Artificial intelligence (AI) that works with mind-reading machines is the type of technology that we’ll need to reproduce what we experience in our dreams.
A well-publicised Japanese research study demonstrated the beginnings of the method in 2023: researchers recorded the brain activity of sleeping participants using functional magnetic resonance imaging (fMRI) scanners and then used machine learning to classify perceived objects, such as a key, a person or a chair from the activity.
However, the study was focused on sleep onset: the first two stages of sleep during which we experience visual imagery (hallucinations) and didn’t look at dreams at all.
They chose this approach so the participants could wake up and quickly describe what they saw.
To reproduce our dreams, we need a lot of detailed fMRI data from dreaming volunteers with which to train a large AI. The volunteers need to be very good at remembering their dreams in a lot of detail so that we’ll know how accurate any prediction may be.
This is likely to be the most difficult part about recording dreams and it’s unclear how this kind of data can be generated reliably.
But we already have a head start in a related area: there have been research studies that have produced huge datasets of fMRI brain activity while conscious participants watch thousands of videos, listen to spoken-word recordings and read text.
Using AIs trained on these datasets we can already predict what a waking person may be watching or reading.
If we assume that, in a few years, we’ll have sufficient data to make such an AI and we have portable fMRI machines that are quiet enough to let you sleep while they scan your dreams, then we already have the methods needed to show the results.
Generative AIs, such as Sora from OpenAI and Lumiere from Google DeepMind, can make very dream-like video sequences already.
Use the dream-analysing AI to provide a clear textual description for the generative AI and you’ll get a video illustrating the likely dream sequence from that person.
A word of warning, though. These AIs are not really reading minds; they’re just matching patterns of brain activity to likely images they’ve seen before. And the generative AI will have no idea if its video resembles your dream – it’ll simply string the images together, perhaps with some rudimentary story.
So, while the result will look spookily dream-like and may contain many of the same elements as the original dream, it won’t be an accurate reproduction – any more than the Tom Hanks movie Cast Away resembles the true story of José Salvador Alvarenga, the fisherman who spent 14 months stranded on a fishing boat adrift in the Pacific.
AI is amazing, clever and spooky. It’s just not always very accurate when it comes to our brains.
This article is an answer to the question (asked by Andrew Taylor, via email) 'How close are we to being able to record our dreams?'
To submit your questions, email us at questions@sciencefocus.com, or message our Facebook, X, or Instagram pages (don't forget to include your name and location).
Check out our ultimate fun facts page for more mind-blowing science.
Read more:
