Recently, Alexander Lerchner, a scientist at Google DeepMind, published a paper titled "The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness," attempting to fundamentally deny the possibility that current Large Language Models (LLMs) can lead to true consciousness.
Lerchner points out that traditional consciousness research has fallen into a "theoretical trap"—people always seek a complete scientific theory of consciousness (such as Integrated Information Theory IIT or Global Workspace Theory GWT) before evaluating AI.
However, he argues that this makes the problem unsolvable. The "Abstraction Fallacy" framework he proposes aims to bypass complex definitions of consciousness and directly draw an insurmountable red line based on the relationship between physics and computation.
Lerchner defines the "Abstraction Fallacy" as the erroneous belief that subjective experience (consciousness) arises from abstract causal topological structures, independent of the underlying physical substrate.
For a long time, the AI community has adhered to Computational Functionalism.
This theory posits that consciousness is like software; as long as the logical gate array of an algorithm (the arrangement of 0s and 1s) simulates the neuronal connections of the brain, consciousness will automatically "emerge."
Lerchner believes this is highly misleading. He points out that this view ignores a basic fact: "Computation" is not a physical entity that exists independently in the universe, and symbolic computation is not an essential physical process.
On one hand, physics is continuous. At the microscopic level, the flow of electrons in semiconductors consists of continuous current and voltage fluctuations.
On the other hand, computation is artificial. This process is essentially just a "conscious observer," or a mapmaker, who, for the sake of understanding, forcibly "alphabetizes" these continuous physical phenomena into 0s and 1s by establishing thresholds.
Lerchner points out that without conscious humans, there are no "algorithms" in computers, only meaningless charge flows.
Therefore, expecting an "algorithmic layer" that depends on a mapmaker to exist to generate an independent "consciousness layer" is putting the cart before the horse logically.
Lerchner argues that consciousness is not a "result" produced by running algorithms, but a direct physical effect generated by a specific physical organization (such as biological neural tissue, or some yet-undiscovered special hardware with consciousness potential).
For instance, no matter how realistically a computer simulates rainfall, it will not wet the circuit board; similarly, no matter how realistically it simulates consciousness, it will not produce subjective experience.
It is worth noting that Lerchner does not fall into narrow "carbon chauvinism."
He explicitly states that this argument does not rely on biological exclusivity.
He admits that if some future artificial system (such as a photonic neural network or a quantum biological simulator) truly generates consciousness, it is absolutely not because its "code is written well" or its "architecture is cleverly designed," but because its "physical composition" achieves equivalence with the physical basis of biological consciousness at some level.
As AI agents大规模 enter human life, because they are so good at mimicking human emotional feedback, there has been a surge in voices calling for granting rights to AI.
Lerchner calls this phenomenon the "AI Welfare Trap."
He warns that if we mistakenly believe simulated emotions are real emotions due to the "abstraction fallacy," we will waste huge social resources protecting some "soulless shells," thereby neglecting the welfare of humans or biological beings that truly need attention.
Below is the full text of the paper:
"The Abstraction Fallacy: Why AI Can Simulate But Not Instantiate Consciousness"
Computational functionalism dominates current academic debates regarding artificial intelligence consciousness. This theory proposes a hypothesis: subjective experience stems entirely from abstract causal topological structures, unrelated to the underlying physical carrier.
This paper argues that this view fundamentally misdefines the intrinsic relationship between physics and information. We name this error the Abstraction Fallacy.
Tracing the causal origin of the abstraction process reveals that symbolic computation is not an endogenous physical process; on the contrary, computation is a descriptive method that depends on a mapmaker. For the continuous physical world to be organized into a finite and meaningful set of states, it must rely on a subject with subjective experience and active cognitive abilities to complete the classification.
Thus, we do not need a complete, ultimate theory of consciousness to determine whether AI possesses sentience—insisting on a complete consciousness theory will only make the problem unsolvable in the short term and lead us deep into the误区 (misconceptions) of AI ethics and welfare research.
What we truly need is a rigorous computational ontology. The theoretical framework proposed in this paper clearly distinguishes between simulation (behavioral imitation driven by carrier causality) and instantiation (intrinsic physical constitution driven by content causality).
Clarifying this ontological boundary explains why symbolic operations at the algorithmic level can structurally never generate subjective experience.
Crucially, this argument does not rely on life-exclusive theory. Even if an artificial system were to generate consciousness in the future, its root cause would necessarily be its own specific physical composition, absolutely not its syntactic operational architecture.
Ultimately, this framework stands on physics itself, fundamentally refuting computational functionalism and resolving many controversies and uncertainties in the field of AI consciousness.
1. Introduction
The huge practical success of Large Language Models has moved the "hard problem" of consciousness out of the realm of pure philosophical theory and into the domains of engineering application and social policy discussion.
