What AI Is Learning About Sleep (2025–2026)

SomniSmart editorial team · Published October 28, 2026 · Sources checked against Crossref, Europe PMC and official pages

A note on "what AI wants to learn." SomniSmart is built with Grok, an AI assistant made by xAI. Grok does not have wishes, and it does not learn from your quiz answers or conversations on this site. When this page talks about what AI "is learning" or "wants to learn next," it means what researchers are training AI models to do, and the open questions the published research points to. Every claim links to its source below.

1. What AI and sleep science learned this past year

One night of sleep may hint at future disease New, Oct 2025 to Oct 2026

The biggest AI sleep paper of the year came from a team led by Stanford researchers. They built SleepFM, a "foundation model" for sleep, published in Nature Medicine in January 2026^[1]. Instead of being trained for one narrow job, the model learned general patterns from more than 585,000 hours of polysomnography (overnight sleep-lab recordings) from about 65,000 people. These recordings combine brain waves, heart rhythm, breathing and muscle signals.

From a single night, the authors report that SleepFM predicted 130 future conditions with a concordance index (C-index) of at least 0.75. The C-index is a 0-to-1 measure of how well a model ranks who gets a disease first, where 0.5 is chance. Reported values included all-cause mortality 0.84, dementia 0.85, heart attack 0.81, heart failure 0.80, chronic kidney disease 0.79, stroke 0.78 and atrial fibrillation 0.78^[1]. It also held up on a Sleep Heart Health Study dataset it had not been trained on. On routine tasks it was competitive with dedicated sleep-staging tools (F1 0.70–0.78), and it detected sleep apnea presence with 0.87 accuracy^[1].

What this does and does not mean: it is a research model run on sleep-lab recordings, not a consumer app or a diagnosis. It shows associations between sleep signals and later disease risk, not that poor sleep causes those diseases. Predicting risk is also not the same as proving that acting on the prediction helps anyone.

AI is helping people talk from inside a dream New, Oct 2025 to Oct 2026

Lucid dreamers know they are dreaming, and they can signal to researchers with agreed eye movements. A 2025 paper in the journal Dreaming tested deep neural networks that read these eye signals from two-dimensional electrooculography (sensors that track eye movement)^[3]. The aim is to make communication from the dream faster and more reliable than counting left-right eye sweeps by hand.

A clearer brain signature of lucid dreaming New, Oct 2025 to Oct 2026

Lucid-dream studies are usually small. A 2025 Journal of Neuroscience study pooled data across laboratories and used a new artifact-cleaning pipeline^[4]. At the scalp, lucid and non-lucid REM sleep looked almost the same. Source analysis, however, found reduced beta power (12–30 Hz) in right central and parietal areas, including the temporoparietal junction, and increased alpha-band connectivity during lucid dreaming^[4].

Why some people remember dreams and others do not New, Oct 2025 to Oct 2026

A February 2025 study followed 217 healthy adults aged 18–70, combining dream diaries with sleep tracking and EEG^[5]. People were more likely to report a dream on waking if they had a positive attitude toward dreaming, tended to mind-wander, and had particular sleep patterns. Age and vulnerability to interference predicted recalling the dream's content. Recall also varied night by night and with the seasons^[5].

2. What is being studied right now

From sleep lab to home. SleepFM was trained on full lab recordings^[1]. A natural next step is to see whether similar models work with the fewer, noisier signals from home tests and wearables. That is still being tested, and we found no peer-reviewed result yet showing SleepFM-level disease prediction from a consumer wearable.
Sleep staging and sleep apnea detection keep improving. Foundation models are now being compared directly against specialised tools such as U-Sleep and YASA^[1].
Two-way dream communication is moving from proof of concept^[2] toward machine-learning decoding of dreamers' signals^[3].
Nightmare treatment continues to rely on established therapy. The 2018 AASM position paper^[10] was formally endorsed again in Sleep Medicine in March 2025^[11].

