We Just Read A Sleeping Person’s Dream

Summary

The video describes a study that used AI to identify which of 15 recently learned objects a sleeping person’s brain appeared to be processing. It presents this as a step toward studying sleep-related memory without waking people—but not as a recording or reconstruction of a dream.

Sleep Interpreter study

• Volunteers learned associations between 15 objects and their sounds, such as a sheep and a bleat.
• While they slept, researchers used a 64-electrode EEG cap and played the associated sounds quietly. The sounds were intended to trigger the learned memories without waking the participants.
• The AI was trained on EEG patterns linked to the same objects while participants were awake and asleep. By comparing the patterns, it learned to distinguish object-related activity from the much larger, slower waves of deep sleep.
• The system reportedly identified objects in new participants better than chance. The video gives a figure of about 25% overall and says accuracy rose to nearly 50% when sounds arrived during the rising phase of a slow brain wave. The exact chance comparison depends on the task; these results do not mean the system could freely read a person’s thoughts.
• In a separate test without sound prompts, the AI’s choices tended to reflect material participants had studied earlier. Stronger sleep-related signals were also associated with better recall the next day—evidence consistent with memory processing during sleep.
• The team reportedly made the code, trained model, and recordings publicly available. The video says the system could run on ordinary graphics cards.

Earlier work and related research

• 2013, Kyoto: Yukiyasu Kamitani’s team repeatedly woke participants from sleep and asked what they had experienced. Using their reports and MRI data, the researchers trained AI to distinguish broad categories, such as people versus places, from brain activity recorded before awakening. The approach depended on participants describing their experiences.
• 2021, lucid dreaming: Ken Paller’s team tested two-way communication with trained lucid dreamers. A participant answered a spoken arithmetic question with deliberate eye movements while asleep. This was communication from within a dream, not AI reading dream content.
• Brain-image reconstruction: The video also discusses work from Michal Irani’s lab that used brain scans from awake people viewing images to generate approximate versions of those images. The video emphasizes that this was a separate project and had not been applied to dreams.

Brain signals and limitations

Deep sleep produces large, slow brain waves that can mask the smaller signals associated with specific memories. The Sleep Interpreter’s approach was to train across awake and sleeping recordings rather than simply suppress the slow waves. The video says the team’s model also performed better at particular points in the slow-wave cycle.

The system’s vocabulary was limited, and it recognized prompted, recently learned objects—not arbitrary thoughts, full dream narratives, or images from a sleeper’s mind. The video describes dream-stage recognition, larger vocabularies, image generation from sleep data, and less cumbersome sensors as future steps, not capabilities already demonstrated.

Privacy issues

The video raises concerns about consumer brain-sensing devices and mental privacy. It reports that a 2024 NeuroRights Foundation review found that many companies’ privacy policies allowed broad access to users’ neural data, with some permitting sharing with third parties. It also cites Chilean legal protections for brain data and a UNESCO agreement on the ethics of neurotechnology. These examples are presented as early safeguards, not a complete legal framework for dream-reading technology.

Researchers and sources featured

Researchers

• Yongjie Liu — Beijing Normal University team
• Jiang Tianzi and He Jiang — researchers named in connection with the sleep-signal analysis
• Yukiyasu Kamitani — ATR research in Kyoto
• Ken Paller — Northwestern University
• Michal Irani — Weizmann Institute
• Marcello Ienca — Technical University of Munich

Institutions, publications, and organizations

• Beijing Normal University and the Chinese Institute for Brain Research
• Neuron
• ATR Research Institute and Science
• Northwestern University and Current Biology
• Weizmann Institute
• MIT Technology Review
• NeuroRights Foundation
• Chile’s Supreme Court
• UNESCO