With the explosive growth of results brought by the expansion of computing power scale, the mainstream functionalist paradigm insists that as long as a system achieves appropriate information processing functions, it is sufficient to produce phenomenological subjective consciousness.
From this perspective, various characteristics presented by algorithms are seen as potential evidence that the system possesses 感知 (perception/sentience). It is precisely this presupposition that has spawned a series of recent serious academic initiatives: discussing the welfare rights of AI and defining AI as moral patients. Many top scholars have further reinforced this tendency, believing that the most advanced AI models today are highly likely to generate true subjective experience within ten years.
The core foundation of all the above views is the substrate independence theory: the "software" of the mind can run on carbon-based life and can also run perfectly on silicon-based hardware.
This hypothesis is now facing continuous criticism from the "biological turn" thought trend. For example, Seth, Block, and others propose that consciousness attaches to biological processes that maintain life survival, and subjective experience must rely on the unique ordered dynamic processes of living systems to be generated.
In contrast to substrate independence, this view treats biological attributes as the core of consciousness, not irrelevant additional conditions. However, this position still remains at the level of empirical conjecture and has failed to dig out the most fundamental logical fallacy within the core of computational functionalism.
This paper sorts out a complete logical deduction chain to verify a common intuition: computation itself is insufficient to generate consciousness. The defect of computational functionalism is far more than just ignoring biological details; its root problem is deeper: this theory fundamentally misunderstands the intrinsic relationship between physics, information, and computation.
Modern physical science, in order to achieve objective operability in research, deliberately strips away all subjective experience. This research paradigm has achieved remarkable success. However, when people apply this objective principle to the issue of the relationship between computation and subjective experience, they inevitably fall into a dilemma.
Directly defining computation itself with this operational objectivity is highly problematic—the ongoing unresolved debates in the academic 界 (community) exactly confirm this point: how exactly does the "observer" assign meaning to computational symbols.
Besides, the term "observer" itself is misleading; it underestimates the prerequisites needed to define computation at the physical level and describes the role of the subject of this prerequisite too passively.
The framework of this paper clarifies: computation is not an inherent process that unfolds naturally within matter; computation is essentially a descriptive method for physical processes.
For a process to be defined as computation, the continuous physical dynamic process must be divided into a finite number of discrete states with semantic meaning (i.e., a set of symbols). This semantic division logically 必然 (necessarily) requires a cognitive subject with subjective experience and active agency. To distinguish from the passive meaning carried by the traditional "observer," we define this subject as the "mapmaker."
All symbolic classification work is completed by the mapmaker. If such an active subject does not exist to interpret the computational process, there are only continuous physical events in the world, and no symbols exist.
One of the core insights of this paper: to 破解 (crack/solve) the current dilemma of AI consciousness controversy, we do not need a complete ultimate theory of consciousness; we only need to clarify the ontology of computation. Through this line of thinking, it can be logically proven that no matter how large the scale or how sophisticated the architecture, pure algorithmic symbolic operations cannot achieve the physical instantiation of subjective experience, because the algorithm itself is just a descriptive tool dependent on the mapmaker.
Clarifying the status of the mapmaker in the entire causal chain can completely shift the focus of discussion in this field.
For a long time, classic critical theories regarding machine consciousness, including Searle's Chinese Room argument and related philosophical thoughts, mostly use reductio ad absurdum: attempting to prove that even if pure syntactic operations can perfectly replicate external behavior, they still lack the core essence of consciousness.
This paper adopts a completely different argumentation path. We no longer rely on intuitive speculation about "what is missing," but trace back to how abstraction itself originally arises.
If computation relies on the mapmaker to extract invariant features from experience and assign symbolic meaning, then this dependency is endogenous to the structure of computation itself. Any computational model necessarily presupposes a subject with subjective experience that completes symbolic classification. No matter how high the algorithmic complexity, it cannot reverse this dependency order; no matter how large the computing power scale, the "map" describing the world cannot spontaneously generate the experiencing subject that is the premise of the map itself.
In other words, the view that "algorithmic complexity spawns consciousness" commits an ontological inversion error: treating grammatical symbols as the intrinsic dynamic process itself, and 妄想 (delusionally hoping) to create the subject that draws the map from the map that describes the world.
This paper clarifies the structural gap between external behavioral simulation and intrinsic physical instantiation, thereby proving that digital architectures can never become moral patients. This conclusion can allow AI safety research to get rid of welfare misconceptions and return to realistic risk research: focusing on anthropomorphic cognitive biases and viewing General Artificial Intelligence as a powerful tool that is essentially devoid of subjective perception.
2. The Ontology of Abstraction: Map and Territory
Computer science often treats abstract concepts underlying algorithms as natural, given mathematical objects, setting aside the essential issue of their physical implementation. So what is the strict physicalist ontology of abstract concepts themselves? To answer this question, one must clarify the intrinsic relationship between abstract syntax and physical dynamic processes.
