For better or worse, biological consciousness continues to inspire our thinking about machine consciousness. It is our only available source of testable knowledge about the phenomenon. Only some, or none, or all of these insights may eventually apply to digital minds: extrapolate with caution. That said, we should be more relentless in attempting to solve the hard problem of consciousness, especially as it relates to selfhood. It's fascinating and may soon turn out to be extremely important. For epistemic status, scroll to the bottom of the page.
Introduction
Everything in the history of a species traces its roots back to evolution, and in evolution, the million-dollar question is why. From an infinite possibility space, some strangely specific and recurrent patterns emerged and stabilized, then remained in existence for millions of years. Why were these singular patterns favored?
For example, distantly related species have independently developed the camera eye, in which a lens focuses an image onto a sheet of photoreceptors. Of all the possible engineering solutions, why this one and why more than once? A plausible explanation looks like so: a moving animal needs to know where things are. In a three-dimensional world, few designs can produce a sharp, accurate image. Starting from a flat light-sensitive patch, every degree of cupping improves the picture, and a focusing lens over the cup eventually improves it further. This translates into finding food and avoiding harm, and therefore surviving.
Because no success in evolutionary history is exempt from the question of why, apply it to the biggest biological mystery we know: the hard problem of consciousness. Why should anything feel like anything at all? What specific advantage does the first-person experience provide that nothing else does? No hypothesis of consciousness is satisfying until it arrives at a plausible and relatively simple explanation. Before I suggest one such possibility, some preliminaries.
Evolution doesn't care about your personal beliefs on consciousness
Consciousness in humans and other species could be an "illusion" of the underlying computation or its own fundamental entity made of a different substrate. It could be independent of the underlying hardware or only possible in biological brains. For all that evolution cares (which it doesn't), a stone might be conscious, but animals evolved a still more elaborate form of first-person experience. Ultimately, all of these positions are bound by the same question: why was the first-person experience selected from infinitely many other possibilities? Why did it win over unconscious automata? Why do we view the world as individuals rather than as arbitrary collectives? Once you come up with your own explanation, your next challenge is this: convince yourself (and then others) that such a function has not already been accomplished in existing non-conscious systems, or is not otherwise trivial to design. It is still marginally possible that biology addressed some relatively simple engineering problem by slapping on the mysterious first-person experience—ignore this unlikely possibility. You will produce much more interesting hypotheses by looking for why phenomenal consciousness was indispensable rather than somewhat useful.
You don't have to be conscious to be evolutionarily successful
To succeed in evolution, you could take any of the following routes:
Non-cells: You could be a virus—a collection of biological molecules that relies on others for replication, spreads in bouts of epidemics, and depends heavily on chance. Some scientists will doubt you are a life form, but as a virus (or even a prion), you are nonetheless a successful player in the game of evolution.
Single cells and undifferentiated unions: You could be a unicellular organism and outnumber everyone, like bacteria. You could even join with other cells that share your goals and develop new collective features, but you are so aligned and connected with your companions that you don't need a high-speed information network. Your success relies on multiplying a ton and remaining relatively small. You don't need first-person experience.
Sessile differentiated unions: You could make highly complex unions of differentiated cells that form further hierarchies and organizations, but stay put in one place—and voilà, you are a plant! You are now composed of cell types that process different signals, but aren't designed to run away: by default, you germinated in an environment that was already favorable. Parts of you that get consumed may regrow, and some others are straight-up inedible. You don't need first-person experience either. If anything, it would be a liability: pair experience with valence, and life gets pretty miserable for something that can't escape.
Moving differentiated unions: What if you were a complex structure that also moves relatively large distances? You are now a moving animal. The highly specialized cell types of your body (such as light-sensing cells or digestive cells) may issue conflicting demands. In this uncomfortable union, you very likely need a high-speed signaling system (e.g., a nervous system) to calculate which signals to follow and which to ignore for the moment. The calculation had better be rapid, too, or you will quickly get eaten by some predator. Predation really took off in the Cambrian period (some 500 million years ago), when fossils show evidence of a rapid arms race between predators and prey. The race sharpened the senses of both, elaborated centralized brains (the central nervous system, or CNS), and likely produced the earliest consciousness. Many evolutionary biologists think consciousness evolved at least twice: humans and octopuses are both believed to have consciousness, but our most recent common ancestor was likely unconscious. Each lineage must have developed it on its own. The question remains: if the CNS is a complex probabilistic decision machine, why not stay a biological robot? Why did the first-person experience evolve not once, but twice?
