Dissociating Artificial Intelligence from Artificial Consciousness

TL;DR

It is demonstrated that, according to IIT, it is possible for a digital computer to simulate the authors' behavior, possibly even by simulating the neurons in their brain, without replicating their experience.

Abstract

Developments in machine learning and computing power suggest that artificial general intelligence is within reach. This raises the question of artificial consciousness: if a computer were to be functionally equivalent to a human, being able to do all we do, would it experience sights, sounds, and thoughts, as we do when we are conscious? Answering this question in a principled manner can only be done on the basis of a theory of consciousness that is grounded in phenomenology and that states the necessary and sufficient conditions for any system, evolved or engineered, to support subjective experience. Here we employ Integrated Information Theory (IIT), which provides principled tools to determine whether a system is conscious, to what degree, and the content of its experience. We consider pairs of systems constituted of simple Boolean units, one of which -- a basic stored-program computer -- simulates the other with full functional equivalence. By applying the principles of IIT, we demonstrate that (i) two systems can be functionally equivalent without being phenomenally equivalent, and (ii) that this conclusion is not dependent on the simulated system's function. We further demonstrate that, according to IIT, it is possible for a digital computer to simulate our behavior, possibly even by simulating the neurons in our brain, without replicating our experience. This contrasts sharply with computational functionalism, the thesis that performing computations of the right kind is necessary and sufficient for consciousness.

Figures (22)