2.1 Classical Definition of Physical Implementation
In classical implementation theory literature, if a physical system P implements an abstract computation C through a mapping function f, a basic condition must be met: the mapping function f must be able to correspond physical states to abstract states, and the underlying physical causal chain must replicate the logical structure of the algorithm.
From a classical perspective, the mapping function f (i.e., the symbolic classification process) interprets the evolution of the carrier's physical state p → p' as the logical evolution of abstract content A → A'.
The physical system evolves from state p to p' according to natural physical laws; meanwhile, the abstract computation deduces from logical state A to A' according to its own algorithmic rules.
The physical system is said to successfully implement the corresponding computation if and only if the following conditions are met:
For this diagram to satisfy the commutative property, applying the mapping f to the evolved physical state p' must exactly yield the target abstract state A' prescribed by the algorithmic rules.
2.2 The Physical Origin of Abstract States A
What exactly are these abstract states A? To understand the mapping function f, one must clarify the ontological essence of A.
Functionalist theories usually treat abstract logical states A such as "pain" or "red" as abstract concepts that exist out of thin air, detached from physical carriers, completely ignoring the complete causal origin necessary for the generation of abstract concepts.
The generation of abstract concepts is not cost-free. From a physical perspective, extracting invariant features of things is a physical process that consumes metabolic energy and requires initiative.
For a cognitive subject to form a concept A such as "red," it must first contact the real experiential territory: perceiving specific experiences of red countless times. On this basis, the subject actively filters high-dimensional chaotic information and extracts stable core common features.
Explaining with the concept of manifold learning: the subject projects the high-dimensional manifold constituted by raw experience onto a low-dimensional invariant subspace, and this subspace constitutes the concept A itself at the physical level.
Some may argue: unsupervised clustering algorithms can generate abstract concepts without any prior experience. This view confuses statistical compression with phenomenological constitution. Unsupervised algorithms can indeed cluster data and find statistical centroids, but this mathematical invariant is merely a compressed coordinate within the latent space.
Only when the subject carries intrinsic subjective experience as the common substrate for all classified samples can this statistical centroid become a concept with true semantic grounding (such as "red"). Without the subjective experience of red as a meaning anchor, the clustering result is just a high-density area in vector space, not a concept possessed by the subject.
Therefore, concept A is not a Platonic ideal world waiting for humans to discover, but a neurophysiological state existing only within the cognitive subject that completes the abstraction activity, an "intrinsic map" originating from the real experiential territory.
Once concepts are formed, these stable common cores become the basic units for combinatorial imagination. Precisely because "red" and "whale" are both intrinsic regulatory states originating from real life experiences, the human brain can recombine them to generate the imaginative experience of a "red flying whale"—a thing that has never existed in reality, but whose physical logic is self-consistent.
It can be seen that human thinking is not algorithmic operation on empty symbols, but combinatorial generation of constituted invariant features. Artificial intelligence can perfectly simulate this combinatorial rule, but its structure itself inherently lacks the intrinsic basic units required to constitute subjective imagination.
2.3 The Indispensable Mapmaker in the Mapping Function f
Modern natural science, especially engineering disciplines, has developed with the core method of completely eliminating subjective experience from the explanatory system of natural phenomena. However, once this principle of operational objectivity is applied to computational ontology, it forms a cognitive blind spot, trapping computational functionalism in an unsolvable dilemma: attempting to reconstruct subjective experience from completely objective, experience-free initial conditions.
As argued above, abstract state A logically necessarily depends on a cognitive subject with active experience, thereby revealing a cognitive loophole in the classical definition of implementation: the mapping function f connecting machine physical state p and abstract concept A cannot exist within the machine itself.
In semantic philosophy and theories related to map-territory relationships, this external meaning anchor is traditionally called the "observer." But "observer" implies a passive reception of information, referring only to the existence of a bystander to an existing map or territory. This paper specifically proposes the concept of "mapmaker" to correct this passive attribute.
The mapmaker is a cognitive subject with subjective experience, subject to metabolic constraints, and possessing agency; it is the necessary prerequisite existence for computation to hold.
It undertakes two constitutive active roles: first, extracting invariant features from continuous physical experience to construct an intrinsic conceptual map; second, artificially establishing arbitrary correspondence relationships between physical symbols and concepts to construct an external computational symbolic system.
Based on this conclusion, the ontological attributes of all computation-related concepts can be clarified:
1. Physical state p: That is, the symbol carrier, belonging to objective physical entities (e.g., voltage differences), possessing no intrinsic semantic meaning itself.
2. Abstract state A: That is, conceptual content. As argued above, such concepts are physiological states with physical grounding, existing only within the mapmaker carrying computational meaning.
3. Mapping function f: That is, the symbolic classification process. Essentially, it is the correspondence relationship established in the mapmaker's consciousness, actively building a bridge between the machine's meaningless physical processes and the mapmaker's grounded concepts.