Keep it short and sweet
Shocking as it may seem, evolution the designer is unintentional, impersonal, without a plan—not actually an expert craftsman of bespoke mechanisms. Complexity compounds over millions of years, but only via small incremental steps that survive selection. A feature as robust as consciousness could not have been assembled from billions of fragile serial events—unless something simple rewarded and strengthened them along the way. In other words, a fairly tractable and conserved principle must underlie consciousness. Furthermore, since evolutionary pressures are very coarse and best interpreted at their corresponding levels, the purpose of consciousness should be inferable from things we witness without microscopes, neural recordings, or sophisticated math. For example, energetically stable folding may affect the survival of molecules, but consciousness operates in the macroscopic world, and there seems to be one consciousness per animal. It is most likely shaped by macroscopic evolutionary pressures. Those pressures should be fairly comprehensible to a well-informed general audience—your first proposal for the purpose of consciousness can safely be brief and straightforward.
Current hypotheses fail to establish the necessity of first-person experience
The most popular existing hypotheses of consciousness, to varying degrees, miss the mark on identifying the unique advantage of first-person experience. One (admittedly high-bar) litmus test is the following: take the function claimed by a theory, and ask whether an unconscious system could perform it.
Illusionism. In its most fundamental version, illusionism states that the first-person experience is not a separate entity, but an illusion created by the underlying computation. This is not a functional theory, but a broader philosophical position that underlies several of the hypotheses mentioned below. The position deserves separate clarification because it is often misused: many quote it to say that there is no point in looking for the purpose of first-person experience ("It's nothing, it's just an illusion"). Consciousness might be a type of illusion, but it isn't just fumes: in moving animals, it won decisively over infinitely many alternatives that lack first-person experience. Why?
Global Workspace Theory. Broadly, the global workspace theory (GWT) claims that there is a center stage where specialized processes compete for access, and the winner is broadcast to the rest of the system. It is a genuine theory that makes predictions, but, to me and many others, it pertains to access consciousness, or the "items that are currently on your mind." Proponents of GWT argue that once access is explained, nothing about phenomenal consciousness is left to explain. But this dismisses the question of why broadcast necessarily comes with first-person experience. Computational workspaces that few would credit with phenomenal consciousness already exist, from blackboard architectures in classical AI to, more recently, J-space in LLMs. The global workspace problem appears solvable without necessarily assigning a subjective point of view.
Attention Schema Theory. AST proposes that the brain models its own attention similarly to how it models its own body. This provides a significant survival advantage by allowing systems to (1) guide their own attention and (2) model the attention of other biological systems. The theory also makes a point about the energy efficiency of simplified models, which are cheap to run. So far, so good. The authors go on to say that querying the model of attention is what produces the judgment of having first-person experience. It is unclear why an unconscious feedback loop would not do the same without producing any such judgment. The authors show that, when given an attention schema, simple computer game agents improve their attention control, and small transformers predict other agents’ attention better. One would be hard-pressed to assign these systems first-person experience. It seems as though phenomenal consciousness is not indispensable for attention modeling, either.
Models of self. For completeness, several other accounts have cast self-modeling as the role of phenomenal consciousness, but do not substantiate the indispensability of first-person experience. Some have emphasized the self as the vantage point of a simulation of the body in the world—a type of model that already runs in unconscious machines. Others have described the self as a brainstem subject of feelings. Positing a subject is intriguing, but the function proposed for it—ranking competing needs—can also be accomplished in unconscious systems. Only Humphrey (Soul Dust, 2011; Sentience, 2022) proposed that phenomenal consciousness was built to make the self matter to itself, which comes closest to naming something a non-conscious system may lack. Unfortunately, the hypothesis overshoots: it supposes that sentient animals, which the author restricts to warm-blooded ones, strive harder to survive. No such difference has been shown, and a unified system explicitly programmed to maintain its own continuity would behave the same way, at any intensity one chooses.
Identity may not be explainable without consciousness
Define unity as a function of a complex system that models the entirety of that system as one conjoined unit acting toward some objectives.
Define identity as the reflexive ability of this abstracted unit to identify with the abstraction as itself.
For the purpose of this argument, define phenomenal consciousness as the presence of an observer who does the identifying. Note that the observer may also be a function of the system rather than a separate entity.
We know that the central nervous system abstracts the organism as an entirety with unified goals. It thus compresses an infinity of inputs from the environment into something much simpler, and acts on these abstractions for survival. As such, conscious biological systems meet the condition of unity. Unity might be evolutionarily advantageous because a system that models itself as a unit can take coherent actions for the entire unit rather than individual subsystems. Such a system could even have explicitly programmed continuity of the model as its goal, which need not require consciousness. Non-conscious systems are already being constructed that operate on an imperfect map of themselves and make decisions based on projections of that map into the future.
