Dreams as Simulators: Could AI Help Design How the Brain Learns?

Imagine someone who has never driven a car. One night, they dream that they are driving. They hold the steering wheel, see other vehicles, press the brakes when danger appears, and make decisions as the road suddenly changes. When they wake up, they do not have the skills of someone who has practiced for months. Yet their brain has just gone through something resembling an experience.

The question is not whether dreams can replace real-world training. It is narrower and more interesting: can experiences in dreams become part of the human learning process?

What we know: sleep processes experience

Experiences we acquire while awake do not immediately become stable long-term memories. During sleep, the brain runs processes that stabilize and organize new information. One key finding is that patterns of neural activity from learning can be reactivated during sleep, as seen in the hippocampus of rodents and, with different methods, in humans. Sleep is not just physical recovery. The brain keeps working on what it has experienced.

Here a distinction often gets blurred. The evidence that sleep supports memory consolidation is strong. The evidence that the experience of dreaming itself plays a causal role is much thinner. A small study by Wamsley and colleagues (2010) found that participants who dreamed about a maze task they had just learned showed greater improvement afterward. The result is intriguing, but not enough to conclude that dreaming caused the improvement.

Dreams as simulation

One of the best-known theories of dream function is Antti Revonsuo's Threat Simulation Theory (2000). It proposes that dreams simulate threats so we can rehearse responses to them. In dreams we are often chased, falling, fighting, lost, or making decisions under pressure. If the theory holds, dreams are a biological simulator that lets the brain run scenarios without real-world consequences.

The theory remains contested. Many dreams are neutral or pleasant, and some studies find threats are less dominant than predicted. Other explanations include the continuity hypothesis (dreams reflect waking concerns), activation-synthesis (dreams arise from random brain activity that is then given meaning), and emotional regulation. Dreams may serve more than one function.

The ability to simulate is not limited to sleep. When we imagine giving a speech or performing an athletic movement, the brain constructs a representation of something not currently happening. For movement, this is called motor imagery. Research on mental practice shows that imagining an action can contribute to motor learning, especially alongside physical practice. Imagining swimming does not make someone a good swimmer, but internal representations clearly can support skill.

Lucid dreaming: a virtual reality built by the brain

Lucid dreaming is the state of knowing you are dreaming. For some people, this awareness allows them to influence the dream: changing the setting, flying, talking to dream characters, or trying particular activities. If virtual reality is a computer-generated environment, lucid dreaming is a biological version produced by the brain itself, with no external device.

Does practicing in a dream improve ability after waking? The evidence is early. A few small studies, such as Erlacher and Schredl (2010) on tossing coins into a cup, report modest improvement after lucid dream practice. These results cannot yet be generalized, and lucid dreaming is hard to induce reliably. Someone who dreams about driving will not necessarily be able to drive. They still need physical coordination, sensory experience, and real practice. Dreams are best viewed as a complement to training, not a replacement.

When AI becomes a designer of experiences

AI can already generate images, sound, video, characters, stories, and virtual environments. Combined with the brain's capacity to simulate, AI could become more than a source of information. It could become a designer of learning experiences.

Consider someone learning Japanese. Instead of vocabulary lists and multiple-choice quizzes, they enter a simulated restaurant. A waiter greets them and asks a question in Japanese, and they must respond. The AI analyzes vocabulary, pronunciation, response time, and errors, then builds the next scenario around their weaknesses. Learning shifts from reading and memorizing to experiencing, trying, failing, correcting, and trying again.

Connecting simulation with sleep

One approach already studied is Targeted Memory Reactivation (TMR). During learning, information is paired with a cue, such as a sound or a scent. The same cue is presented again during sleep to support consolidation. An early study by Rasch and colleagues (2007) used a rose odor, and later research has shown that TMR can improve recall, with moderate effects that vary across studies. TMR does not mean knowledge can be inserted into dreams. It shows that processes during sleep can still interact with information from the environment.

That interaction has also been shown in dreams. Konkoly and colleagues (2021) demonstrated simple two-way communication with lucid dreamers: researchers asked questions, and dreamers answered with eye or facial muscle movements. This is a real milestone, but it is very far from designing dream content precisely. Research on dream incubation at MIT (Horowitz and colleagues, 2023) suggests that themes can be nudged around sleep onset, though control remains crude.

An imagined learning cycle

From these pieces, we can imagine a cycle:

AI identifies a learning goal and a weakness, such as difficulty with public speaking.
AI designs a simulation matched to the person's level.
The person practices during the day, with a cue paired to the experience.
During sleep, the same cue is replayed to support consolidation.
The brain processes the experience and may generate related dreams.
After waking, the person practices again in the real world.
AI evaluates the results and designs the next simulation.