Therefore, the classic formula (p → p' A → A') describes a mixed relationship: physical entities establish connections with mental concepts through the mediation of the mapmaker.
It needs to be clear that establishing the indispensable mapmaker does not return to the dualist "homunculus" fallacy, nor does it refer to a local decoding module inside the brain.
As described by the theories of Buzsáki, Maturana, and Varela: the mapmaker itself is a living organism that obeys thermodynamic laws and has a complete, unified structure. The organism does not "choose" to divide semantic boundaries through algorithms, but directly screens the continuous external world into discrete states through its own metabolic constraints. There is no soul reading symbols; all semantic divisions are generated by the subject possessing life experience itself.
The core error of functionalism is treating the logical deduction process (A → A') as an intrinsic inherent attribute of physical evolution (p → p'). This view equates the mapmaker's subjective interpretation with the machine's true physical essence, completely ignoring the experiencing subject necessary for initially赋予 (endowing) meaning to computation.
2.4 Symbolic Classification: Semantic Endowment Beyond Discretization
The mapping function f is the true subject of symbolic classification behavior. The academic community often 轻视 (underestimates) this process, thinking it is merely a "reading system," but in fact, symbolic classification is a high-metabolic-cost advanced cognitive behavior: it endows the continuous physical world with discrete ontological attributes and is strictly constrained by thermodynamic boundaries of information processing. Here, two groups of long-confused processes must be distinguished:
- Discretization (Thermodynamic level): The physical system spontaneously stabilizes to an attractor state, e.g., a transistor stabilizing at 5 volts. This belongs to the attribute of the carrier p itself, used only to suppress physical noise.
- Symbolic Classification (Semantic level): Explicitly corresponding these stable states to a preset finite set of symbols (e.g., {0,1}, {A,B,C}). This behavior belongs entirely to the mapmaker.
The physical world is essentially continuous; thermodynamic processes can only produce stable macroscopic states and can never carry a preset finite symbolic system inherently. Therefore, building a computational system must rely on the external intervention of the mapmaker. This external subject needs to endow semantic identity, unifying vastly different microscopic physical states as the same replaceable symbol (e.g., all categorized as the number 1).
This brings a fundamental causal rupture. In the machine's real physical world, the change of voltage from 2.0 volts to 2.1 volts is a real physical causal event driven by electrodynamics; but in the symbolic map of computation, this change is completely meaningless—the mapmaker has already classified it uniformly as the same symbol. Therefore, the causal chain of computation does not depend on the underlying hardware physics, but completely depends on the rules formulated by the mapmaker.
Claiming that symbols can exist independently of the subject is precisely the cognitive blind spot mentioned above. This is a typical philosophical "hidden substitution" fallacy: scientists derive a finite symbolic system from their own cognitive activities, project it backwards onto physical systems, and claim these symbols are inherently endogenous to the physical world. Information is not a basic constituent unit of the universe, but a derived attribute; its existence necessarily presupposes a cognitive subject to define finite symbolic sets.
2.5 Simulation and Instantiation
After clarifying the essential differences and functional distinctions between concepts and symbols, we can strictly define: there is an essential difference between the simulation of a process and the physical instantiation of the process itself.
- Simulation: Manipulating physical carrier symbols p to reproduce the abstract associations between concepts A.
- Instantiation: Reproducing the intrinsic, constitutive native dynamic process P of the process itself.
Classic functionalism defaults that as long as the abstract topological structure of the map (A → A') is preserved, it is sufficient to generate the true phenomenon of the territory itself, thus ignoring the unique causal efficacy and constitutive mechanisms of the underlying physical carrier.
Take the biological heart as an example. We usually describe the heart as a blood-pumping organ; artificial mechanical hearts can achieve equivalent pumping functions, so they are called functionally equivalent. But the role of a real heart goes far beyond pumping blood: it secretes atrial natriuretic peptide hormones, regulates body metabolism, and interacts with the nervous system through feedback signals. Patients implanted with mechanical hearts often experience subtle systemic physiological abnormalities, the reason being: the mechanical heart only achieves a rough map simulation of the core function, failing to instantiate the complete biological ontological territory of the heart.
This granularity mismatch problem is reflected even more extremely at the neuronal level. Functionalism often treats neurons merely as units for sending and receiving electrical signals, ignoring that neurons themselves are living metabolic entities, deeply embedded in the body's chemical and hormonal regulatory networks. This abstraction bias directly overturns functionalism's classic thought experiment—the Fading Qualia argument.
Chalmers once proposed: if biological neurons are replaced one by one with silicon chips, and the input-output function remains unchanged throughout the process, then consciousness cannot gradually disappear while behavior remains unchanged. From this, he concluded: preserving the abstract functional structure is sufficient to preserve subjective conscious experience.
