By
Stuart Kerr, Technology Correspondent,
LiveAIWire

Sleep science has established that dreaming is associated with
memory consolidation, emotional processing, and creative problem-solving in
ways that make it functionally significant rather than merely epiphenomenal.
What the field has not established is a reliable method of decoding dream
content from neural signals with enough precision to make that content
legible to external observers. AI is changing that calculus, not by solving
the hard problem of dream interpretation, but by making possible the analysis
of neural and physiological data at scales and resolutions that were
previously impractical, and by identifying patterns in that data that
correlate with reported dream content in ways that are beginning to achieve
statistical significance.

Research published in Science
in 2013 documented an AI system trained to decode the
semantic content of dream reports from fMRI brain activity during sleep,
achieving above-chance accuracy in identifying broad categories of dream
content from neural data. The system was not reading dreams in any
comprehensive sense. It was identifying neural signatures associated with
categories of visual and semantic content that appeared in subsequent verbal
reports of dream experience. The gap between that capability and a technology
that could reliably reconstruct the specific narrative content of a dream
remains substantial. But the direction of travel is clear, and the rate of
progress in neural decoding more broadly suggests that the gap will
narrow.

What Dream Decoding Could Be Used For

The research applications of AI-assisted dream analysis are
legitimate and potentially significant. Understanding the neural basis of
dreaming contributes to sleep science, to the study of disorders including
PTSD in which nightmares are a primary symptom, and to basic neuroscience
research on consciousness and memory consolidation. The therapeutic
application of dream content in psychotherapy has a long clinical history,
and AI tools that can provide more precise and verifiable accounts of dream
content could potentially enhance therapeutic work with patients for whom
verbal dream recall is inconsistent or incomplete.

The surveillance applications are more concerning and are not
hypothetical. Employer wellness programmes that monitor sleep quality and
patterns are already deployed in some corporate contexts. Military and
security agencies have funded research into neural decoding with obvious
potential applications in interrogation and intelligence. The journey from AI
systems that identify broad dream content categories to systems that provide
commercially or coercively useful information about an individual’s mental state
during sleep is not technically resolved, but it is in progress. As our
analysis of how
AI genomic analysis creates privacy consequences for people who did not
consent to data collection found, the most significant privacy
risks from AI arise from the extension of capability into domains where the
subject’s ability to consent is structurally limited. Dream data sits at an
extreme of that spectrum.

The Consent Architecture Problem

The consent problem in AI dream analysis is structurally similar
to the consent problem in genetic genealogy: the data being collected and
analysed is intimate, reveals information the subject may not have intended
to disclose, and is generated in a state of reduced capacity. A person asleep
and dreaming is not in a position to evaluate or withdraw consent for
monitoring of their neural activity. The unconscious mind during dreaming may
reveal fears, desires, and associations that the person would not voluntarily
disclose and whose disclosure could be used against their interests in
employment, insurance, or legal contexts.

The regulatory frameworks for neurotechnology and sleep monitoring
are at early stages in all major jurisdictions, and the pace of capability
development in neural decoding is likely to outpace them. The OECD’s
Recommendation on Responsible Innovation in Neurotechnology,
adopted in 2019 and the first international standard in this area,
establishes principles including mental privacy and non-discrimination, but
its implementation across member states has been partial and its application
to commercial sleep monitoring products remains contested.

The Interpretation Limit

A final consideration that tempers both the promise and the alarm
around AI dream decoding is the interpretive limit of what neural correlates
of dream content can tell us. Dreams are not encrypted messages that decode
to precise propositional content. They are multimodal, emotionally saturated
experiences whose meaning is contextual, personal, and constructed in the act
of recollection and interpretation rather than inherent in the neural
activity that generated them. An AI system that identifies the neural
signature of a dreamed face is not identifying who the dreamer loves or
fears. It is identifying a perceptual pattern whose emotional and semantic
significance requires interpretation that the system cannot
perform.

This limit does not eliminate the privacy risk of dream
surveillance. Employers, insurers, and security agencies do not need to
decode dreams accurately in order to use dream data in ways that harm
individuals. Correlation between dream patterns and subsequent behaviour,
health outcomes, or stress indicators may be commercially or coercively
useful even in the absence of accurate dream content decoding. As our
coverage of how
AI creates governance gaps in domains where its effects are least
visible found, the most consequential governance challenges in AI
develop fastest in the spaces where public awareness and regulatory attention
are lowest. The unconscious mind is the most private domain that AI is now
beginning to reach.

The Research Opportunity

The most constructive near-term use of AI dream research is in
clinical contexts where understanding dream content has established
therapeutic significance. PTSD treatment using trauma-focused therapies
relies partly on patients’ ability to recall and process nightmare content,
and AI tools that can provide more precise accounts of what neural activity
during nightmares looks like could enhance the targeting and evaluation of
treatments. Sleep disorder medicine similarly benefits from better
understanding of the neural processes associated with different sleep stages
and their disruption, and AI analysis of polysomnography data is already
improving diagnostic accuracy in ways that directly benefit
patients.

These applications use AI to enhance clinical understanding within
established ethical frameworks, with patient consent, clinical oversight, and
clear therapeutic purpose. They represent the most defensible use of AI dream
research in the near term, because they deliver benefit to identifiable
patients under conditions that governance frameworks are equipped to
evaluate. The broader surveillance and commercial applications of dream data
analysis represent a different category that requires governance frameworks
not yet in place. As our analysis of how
AI applications targeting private human experience require more protective
governance than consumer product frameworks typically provide
found, the precautionary principle applies most strongly where the data being
collected is most intimate, where the subject’s capacity to consent is most
limited, and where the potential for misuse is most consequential. Dream data
meets all three criteria.

The Ethical Framework That Is Missing

The ethical framework for AI dream research and its downstream
applications does not yet exist in a form adequate to the capabilities being
developed. Informed consent in the context of neural data collection during
sleep presents challenges that standard research ethics frameworks were not
designed to address. The boundary between sleep research and surveillance,
clear in principle but blurry in practice when the same neural monitoring
technology serves both functions, requires governance that regulators have
not yet developed.

The precedents being set in research contexts will shape how AI
dream analysis is regulated when it moves into commercial and law enforcement
applications, which the history of AI technology suggests will happen faster
than governance frameworks can be developed reactively. The appropriate
moment for establishing the ethical and regulatory boundaries of AI dream
analysis is now, when the technology is developing but not yet deployed at
scale, rather than after deployment has created commercial interests and
operational dependencies that make retroactive regulation significantly
harder.

For related coverage, see our analysis of the
limits of what AI can reliably know, our look at how
people are resisting AI surveillance, and our broader examination
of what
AI governance currently looks like.

About the Author

Stuart Kerr is the Technology Correspondent for LiveAIWire. He
writes about artificial intelligence, emerging technology, and the forces
reshaping work, business, and society.