Such a system would not merely make dreams more interesting. It would be a personal cognitive training system, combining the brain's biological ability to simulate with AI's ability to tailor practice.

To be clear, for the sleep and dream stages this cycle is still speculation. The pieces are real, but there is no evidence yet that the whole works.

The brain, machines, and the limits of reading dreams

We are far from being able to read dreams. EEG does not display an image labeled "this is a house" or "this is a face." It records electrical activity from large populations of neurons. With AI, researchers can look for relationships between brain activity patterns and the information being processed, a field called neural decoding. But recognizing certain patterns is not the same as reading a mind. The brain is far more complex than a signal that translates into a single word or image.

Brain-Computer Interfaces (BCIs) show that the link between brain and machine is gradually becoming real. BCIs record brain activity with sensors, then use algorithms to learn patterns and translate them into commands. Research has demonstrated BCIs that control cursors, robotic arms, and assistive technologies. Recent studies (Willett et al., 2023; Metzger et al., 2023) have even translated neural activity in paralyzed patients into speech.

A further step comes when communication flows not only from brain to machine but back again. Imagine someone controlling a robotic hand through brain activity while sensors on the hand send touch information back to the nervous system, producing sensation. This is a closed-loop neuroprosthesis.

What about telepathy and memory transfer?

Biological telepathy, the direct exchange of thoughts without any device, has no convincing scientific support. Technology, however, could produce something functionally similar:

Brain A → BCI → Computer/AI → Network → BCI → Brain B

This would not be paranormal telepathy but technology-mediated brain-to-brain communication. Existing demonstrations are very limited, such as sending simple binary signals between humans.

Transferring memories is far harder. People imagine memory as a file stored somewhere that could be copied. In reality, memory involves changes and activity across many neural networks, spanning vision, sound, emotion, bodily sensation, place, language, and links to other experiences. Memory is also reconstructive: it is rebuilt and altered each time it is recalled. "Downloading a memory" is the wrong metaphor.

Progress will most likely be gradual: brain to computer, computer to brain, two-way communication, and then a brain-AI relationship resembling a cognitive copilot.

Fact, research, and speculation

Supported by evidence: sleep plays a role in memory consolidation; the brain stays active during sleep; mental imagery can contribute to training; lucid dreaming allows some awareness within dreams; certain stimuli can influence memory processes during sleep; BCIs can translate some brain activity into machine commands.

Under active research: more advanced neural decoding, neuroprostheses with sensory feedback, brain-to-speech systems, AI-assisted BCIs, and more direct communication between brains and devices. In the dream domain: whether dream experience contributes to learning, and methods for influencing dream content.

Still speculative: fully controlling dreams, inserting complex skills through dreams, downloading memories, directly transferring skills, sharing subjective experience, and high-bandwidth technological telepathy.

Imagining the third category is not wrong. Speculation helps us explore where technology might lead. What matters is not confusing possibility with established fact.

Ethical considerations

Technology that touches sleep and dreams carries its own risks.

Sleep quality. Stimulation during sleep can disrupt it, and disrupted sleep harms memory and health. Any system must prove it does no net harm.
Privacy. Dream data is among the most intimate mental data there is. Who owns it, and who may profit from it?
Manipulation. A tool that trains speaking skills could, in principle, be used to shape attitudes or associations while someone is in a vulnerable state.
Therapy may come first. Imagery rehearsal therapy for nightmares and PTSD is already used clinically. Deliberate influence over dreams will probably prove useful there before it does for skill training.
Conclusion

Long before virtual reality existed, the human brain could build its own worlds during sleep. We can suddenly find ourselves somewhere we have never been, speak with people who are not present, feel fear, joy, and loss, and even make decisions in a world generated entirely by brain activity. Dreams may not literally be biological computers, but they reveal something remarkable: the brain can build a model of the world and run simulations from within.

So the more important question may not be "can humans control their dreams?" but rather: how far can humans use their own brains as simulators for learning?

We are not there yet. We cannot insert complex skills into dreams, download memories, or freely read thoughts. But the direction is becoming visible: the brain produces signals, machines learn to recognize their patterns, AI helps interpret them, and devices are starting to respond.

If AI one day learns to understand this biological simulator, the ultimate question may no longer be whether computers can imitate humans. It may be how far humans and AI can learn together through the simulator that has always existed inside us: the human brain.