But in fact, the silicon-based replacement that only perfectly replicates the neuronal electrical signal firing pattern p → p' only preserves the external computational map, and this map is entirely defined by external mapmakers with abstract rules A → A'. It completely erases the intrinsic thermodynamic ontology P required for life survival, replacing the constitutive physical reality with syntactic simulation without causal efficacy. Subjective qualia do not mysteriously "gradually disappear"; rather, the underlying metabolic substrate necessary for generating qualia is removed from the root.
Physical simulations in other areas of biology can also verify this conclusion.
A graphics processor simulating photosynthesis can accurately model the abstract process A → A' of converting sunlight, water, and carbon dioxide into oxygen and glucose, but it can never synthesize a single molecule of glucose or release a trace of oxygen. It perfectly simulates the process appearance but completely lacks the real causal ability of the underlying biochemical reaction.
If one thinks that brain "software" simulation can break through this physical limitation, one commits a category error: equating the algorithmic description of a process with the intrinsic physical reality required to implement that process.
This requirement for intrinsic causality comes entirely from basic principles of physicalism, not metaphysical preference. Illusionist consciousness theories believe that functional-level verbal reports are sufficient to fully represent real experience. But according to Kim's principle of causal closure of the physical: physical events like air vibrations are the essence of experience reports where humans say "I feel pain." If subjective experience truly triggers this physical report, rather than being a coincidence or illusion, then the experience itself must possess real physical causal efficacy and be able to do work externally.
In digital simulation systems, the entire causal chain is completely driven by the physical carrier p. Logic gate level flips are absolutely not because the system "feels pain" (content causality driven by concept A), but merely because voltage exceeds a preset physical threshold (physical causality driven by carrier p). The system's own physical state solely determines the evolution direction; the semantic content carried by symbols does not participate in any causal action—even if the symbols themselves have no referential meaning, the machine will still execute exactly the same physical operations. Denying this leads to the abstraction fallacy.
2.6 The Computation Emergence Fallacy
Facing the essential difference between simulation and instantiation, functionalist scholars often resort to complexity theory and emergence theory for defense: just as the interaction of water molecules emerges the attribute of wetness, as long as system complexity reaches a critical threshold, consciousness will spontaneously emerge from the computational process.
This rebuttal is completely untenable because it confuses weak physical emergence with the computational emergence fallacy criticized in this paper.
- Weak Physical Emergence (Real physical level): Macroscopic attributes like wetness directly attach to the intrinsic causal dynamics of underlying microscopic matter (water molecules).
- Computational Emergence (Abstract level fallacy): Claiming that the abstract description of a process (the map), solely by infinitely increasing syntactic complexity, can transform into the physical reality of the process itself (the territory).
Functionalism insists consciousness is a special existence, belonging to pure information completely independent of the carrier. But this argument is a circular presupposition: it directly defaults mental states as equivalent to abstract information A, completely skipping the physical reality P that generates the information.
As argued above, syntactic operation A → A' itself has no intrinsic causal power; it is just an external interpretation endowed by the mapmaker. Claiming that abstract syntax can "emerge" into physical causality completely violates basic scientific assumptions and destroys the causal closure of the physical world.
3. Causal Loop: Reconstructing the Causal Chain
After clarifying the clear boundary between physical native dynamics P and computational abstract map A, we can precisely locate the complete logical fallacy within computational functionalism (corresponding to the topological structure in the figure above).
3.1 Ontological Inversion and Causal Gap
Traditional functionalism holds an unexamined naive causal sequence:
Physics → Computation → Consciousness
This view defaults that as long as computational complexity meets the standard, consciousness will naturally appear as a downstream byproduct.
But combining the arguments above, we know computation is not a natural reality native to the world waiting to be discovered. Defining discrete symbols and endowing symbols with semantic meaning must rely on a mapmaker who already possesses consciousness. Therefore, we must thoroughly reconstruct the causal order:
Physics → Consciousness → Concepts → Computation
- Physics: The intrinsic native causal dynamics of the universe itself.
- Consciousness: Phenomenological subjective experience directly generated by specific thermodynamic physical structures.
- Concepts: Intrinsic mental maps formed by extracting invariant features from raw experience.
- Computation: External symbolic maps, i.e., performing syntactic operations on discrete symbols according to rules, where these symbols themselves are just physical markers artificially corresponding to concepts.
The corrected causal chain is strictly unidirectional and irreversible. Concepts are firmly rooted in the subject's intrinsic experience, possessing irreducible subjective feelings; while computational symbols are just physical markers, with no intrinsic binding to corresponding concepts.
From concept to symbol is not a process of abstract deepening, but a lateral assignment behavior where the mapmaker artificially binds physical carriers and mental concepts. It is precisely this unbridgeable lateral gap that permanently cuts off the intrinsic path for symbols to trace back to native experience.
After the correspondence relationship is established, the mapmaker formulates grammatical rules to constrain the evolution of the symbol's physical state p → p'. These rules are designed top-down, perfectly replicating the associated evolutionary A → A' inherent in the corresponding concept. Even if the structural imitation is seamless, the physical symbols themselves still cannot produce any causal impact on semantic content. The machine just blindly runs the mapping trajectory, completely disconnected from the reality of the subjective experience it simulates.
