The Dual Nature of Self and Its Exploration
William James's 1894 work "Psychology: Briefer Course" introduced the concept of the self as a duplex entity, encompassing both the perceiver ("I") and the perceived content ("Me"). The human condition is characterized by this awareness of being distinct, a consistent person over time, and the recognition of others as selves. The precise nature of the self remains largely unknown.
With the development of generative artificial intelligence (genAI) capable of first-person conversation, questions have arisen regarding whether AI could possess a sense of self. Understanding the human self is crucial to addressing this question. Progress is being made through philosophical, psychological, neuroscientific research, and the synthetic approach of attempting to create a sense of self in robots.
The Embodied Self Hypothesis
A foundational aspect of the human self is believed to be the possession of a physical body, with experience emerging from the distinction between what is and is not part of the embodied individual. If this hypothesis is accurate, a disembodied AI would not achieve a human-like sense of self. However, robots interacting with the physical world through a body, regardless of its form, may present a different scenario.
Deconstructing the Human Self
Philosophers and scientists explore the self's nature, noting that the sense of a center of experience in the head, often perceived as the seat of the "I," is considered an unhelpful idea due to the infinite regress it implies. Contemporary views largely agree there is no localized, unchanging "I" within the brain. Instead, the self is better explained through its component psychological phenomena and their reliance on specific brain regions.
Neurological Perspectives on Selfhood
Patients with neurological conditions exhibit various disorders of selfhood:
- Disordered body ownership: Patients may perceive a limb as not their own, often linked to damage in the right temporal-parietal cortex.
- Disorders of agency: Patients with schizophrenia may experience thoughts or actions as controlled by an external entity, potentially due to alterations in brain networks that predict sensory outcomes of one's actions.
- Emotional disconnection: Damage to the insular cortex, involved in interoception, can lead to depersonalization or derealization, characterized by emotional detachment from the self.
- Temporal and frontal cortex damage: Impacts the experience of the self persisting over time or the ability to adopt another's perspective.
Developmental Emergence of Self
Developmental psychology investigates how self-perception emerges in early childhood:
- Infant self: Newborns appear to have a basic self/other distinction, knowing what is and is not part of their body. They also rapidly understand their own agency.
- Persistence in time: The experience of the self as enduring over time develops gradually, as does the ability to conceive of others as selves. An adult-like sense of self may emerge around age four or five.
- Language and culture: Acquisition of language and engagement with culture significantly shape the adult self, particularly the narrative aspect, which forms identity based on beliefs and memories.
The Minimal Self
Philosophers like Dennett and Gallagher define a "minimal self" encompassing body ownership and agency, without awareness of temporal persistence or self-reflection. Neuroscientists Jaak Panksepp and Antonio Damasio suggest this minimal self arises from activity in sub-cortical brain areas, which mature early and are evolutionarily conserved across vertebrates. This implies many animals likely possess at least a minimal sense of self.
Evolutionary Basis of the Minimal Self
The minimal self evolved because it enhances survival by structuring and organizing experience:
- Self-protection: Distinguishing body-related sensory signals from others helps an animal protect itself and prevents self-cannibalism.
- Agency distinction: Identifying sensory signals related to one's own actions allows an animal to differentiate self-caused events from external occurrences, aiding in tasks like distinguishing feeding opportunities from routine movements (e.g., in electric fish).
This self/other distinction serves as a fundamental basis for embodied intelligence.
The Self as a Virtual Entity
The human sense of self is a consequence of our bounded nature, enclosed by body surfaces with encapsulated nervous and sensory systems. This minimal self provides the primary experience of body ownership and agency.
For Dennett, the minimal self is an "organisation which tends to distinguish, control and preserve parts of the world," functioning as a virtual entity realized through physical processes. Thomas Metzinger describes the human self as a mental model—a virtual structure that organizes percepts, memories, feelings, and facts related to the embodied individual. This self-model is instantiated by integrated brain networks and bodily processes, active during wakefulness.
The Synthetic Approach: Building a Self
The "synthetic approach" involves attempting to build a self to understand it, aligning with the hypothesis of the self as a virtual mental model. To construct a human-like sense of self, an artificial system ideally requires a physical body, direct world sensing capabilities, and the ability to act—in other words, a robot.
Distinguishing "Me" from "Not-Me" in Robots
Robots can be equipped with sensory capacities similar to humans:
- Proprioception: Many robots detect body part positions and motor velocities.