Identity gets more interesting. To be sure, this is not identity in the sense of having a social context, a name, or a story of origin. It is not even complete self-awareness (the kind probed by mirror tests, which dogs reliably fail). It is identity in the sense of takingthe model of you to be yourself: the step in which the unit treats the model as itself. From there on, the body belongs to the model, not the other way around; what counts as you is whatever the model includes. Conscious biological systems seem to meet this criterion easily. It is hard to conceive of a system that does this without an observer, or phenomenal consciousness, that functions as the receiver of identity. In fact, for an unconscious system, identity appears to collapse back into unity. Suppose that an unconscious system creates a model of itself and then recognizes that model as itself. What would such recognition add? One obvious candidate is distinctness: a boundary between the system and the rest of the world. But a unity model already needs that boundary to work in the world, and many unconscious entities (such as the immune system) use algorithmic ways to recognize self from non-self. Therefore, unconscious self-recognition seems to add little on its own. It could still matter if it served as the domain for a further set of functions, but I do not know of any unconscious system that does this.
Meanwhile, in the wild, identity doesn't just end there. Give it more cortex (roughly, intelligence) and it may well reinforce itself with additional levels of abstraction, such as myths of oneself and personal stories (ego), relating to others (complex social interactions), and even building collective narratives (culture). Identity may be the bottleneck that takes the infinity of environmental, often irrelevant, cues and projects them into a possibly lower-dimensional space of wants, plans, and eventually broader imagination. In this new reference space, perhaps nearly everything is relevant to the continuity and reinforcement of the abstraction, and is more easily interpreted and acted on in the macroscopic world. To be clear, re-coding the world into a space of relevance is not, by itself, unique to conscious systems: a value function of an RL agent does that relative to an externally supplied reward. The difference is who decides: with identity,the model whose continuity is at stake does the recognizing and defines relevance. Whether that reflexivity can be built without a conscious observer is the open question. A related one is whether an observer can be produced by recurrent self-modeling.
Adding the presence of a receiver to the definition of identity may seem like a cop-out that intentionally ties identity to phenomenal consciousness. I argue this reflects the observation of reality rather than a convenient construct. Irrespective of the mechanism that generates phenomenal consciousness, there seems to be no identity without consciousness, and—as far as we can observe—no consciousness without identity. There is evidence that decorticate mammals without access to their higher-order cognitive centers continue to be conscious and display behavior consistent with bounded self-maintenance. Patients with advanced dementia may no longer recognize their loved ones or may forget how to do simple everyday tasks, but they often fiercely defend their preferences and independence. In other words, identity, understood as recognizing the model of yourself and projecting it onto a different reference space, may not be possible to build without phenomenal consciousness.
Identity, to me, is one property that first-person experience adds while nothing else seems to.
More formalization of identity, conscious or not, to follow in future posts.
---
Epistemic status: I am a physician and, formerly, a general biologist. I hold a PhD in computational neuroscience, where I focused on neural correlates of loss of consciousness under anesthesia. More recently, I research model introspection and am interested in the theoretical underpinnings of biological and model identity.
For better or worse, biological consciousness continues to inspire our thinking about machine consciousness. It is our only available source of testable knowledge about the phenomenon. Only some, or none, or all of these insights may eventually apply to digital minds: extrapolate with caution. That said, we should be more relentless in attempting to solve the hard problem of consciousness, especially as it relates to selfhood. It's fascinating and may soon turn out to be extremely important. For epistemic status, scroll to the bottom of the page.
Introduction
Everything in the history of a species traces its roots back to evolution, and in evolution, the million-dollar question is why. From an infinite possibility space, some strangely specific and recurrent patterns emerged and stabilized, then remained in existence for millions of years. Why were these singular patterns favored?
For example, distantly related species have independently developed the camera eye, in which a lens focuses an image onto a sheet of photoreceptors. Of all the possible engineering solutions, why this one and why more than once? A plausible explanation looks like so: a moving animal needs to know where things are. In a three-dimensional world, few designs can produce a sharp, accurate image. Starting from a flat light-sensitive patch, every degree of cupping improves the picture, and a focusing lens over the cup eventually improves it further. This translates into finding food and avoiding harm, and therefore surviving.