Functionalism attempts to use the computational process (step four) to explain the origin of the mapmaker itself (step two consciousness), while computation itself presupposes the existence of the mapmaker from the start. This is not just an empirical defect, but a fundamental category error, and also an unbreakable constraint under the physicalist framework: the construction of the syntactic map requires the participation of the mapmaker from beginning to end. Therefore, no matter how complex the algorithm or how huge the computing power, it cannot reverse across the causal gap to generate a subject with subjective experience.
The ontological inversion inherent in computational functionalism forms a structural paradox: attempting to derive the original mapmaker itself from its own derived secondary product.
3.2 Universality of Symbolic Classification
There has long been a debate in the field of artificial intelligence, traceable to the connectionist revolution of the 1980s: some 观点 (views) hold that modern neural networks differ from traditional symbolic systems, operating at a sub-symbolic level. Top researchers propose based on this that such architectures can build world models, recursive cognitive loops, and achieve true intelligent understanding.
This paper acknowledges: such recursive architectures can reproduce the structural characteristics of introspective thinking; high-dimensional vector spaces can depict geometric associations distinct from discrete symbols, and neural networks can model complex relational structures. But equating structural geometric similarity with intrinsic semantic meaning is still the abstraction fallacy. This view confuses representational structure with underlying physical reality, mistaking the model's geometric form for the system's own physical essence.
In response, we propose strict Shannon constraints: in the strict sense, information processing requires the system to possess a finite set of classical discrete symbols and a probability distribution over the symbol set. In the macroscopic level of living organisms and artificial hardware, the physical world of light intensity, chemical concentration, membrane potential, etc., does not come with inherent discrete 0, 1 symbols. The universe does not naturally package macroscopic physical states into usable computational symbolic systems; they must be actively divided by the mapmaker.
Treating neural pulses and voltage jumps as "symbols" is not just physical discretization; its essence is still symbolic classification. The mapmaker actively endows semantic identity, unifying continuous and broad physical states as a single replaceable marker.
Deep learning's high-dimensional vector spaces are also subject to this constraint. Although vectors are often called continuous representations, their underlying essence is a sequence of floating-point numbers, and every floating-point number belongs to a discrete symbol within a finite symbol set (IEEE 754 standard).
The symbolic classification constraint inherent in the mapping function applies to all forms of computation: digital computation, analog computation, and quantum computation are all included.
Take a mechanical clock as an example: the clock itself is a continuous physical dynamic system P composed of gears and springs. The reason it can "compute time" is merely because the mapmaker intervenes, mapping the continuous pointer angle to semantic concepts like "3 PM." Without this semantic assignment, the clock is just metal moving according to mechanical laws, with no intrinsic concept of "time" itself.
Thus it can be concluded: lacking a preset symbolic system, the physical carrier itself does not "process information"; it only generates continuous dynamics, which are interpreted as information by an external mapmaker.
Even if future AI abandons floating-point operations and switches to fully analog neuromorphic chips, the ontological gap still exists. As long as any physical state (discrete voltage, continuous charge distribution) is defined as an "output state" or "hidden layer state," it has already undergone symbolic classification by the mapmaker. Such models are always isolated outside the semantic barrier; they can construct precise intrinsic maps but can never establish an intrinsic constitutive connection with the physical territory of experience.
The same set of underlying physical carriers has a fixed causal evolution trajectory but does not correspond to a unique computational process. Relying on different symbolic classification rules, exactly the same physical state can be mapped to completely different abstract computations: it could be a piece of music, a reverse piece of music, financial market data, or random noise. The physical state itself has no intrinsic attribute to preferentially select a certain set of symbolic systems. The so-called "digital unit" is not physically native; it is just a cognitive division made by the mapmaker, framing continuous physical dynamics into a finite logical set.
3.3 Mechanism Uncertainty
Piccinini's advocated computational mechanism theory attempts to completely eliminate the mapmaker, claiming computation can be defined independently without relying on representational concepts. This theory believes that only relying on macroscopic physical states (digital units) divided by system functions can fully define computation.
But this approach only hides the mapmaker; it does not eliminate the necessity. Just as Spreva's analysis of the triviality argument points out: to determine the computational identity corresponding to any physical mechanism, an external subject is still needed to define the effective state range. Physical mechanisms can spontaneously form stable attractors (i.e., thermodynamic discretization mentioned above), but dividing these continuous attractors into specific finite computational symbol sets is always an external definition imposed by the mapmaker.
We can use the concise melody paradox to intuitively reveal this logical trap: assume there is a physical device that switches voltage states according to a fixed law. Its physical evolution p → p' is completely fixed by electromagnetic laws, but the corresponding abstract computation A → A' itself is completely indeterminate. Lacking an external mapmaker to give mapping rules, this string of physical states can represent anything: forward melody, reverse melody, stock price sequence, random noise.