• Figure 1: A target system for simulation. (A) PQRS is a substrate of four units with binary states that update synchronously at discrete intervals. The system is fully connected: each unit receives input from every unit, including itself. Here, the system is shown in state 0101, also written as $pQrS$. (B) The units P, Q, R, and S implement custom boolean functions of four inputs which, despite not resembling familiar logic gates, are described by a truth table. The truth table yields, for any possible state of PQRS (left), what its subsequent state would be (right). (C) Since all of PQRS's state transitions are deterministic, this truth table is also equivalent to PQRS's Transition Probability Matrix (TPM; right). (D) Applying IIT's postulates to $pQrS$ and its TPM reveals a single, maximally irreducible complex ($\varphi_s(pQrS) = 1.51$ ibits). This complex is, by definition, more irreducible than all other candidate systems over the same substrate, including $QrS$ ($\varphi_s(QrS) = 0.42$ ibits; grey dashes) and $p$ ($\varphi_s(p) = 1.00$ ibits; grey dashes). All other candidate systems (e.g., $pQr$, $pQS$, $prS$, $pQ$, $pr$, etc.; not shown) are fully reducible, with $\varphi_s = 0$. (E--F) The unfolded cause--effect structure of $pQrS$ consists of 13 distinctions ($p$, $Q$, $S$, $pQ$, $pS$, $Qr$, $QS$, $rS$, $pQr$, $pQS$, $prS$, $QrS$, $pQrS$) and 8184 relations among these 13 distinctions. Relations bind state-congruent overlaps among the causes and effects of distinctions. Only 2-relations (edges) and 3-relations (triangles) are shown here. For example (D) , mechanism $pQ$ specifies a maximally irreducible cause $P$ and a maximally irreducible effect $Rs$, yielding a distinction. Distinctions $pQ$ and $pS$ jointly specify overlapping and congruent causes or effects (in this case, a cause $P$), yielding a relation. According to IIT, the cause--effect structure (E) corresponds to the content, or quality, of the experience supported by $pQrS$. The quantity of consciousness is measured by $\Phi$ (here equal to 391.25 ibits), a non-negative number equal to the sum of the irreducibility values ($\varphi$) of all distinctions and relations in its cause--effect structure Albantakis2023.
• Figure 2: Four-bit computer that simulates PQRS indefinitely. (A) A simple computer can be used to simulate PQRS for an arbitrary number of time steps. The computer comprises a clock with frequency dividers, a program that encodes PQRS's transition rules, four one-bit data registers that store PQRS's state, and a multiplexer-like processing unit. Each of the computer's 117 units implements a Boolean function (COPY, AND, OR, or XOR) over its inputs, and is either OFF (white) or ON (black). To reduce visual clutter, colored arrows (teal or tan) are used in place of black arrows to indicate certain connections. For example, the rightmost AND gate in the clock and frequency dividers outputs directly to each AND gate in the data registers. (B) The computer can be programmed by setting the initial state of its data registers P', Q', R', and S' to encode the current state of PQRS. (C) The initial states of the program units are set to encode the state transition rules of PQRS. The data that the program operates on are the current states of P', Q', R', and S'. (D) Each register updates its state every 8th timestep, when the computer begins the next iteration of its simulation. For detailed information about how the computer works, see the \ref{['sec:step_by_step']} supplement.
• Figure 3: Identifying the computer's complexes and unfolding their cause--effect structures. (A) The computer as a whole is not integrated, regardless of its state ($\varphi_s = 0$ ibits), due in part to the presence of purely feedforward modules. Therefore, the computer is not a complex and has no cause--effect structure. Instead, it fragments into twenty-four disjoint small complexes, each one of which is maximally irreducible (shown in blue). None of these complexes specify a cause--effect structure equivalent to that specified by $pQrS$. (B) Shown here for illustrative purposes are the cause--effect structures of nine complexes depicted in panel A. Due to lack of space, the cause--effect structures of fifteen complexes in the program memory are not shown, but they are essentially identical to the one shown (top left). The cause--effect structures of the eight complexes in the clock and data registers are labeled (1-8). All complexes specify cause--effect structures that are simple, with only first-order relations, and not as rich as that of $pQrS$ (see Fig. \ref{['fig:pqrs_intro']}F). (C) Feedback connections can be added to the computer without disrupting its function; the computer's ability to indefinitely simulate any four-unit Boolean system is unimpaired. Each line of the program has a single COPY unit replaced by an OR unit, to which a feedback connection from the instruction register is wired (red; left). Similarly, the clock has a single COPY unit replaced by an OR unit, to which a negated feedback connection from each of the data registers is wired (red; right).
• Figure 4: Identifying a system's intrinsic units based on maximally irreducible cause--effect power. (A) Identifying a system's intrinsic units by macroing over units and updates. Macro states can be defined not only over sets of units (top), but also over sets of updates (bottom). Macroing can be performed recursively across "meso" units and updates. (B) An example of a system whose intrinsic cause--effect power is maximal at a macro grain. There are combinatorially many ways to macro the eight micro units $ABCDEFGH$. Shown here is one possible macroing into two macro units $\alpha\beta$ over two updates. The corresponding macro TPM captures the intrinsic causal powers of $\alpha\beta$ at this grain, and is analyzed using the same formalism for micro systems. $\alpha\beta$ in state $11$ has system integrated information $\varphi_s = 1.118$ ibits, greater than all of the micro-grain candidate systems (max $\varphi_s = 0.135$ ibits, not shown), including the system $ABCDEFGH$ ($\varphi_s = 0.017$). (C) A complex and its intrinsic units must comply with IIT's postulates of physical existence. Consider a system of four units $abcd$, in which there are no connections between vertical neighbors. Each micro unit has $\varphi_s = 1$ ibits on its own (left), while each pair of vertical neighbors has $\varphi_s = 0$ ibits (middle). Because each pair of vertical neighbors is reducible, they do not satisfy the integration postulate, and are not valid macro units (right). (D) Macro units must also be definite, satisfying the exclusion postulate. This translates into the requirement that a macro unit must be "maximally irreducible within" (have greater $\varphi_s$ than any subset of its constituents) Marshall2024. Consider $abcd$ when weak connections are introduced between vertical neighbors (left). Although each pair of vertical neighbors is now weakly integrated with $\varphi_s = 0.01$ ibits (middle), they are not maximally irreducible within (e.g., $\varphi_s(ac) < \varphi_s(a)$). Thus, the vertical neighbors are not valid macro elements (right). IIT's postulates of physical existence do not allow for something (a unit) to be built out of nothing (panel C) or nearly nothing (panel D). Adapted from Marshall2024.
• Figure 5: The computer does not replicate the target's cause--effect structure at any macro grain. One might consider macroing the computer as shown: Each line of the stored program and its corresponding instruction register unit is grouped into its own unit ($\alpha$-$\pi$), the multiplexer units are grouped together ($\omega$), and each data register is grouped into its own unit ($\rho$, $\sigma$, $\varsigma$, $\upsilon$). However, not all of these units qualify as intrinsic units (see text). Moreover, integration occurring in the background conditions, extrinsic to a candidate complex, does not contribute to its irreducibility Marshall2024. The functional integration among the units constituting the candidate complex $\rho\sigma\varsigma\upsilon$ is entirely mediated by the multiplexer, and must be discounted (see text).

Theorems & Definitions (79)

• Definition 1: Micro connectivity
• Definition 2: Paths
• Definition 3: Generalized connectivity
• Definition 4: Generalized connectivity matrix
• Definition 5: Cut matrix
• Definition 6: $\Gamma$-notation
• Claim 1
• proof
• Definition 7: Externally determined units
• Claim 2