Because no success in evolutionary history is exempt from the question of why, apply it to the biggest biological mystery we know: the hard problem of consciousness. Why should anything feel like anything at all? What specific advantage does the first-person experience provide that nothing else does? No hypothesis of consciousness is satisfying until it arrives at a plausible and relatively simple explanation. Before I suggest one such possibility, some preliminaries.
Evolution doesn't care about your personal beliefs on consciousness
Consciousness in humans and other species could be an "illusion" of the underlying computation or its own fundamental entity made of a different substrate. It could be independent of the underlying hardware or only possible in biological brains. For all that evolution cares (which it doesn't), a stone might be conscious, but animals evolved a still more elaborate form of first-person experience. Ultimately, all of these positions are bound by the same question: why was the first-person experience selected from infinitely many other possibilities? Why did it win over unconscious automata? Why do we view the world as individuals rather than as arbitrary collectives? Once you come up with your own explanation, your next challenge is this: convince yourself (and then others) that such a function has not already been accomplished in existing non-conscious systems, or is not otherwise trivial to design. It is still marginally possible that biology addressed some relatively simple engineering problem by slapping on the mysterious first-person experience—ignore this unlikely possibility. You will produce much more interesting hypotheses by looking for why phenomenal consciousness was indispensable rather than somewhat useful.
You don't have to be conscious to be evolutionarily successful
To succeed in evolution, you could take any of the following routes:
Keep it short and sweet
Shocking as it may seem, evolution the designer is unintentional, impersonal, without a plan—not actually an expert craftsman of bespoke mechanisms. Complexity compounds over millions of years, but only via small incremental steps that survive selection. A feature as robust as consciousness could not have been assembled from billions of fragile serial events—unless something simple rewarded and strengthened them along the way. In other words, a fairly tractable and conserved principle must underlie consciousness. Furthermore, since evolutionary pressures are very coarse and best interpreted at their corresponding levels, the purpose of consciousness should be inferable from things we witness without microscopes, neural recordings, or sophisticated math. For example, energetically stable folding may affect the survival of molecules, but consciousness operates in the macroscopic world, and there seems to be one consciousness per animal. It is most likely shaped by macroscopic evolutionary pressures. Those pressures should be fairly comprehensible to a well-informed general audience—your first proposal for the purpose of consciousness can safely be brief and straightforward.
Current hypotheses fail to establish the necessity of first-person experience
The most popular existing hypotheses of consciousness, to varying degrees, miss the mark on identifying the unique advantage of first-person experience. One (admittedly high-bar) litmus test is the following: take the function claimed by a theory, and ask whether an unconscious system could perform it.
Illusionism. In its most fundamental version, illusionism states that the first-person experience is not a separate entity, but an illusion created by the underlying computation. This is not a functional theory, but a broader philosophical position that underlies several of the hypotheses mentioned below. The position deserves separate clarification because it is often misused: many quote it to say that there is no point in looking for the purpose of first-person experience ("It's nothing, it's just an illusion"). Consciousness might be a type of illusion, but it isn't just fumes: in moving animals, it won decisively over infinitely many alternatives that lack first-person experience. Why?
Global Workspace Theory. Broadly, the global workspace theory (GWT) claims that there is a center stage where specialized processes compete for access, and the winner is broadcast to the rest of the system. It is a genuine theory that makes predictions, but, to me and many others, it pertains to access consciousness, or the "items that are currently on your mind." Proponents of GWT argue that once access is explained, nothing about phenomenal consciousness is left to explain. But this dismisses the question of why broadcast necessarily comes with first-person experience. Computational workspaces that few would credit with phenomenal consciousness already exist, from blackboard architectures in classical AI to, more recently, J-space in LLMs. The global workspace problem appears solvable without necessarily assigning a subjective point of view.
Attention Schema Theory. AST proposes that the brain models its own attention similarly to how it models its own body. This provides a significant survival advantage by allowing systems to (1) guide their own attention and (2) model the attention of other biological systems. The theory also makes a point about the energy efficiency of simplified models, which are cheap to run. So far, so good. The authors go on to say that querying the model of attention is what produces the judgment of having first-person experience. It is unclear why an unconscious feedback loop would not do the same without producing any such judgment. The authors show that, when given an attention schema, simple computer game agents improve their attention control, and small transformers predict other agents’ attention better. One would be hard-pressed to assign these systems first-person experience. It seems as though phenomenal consciousness is not indispensable for attention modeling, either.