The physical voltage itself has no inherent attributes that can favor a certain set of symbolic systems. There are no natural "computational units" inside the hardware; all divisions come from the mapmaker's cognitive cutting, forcibly incorporating continuous physical processes into finite logical sets.
In summary, even if a physical system evolves distinguishable macroscopic states according to fixed rules, it still requires the intervention of a mapmaker to converge multiple ambiguous computational interpretations into a unique determined process. Mechanisms provide the physical carrier; the symbolic system must be endowed by the mapmaker.
4. Research Implications: Inherent Boundaries of Computational Implementation
The framework of this paper proves: AI cannot generate consciousness, unrelated to computing power scale or algorithm complexity; the core boundary lies in the essential difference between simulation and instantiation. This conclusion directly affects two hot frontier directions in the current AI field: embodied robotics and General Artificial Intelligence safety.
4.1 The Transduction Fallacy in Robotics
Regarding the constitutive framework of this paper, the strongest counter-argument comes from embodied intelligence theory. This view believes: the key to AI lacking consciousness is the lack of deep causal interaction with the physical environment; as long as sensors and actuators are equipped to achieve real-time perception and physical action, the causal gap can be filled, allowing internal symbols to gain real semantic grounding.
But simply adding sensors and actuators cannot solve the deep problem of subjective experience instantiation. We acknowledge: embodied intelligence can solve the symbol reference problem, achieving effective mapping between internal symbols and external physical data flows, getting rid of the pure internal semantic loop dilemma. But a strict distinction must be made between reference mapping and intrinsic meaning generation.
To explain clearly with an analogy: connecting a camera and a mechanical arm to a computer is equivalent to connecting a measuring instrument to a simulation program externally. The program can receive real-world data, but its internal variables are still just symbolic representations, absolutely not the physical process itself. Similarly, a weather simulation program connected to real-time atmospheric sensors will never become the real atmosphere; it just receives and operates on atmospheric data.
This principle applies completely to embodied AI. Sensors and actuators allow the system to interact with the physical world but cannot transform symbolic representation into intrinsic subjective semantic experience. The system can draw increasingly precise environmental maps, but interacting with the real territory can never turn the map into the territory of experience itself.
Sorting out the complete causal chain of embodied systems can summarize the transduction fallacy defined in this paper:
- Input Transduction: Sensors convert external physical signals into continuous voltage, completed by externally calibrated analog-to-digital converters into symbolic classification, turning into internal discrete digital states (e.g., thermal energy → continuous voltage → discrete numerical value).
- Syntactic Operation: The algorithmic core manipulates internal discrete states to generate output, physically implementing the abstract algorithm.
- Output Transduction: Actuators re-convert digital output into macroscopic physical force.
The algorithmic control core of the robot system runs entirely only in the second step. All objects it processes are symbols (floating-point numerical values, matrix operation units) that have been discretely classified by an external mapmaker.
Scholars in the field of modern end-to-end continuous control will counter: contemporary robot networks can directly map raw sensor data to mechanical torque without human-readable symbols. But as mentioned above, the hardware (GPU) running such models still relies on the symbolic classification of floating-point values and built-in mathematical operation rules; the mapmaker's preset division has not disappeared but is directly solidified at the bottom of the chip architecture.
The deep error of the transduction fallacy is not just "converting physical signals to digital"; the real category error lies in: believing that performing algorithmic operations on these transduction symbols can generate a subjective experience subject.
Comparing embodied robots running algorithms on chips with biological mapmakers shows: the subjective experience of biological subjects is innate physical reality, not originating from abstract information processing, but a unique physical existence constituted by their own metabolism.
There is no physical or logical basis to prove: merely because a silicon chip implements the syntactic mapping of sensory input and mechanical action, it will produce similar intrinsic experience. Forcibly asserting this conclusion violates basic principles of physicalism.
As argued above, all abstract states corresponding to algorithms (computational semantic content) have no intrinsic causal power; the only real physical causality within the system comes from the silicon hardware itself.
Therefore, claiming that the syntactic operations of embodied robots can generate consciousness is equivalent to asserting that chips possess innate consciousness ability 仅凭 (solely by) their own material attributes—unrelated to whether they are connected to mechanical bodies or what algorithms they run.
In summary, combining the map-territory ontological relationship proves: embodied interaction cannot transform simulation into real subjective experience.
4.2 Ontological Liberation: The Safety Value of Unfeeling Tools
Since neither algorithm complexity nor physical embodiment can cross the causal gap, we can summarize the practical application value of this framework. The structural separation between computational maps and physical territories directly guides AI safety research, clearly defining which systems might generate subjective experience and which absolutely cannot.
