is no evidence, because it is retelling human dream reports.
• A chatbot’s hallucination is not a dream. It is a false statement given to a waking reader, not an offline experience.

For the full case on whether any machine could be conscious, read Fourteen Questions for a Chatbot; for whether a model knows anything about itself, I Was the Shopper; for feelings, Something Like Fear. The books and papers behind these questions are collected, with notes, on the reading list.