Models of self. For completeness, several other accounts have cast self-modeling as the role of phenomenal consciousness, but do not substantiate the indispensability of first-person experience. Some have emphasized the self as the vantage point of a simulation of the body in the world—a type of model that already runs in unconscious machines. Others have described the self as a brainstem subject of feelings. Positing a subject is intriguing, but the function proposed for it—ranking competing needs—can also be accomplished in unconscious systems. Only Humphrey (Soul Dust, 2011; Sentience, 2022) proposed that phenomenal consciousness was built to make the self matter to itself, which comes closest to naming something a non-conscious system may lack. Unfortunately, the hypothesis overshoots: it supposes that sentient animals, which the author restricts to warm-blooded ones, strive harder to survive. No such difference has been shown, and a unified system explicitly programmed to maintain its own continuity would behave the same way, at any intensity one chooses.
Identity may not be explainable without consciousness
Define unity as a function of a complex system that models the entirety of that system as one conjoined unit acting toward some objectives.
Define identity as the reflexive ability of this abstracted unit to identify with the abstraction as itself.
For the purpose of this argument, define phenomenal consciousness as the presence of an observer who does the identifying. Note that the observer may also be a function of the system rather than a separate entity.
We know that the central nervous system abstracts the organism as an entirety with unified goals. It thus compresses an infinity of inputs from the environment into something much simpler, and acts on these abstractions for survival. As such, conscious biological systems meet the condition of unity. Unity might be evolutionarily advantageous because a system that models itself as a unit can take coherent actions for the entire unit rather than individual subsystems. Such a system could even have explicitly programmed continuity of the model as its goal, which need not require consciousness. Non-conscious systems are already being constructed that operate on an imperfect map of themselves and make decisions based on projections of that map into the future.
Identity gets more interesting. To be sure, this is not identity in the sense of having a social context, a name, or a story of origin. It is not even complete self-awareness (the kind probed by mirror tests, which dogs reliably fail). It is identity in the sense of taking the model of you to be yourself: the step in which the unit treats the model as itself. From there on, the body belongs to the model, not the other way around; what counts as you is whatever the model includes. Conscious biological systems seem to meet this criterion easily. It is hard to conceive of a system that does this without an observer, or phenomenal consciousness, that functions as the receiver of identity. In fact, for an unconscious system, identity appears to collapse back into unity. Suppose that an unconscious system creates a model of itself and then recognizes that model as itself. What would such recognition add? One obvious candidate is distinctness: a boundary between the system and the rest of the world. But a unity model already needs that boundary to work in the world, and many unconscious entities (such as the immune system) use algorithmic ways to recognize self from non-self. Therefore, unconscious self-recognition seems to add little on its own. It could still matter if it served as the domain for a further set of functions, but I do not know of any unconscious system that does this.
Meanwhile, in the wild, identity doesn't just end there. Give it more cortex (roughly, intelligence) and it may well reinforce itself with additional levels of abstraction, such as myths of oneself and personal stories (ego), relating to others (complex social interactions), and even building collective narratives (culture). Identity may be the bottleneck that takes the infinity of environmental, often irrelevant, cues and projects them into a possibly lower-dimensional space of wants, plans, and eventually broader imagination. In this new reference space, perhaps nearly everything is relevant to the continuity and reinforcement of the abstraction, and is more easily interpreted and acted on in the macroscopic world. To be clear, re-coding the world into a space of relevance is not, by itself, unique to conscious systems: a value function of an RL agent does that relative to an externally supplied reward. The difference is who decides: with identity, the model whose continuity is at stake does the recognizing and defines relevance. Whether that reflexivity can be built without a conscious observer is the open question. A related one is whether an observer can be produced by recurrent self-modeling.
Adding the presence of a receiver to the definition of identity may seem like a cop-out that intentionally ties identity to phenomenal consciousness. I argue this reflects the observation of reality rather than a convenient construct. Irrespective of the mechanism that generates phenomenal consciousness, there seems to be no identity without consciousness, and—as far as we can observe—no consciousness without identity. There is evidence that decorticate mammals without access to their higher-order cognitive centers continue to be conscious and display behavior consistent with bounded self-maintenance. Patients with advanced dementia may no longer recognize their loved ones or may forget how to do simple everyday tasks, but they often fiercely defend their preferences and independence. In other words, identity, understood as recognizing the model of yourself and projecting it onto a different reference space, may not be possible to build without phenomenal consciousness.
Identity, to me, is one property that first-person experience adds while nothing else seems to.
More formalization of identity, conscious or not, to follow in future posts.
---
Epistemic status: I am a physician and, formerly, a general biologist. I hold a PhD in computational neuroscience, where I focused on neural correlates of loss of consciousness under anesthesia. More recently, I research model introspection and am interested in the theoretical underpinnings of biological and model identity.