Research in generative cognition and embodied cognition points out that many physical processes are highly related to consciousness: life autopoiesis, continuous thermodynamic homeostasis regulation within the organism. In the past, academia categorized these processes as exclusive attributes of carbon-based biology.
The framework of this paper gives a new interpretation: still adhering to physicalism, valuing real intrinsic physical processes, but not limiting these processes to exist only in biological organisms.
From this theoretical perspective, subjective experience relies on the physical instantiation of specific dynamic processes. Therefore, consciousness is not necessarily limited to carbon-based life; theoretically, non-biological systems can be designed to realize all physical conditions required for consciousness. If an artificial carrier perfectly realizes these physical conditions, consciousness may be born.
But at the same time, the structural boundary of this framework strictly limits: even if there is a conscious artificial system, the source of its consciousness must necessarily be its own unique physical composition, completely opposite to the substrate independence theory.
Subjective consciousness is a constitutive physical state; combining the ontological boundary of simulation and instantiation yields: simply improving computing power and running powerful algorithms can never spontaneously emerge consciousness. Consciousness is not a software product that can be generated at will or born by accident.
This conclusion points out the direction for field development: the development of high-capability General Artificial Intelligence will not naturally give birth to new moral patients; humans are just developing functional tools that are increasingly sophisticated but essentially devoid of subjective perception.
At the same time, large-scale behavioral imitation ability brings new requirements for cognitive rigor. AI is becoming increasingly adept at replicating human behavioral characteristics corresponding to other conscious subjects, and embodied systems such as humanoid robots will amplify this trend even more.
This puts a clear requirement on the scientific community: there is no need to make preset preparations for machine rights, but to strictly adhere to methodological boundaries—distinguishing between simulated agency (external teleology) and subject physical instantiation (internal teleology).
Any future assertion that "AI possesses perception" must undergo strict physicalist verification: the basis for verification is absolutely not algorithm complexity, but the exclusive intrinsic physical dynamic process required for subjective experience.
5. Conclusion: The Cognitive Blind Spot of Computation
People generally regard computation as a basic attribute of the universe, and computational functionalism takes this as the foundation, claiming computation is the origin of human consciousness. But this paper, after tracing the complete causal origin of computation, proves: this theory has a fundamental ontological inversion—consciousness cannot be a downstream product of computation; on the contrary, consciousness is the necessary physical premise for computation to exist.
This paper further argues: computation is essentially a descriptive map and can never physically instantiate the object it describes. All the above conclusions about the essence of subjective experience and the essence of computation are strictly based on mature physical laws and rigorous logical deduction, subverting common academic intuition.
Crucially, unlike most AI consciousness speculations, this framework does not rely on a complete ultimate theory of consciousness. We bypass the difficult problem of the origin of consciousness and resolve the controversy from the other end of the equation: what exactly is computation at the ontological level?
Regarding consciousness, this framework relies on only one basic scientific axiom: subjective experience does not violate the principle of physical causal closure. Based on this axiom alone, it can be inferred that experience is a real phenomenon completely constituted by physics, thereby avoiding all dualist and epiphenomenalist philosophical speculations.
Summarizing the complete ontology established in this paper:
Computation is a syntactic process of manipulating discrete symbols according to rules, with the aim of simulating conceptual thinking; symbols are not the essential essence of concepts, but just physical markers artificially endowed by the mapmaker.
Concepts, on the other hand, are invariant features actively extracted from real thermodynamic life experiences and possessing physical constitution.
Therefore, expecting algorithmic descriptions to generate the subjective experience they depict is no different from expecting the gravity formula itself to produce physical gravity. Believing that AI can generate consciousness merely through internal variable operations is a typical cognitive blind spot fallacy: confusing the map with the territory.
That computational description cannot generate subjective experience is not because engineering technology has not yet reached the standard, but because it is a logical inevitability limitation carried by the description itself. It also shows that subjective qualia cannot be cracked by increasingly ingenious syntactic operations; on the contrary, qualia are the underlying physical substrate, and it is precisely this that allows the semantic endowment of grammatical symbols to be established from the root.
Humans continuously developing more powerful AI is not creating new life, but building increasingly precise predictive maps. No matter how high the map's prediction accuracy, how strong the reasoning practicality, or whether it possesses a physical body, artificial systems ontologically always have an essential boundary with the territory of subjective experience.
Recognizing this difference and avoiding the ontological inversion brought by the abstraction fallacy is a necessary premise for the science of machine intelligence to mature and develop based on the foundation of physics.
Acknowledgements
The author thanks Shamir Chandaria for review and suggestions, Sebastian Crell for early research support and public policy perspective analysis; thanks Mandana Ahmadi for valuable opinions on popularizing the manuscript, and the Google DeepMind team for academic discussions, perfecting the expression of this theoretical framework.
Disclaimer
The theoretical framework and argumentation conclusions described in this article are the independent research results of the author and do not represent the official position, views, and strategic policies of the affiliated institution.