3. Open questions: what researchers want AI to learn next

These are open research questions, not results:

Does it generalise? Will sleep foundation models stay accurate across age groups, ethnic groups, sleep clinics and devices they never saw in training?
Why does it predict? Which sleep features drive a dementia or heart-failure prediction, and are they something a person or clinician can change?
Does acting on it help? Risk prediction needs trials showing that early action changes outcomes.
Can AI read dreams? Decoding from brain scans so far shows broad categories of content at sleep onset^[6], not full dream narratives.
Can we guide memory processing during sleep safely and at home? See targeted memory reactivation and closed-loop stimulation below^[8]^[9].
Privacy. If one night of sleep data can hint at disease risk^[1], who should be allowed to see it?

4. Neuralink, brain-computer interfaces and sleep

What is verified: Neuralink's US PRIME study is an early feasibility trial of its N1 implant and R1 surgical robot. It is for people with tetraparesis or tetraplegia and aims at controlling external devices such as a computer^[12]. The registry lists a start date of January 9, 2024, status "recruiting", an estimated 15 participants, primary completion in June 2026 and overall completion in January 2031^[12]. In the UK, University College London Hospitals reported on January 29, 2026 that seven GB-PRIME patients had surgery between October and December 2025^[13].

Neuralink and sleep: as of October 28, 2026 we found no announced sleep application from Neuralink, and its trials are about device control, not sleep. Participant totals across countries are reported mainly by the company and the press, so we have not repeated them here.

Where brain-computer techniques and sleep do meet is in research labs using non-invasive tools:

Closed-loop acoustic stimulation. Soft sounds were timed to the "up-states" of slow brain waves during deep sleep. This strengthened the slow oscillations and sleep spindles and improved next-day word-pair memory, while sounds played out of phase did not help^[9].
Targeted memory reactivation (TMR). A cue such as a sound paired with something learned is replayed during sleep. In a 2009 study, cued items were remembered better^[7]. A 2020 meta-analysis of 91 experiments (2,004 participants) found an overall effect of Hedges' g = 0.29, meaning a small but reliable benefit. The effect appeared in N2 and slow-wave sleep but not in REM sleep or wakefulness^[8].

5. Dream science

Background Two-way communication with lucid dreamers. Across four labs, 36 people in polysomnography-verified REM sleep heard questions, such as simple maths problems, and answered with eye movements or facial muscle twitches. They answered correctly on 29 occasions across 6 of the participants^[2].
Background Dream decoding. A Japanese team trained machine-learning models on fMRI brain activity and predicted the visual content people reported just after being woken at sleep onset^[6]. This suggests dream imagery uses the same brain patterns as seeing.
New, Oct 2025 to Oct 2026 Lucid dreaming signatures^[4] and AI-read dream signals^[3] are summarised above.
New, Oct 2025 to Oct 2026 Dream recall depends on attitude, mind-wandering, age, sleep patterns and even the season^[5].
Background What REM sleep is for is still debated. Leading ideas include memory processing, emotional regulation and simulating threats. Notably, the TMR meta-analysis found cueing effects in non-REM sleep, not REM^[8], which adds to the evidence that much memory consolidation happens in deep non-REM sleep.
Background Nightmare treatment. For adults with nightmare disorder, the AASM position paper lists image rehearsal therapy as recommended. In this therapy you rewrite a recurring nightmare while awake and rehearse the new version^[10]^[11]. Nightmare disorder affects about 4% of adults^[10].

What this means for your sleep tonight

None of this research replaces the basics: a supportive mattress, a pillow that keeps your neck aligned, and a cool, dark, quiet room. Our advisors turn the evidence into a personal recommendation:

Not medical advice. This page summarises research for general education. Snoring with pauses in breathing, severe daytime sleepiness, acting out dreams, or frequent nightmares deserve a visit to a doctor or sleep specialist.

Sources

Some links on SomniSmart are affiliate links; they never change what we recommend. Spine Align (Chirobed) is our own brand and is always labelled. Affiliate disclosure.

What AI is learning about sleep