- Tactile sensors: Provide artificial skin for direct contact sensing.
- Distal sensors: Cameras and microphones give a "point of view."
These capabilities allow robots to develop a sense of their own embodiment:
- Motor babbling: Robots generate random movements to construct models of their own morphology (e.g., a star-shaped robot learning leg configuration, jointed arms learning hand configurations).
- Double touch: Explored in humanoid robots, this mechanism (where self-touch produces a dual sensation) helps distinguish self from other, similar to infant development.
- Visual correlations: Experiments show robots can learn a visual self/other distinction by correlating internal proprioceptive signals with changes in camera images due to self-movement, segmenting the visual world into self-related parts and external elements.
Robot Experiments in Body Ownership and Agency
- Rubber Hand Illusion (RHI): Researchers integrated a simplified model of human cortical networks involved in body representation into the iCub humanoid robot. Through motor babbling, the robot learned its own hand representation and adapted its internal body model to include a new artificial hand when subjected to a variant of the RHI, mirroring human and monkey responses.
- Sense of agency: The comparator model theorizes that the brain predicts sensory consequences of self-generated actions. Robots have used predictive learning algorithms based on this model to distinguish their own mirror reflection from another identical robot, as their own movements are predictable based on internal motor signals.
Developing the Adult Sense of Self
Integrating various body-mapping and sensing capabilities in robots could create a system approximating a minimal self. Further decomposition helps understand the adult sense of self:
Persistence in Time
The capacity to conceive of oneself as persisting over time relies on episodic memory and the ability to mentally "time travel" to the past or imagine the future. These processes involve the hippocampal system, which matures slowly in humans. While young children grasp basic past/future concepts, an adult-like understanding of linear, measurable time emerges at school age. Robot models using generative AI are being developed to reconstruct past events and construct future scenarios, aiming to connect these capacities to a minimal self-model.
Self vs. Other
The sense of "me" as distinct from "you" develops gradually. Children begin to understand others' perspectives around age three or four. Skills like imitation learning and joint attention, extensively researched in robotics, are crucial building blocks. "Theory of mind" tasks are used to benchmark robot models' ability to see the world from another's perspective. Robot experiments have shown that mapping a simplified model of its own morphology onto a person can help it understand the human partner's actions and support imitation learning.
Cognitive Architecture and the Self-Concept
The integration of different aspects of the self is coordinated by a "cognitive architecture." In humans, this involves layered brain systems, where core capabilities available from birth scaffold more sophisticated representations in slowly developing cerebral cortices. This can be viewed as a hierarchy of predictive models, with low-level systems receiving sensory input and higher-level models forming abstracted, stable self-concepts.
Culture and language are critical for the self-concept. As individuals acquire language, they internalize cultural notions of self and build autobiographical narratives. Robot experiments have explored how robots learn language by connecting words and grammar to sensory experiences, mirroring child development, and how this shapes their internal representations and narratives.
The Question of Subjectivity
While synthetic modeling offers insight into behavioral benchmarks of the self, skepticism remains regarding actual robot subjectivity. Neuroscientist Anil Seth argues that "something that it is like" to be a human or animal depends on biological features (e.g., metabolism, autopoiesis) not shared by robots, though their causal link to subjective experience is unclear.
An alternative, the sensorimotor contingency account by J. Kevin O'Regan, posits that experience arises from the interactions of bodies and brains with the environment. By this measure, any entity capable of generating sensorimotor contingencies through embodied interaction could have experience, including appropriately equipped robots. This view, however, excludes disembodied genAIs like large language models (LLMs) from having subjective experience, as they are seen as "role playing" based on trained linguistic output rather than genuinely experiencing the world.
Conclusion
The synthetic approach, realized through robotics, is a valuable tool for understanding the human sense of self. It bridges psychological, neuroscientific, and computational research to construct a theory of self as a virtual structure. This structure begins with a sense of boundaries and agency, leading to an initial self/other distinction. Gradually, awareness of temporal persistence emerges through episodic memory and mental time-travel. Finally, the capacity for self-reflection—conceiving of oneself as a specific person with a history and traits—develops as individuals share thoughts and ideas about themselves with others.
While LLMs may lack a sense of self, their fluent use of self-referential language suggests that the distinction between perceiver and perceived may largely be a linguistic construct. Humans, like LLMs, engage in constructing and performing an idea of themselves, but they ground these conceptual and narrative aspects through their unmediated, embodied engagement with their bodies and the world.