Consciousness in Artificial Intelligence: Insights from the Science of Consciousness — part 7/14

Cleaned-text code points [93023, 108148). Bibliography omitted here; source.md includes it.

In keeping with the “theory for consciousness” idea, the PP framework has been employed in developments of GWT and HOT. Hohwy (2013) and Whyte (2019) propose that global broadcast (via ignition) takes place when the process of perceptual inference settles on some representation of the state of the environment as most probable, then makes this representation available as the basis for active inference. Meanwhile, the higher-order state space theory adopts the PP framework (Fleming 2020).

#### 2.4.3 Midbrain Theory

While the neuroscientific theories of consciousness we have discussed so far focus primarily on cortical processes, Merker (2007) argues that the cortex is not necessary for consciousness. This view has been particularly influential in recent discussions of consciousness in non-human animals. Merker’s proposal is that activity in parts of the midbrain and basal ganglia constitute a “unified multimodal neural model of the agent within its environment, which is weighted by the current needs and state of the agent” (Klein & Barron 2016), and that this activity is sufficient for subjective experience. In Merker’s account, one midbrain region, the superior colliculus, integrates information from spatial senses and the vestibular system to construct a model of the organism’s position and movement in space. Other regions including the hypothalamus, periaqueductal gray, and parts of the basal ganglia bring in information about the organism’s physiological state and contribute to identifying opportunities and selecting actions. In functional terms, the midbrain theory claims that consciousness depends on “integrated spatiotemporal modeling” for action selection (Klein & Barron 2016).

Birch (2022b) summarises and criticises evidence for the midbrain theory. For our purposes, it is notable because it offers a particular perspective on the significance of cognitive integration for consciousness. The midbrain theory proposes that the integration which is necessary for consciousness arose to solve the biologically ancient problem of decision-making in complex mobile animals, particularly the need to distinguish the effects of self-caused motion on perceptual input (Merker 2005). Hence the midbrain theory emphasises the need to integrate specific kinds of information, such as spatial, affective, and homeostatic information, into a single common model. This theory, therefore, gives us additional reason to believe that systems for purposeful navigation of a body through space are necessary for consciousness, and thus contribute to the case for the indicators we present in section 2.4.5.

#### 2.4.4 Unlimited Associative Learning

Another influential theory in animal consciousness literature is Ginsburg and Jablonka’s (2019) unlimited associative learning framework (Birch et al. 2020). The proposal here is that the capacity for unlimited associative learning (UAL) is an evolutionary “transition marker” for consciousness: a single feature that indicates that an evolutionary transition to consciousness has taken place in a given lineage. Ginsburg and Jabolanka identify UAL as a marker on the grounds that it requires a combination of several “hallmarks” that are argued to be jointly sufficient for consciousness in living organisms. The UAL framework, therefore, brings together a list of features that seem to be related to consciousness and argues that they are unified by facilitating UAL.

Somewhat like this report, the UAL project aims to identify a set of jointly sufficient conditions for consciousness. This set is summarised in the table below (which follows the presentation of these conditions in Birch et al. 2020).

Hallmarks of consciousness according to UAL Global accessibility integrating sensory, evaluative and mnemonic information Selective attention Integration over time through forms of short-term memory Embodiment and agency Self-other registration, used in constructing a representation of the moving body in space Flexible value system capable of revaluation and weighing needs Binding/unification of features to form compound stimuli, enabling discrimination of complex patterns Intentionality, i.e. representation of the body and environment

This list of conditions is similar to the conditions which are emerging from our work in this section (like us, Ginsburg and Jablonka examined scientific theories of consciousness in developing their list). The global accessibility and selective attention conditions are met by systems that implement global workspace architectures. The integration over time condition may be as well; we discuss integration over time in section 2.4.6. Several of the other conditions are related to agency and embodiment, which we discuss in section 2.4.5. These include not only the embodiment and agency condition itself but also self-other registration (which also aligns with the midbrain theory) and having a flexible value system. We take it that the binding/unification condition will be met by any neural network-based AI system because these systems are designed to learn to discriminate complex patterns. The final condition, that the system should represent its body and the environment, raises philosophical questions about intentionality which we will not go into here. But any naturalistic theory of intentionality is likely to entail that systems that meet the other conditions will also meet this one because the other conditions entail the performance of the functions which might ground intentionality.

What about the capacity for unlimited associative learning itself? This could be argued to be an indicator of consciousness in artificial systems. In the UAL framework, it is characterised as an open-ended capacity for associative learning. In particular, to have this capacity an organism must be capable of conditioning with compound stimuli and novel stimuli and of quickly and flexibly updating the values it associates with stimuli, actions and outcomes. It must also be capable of second-order conditioning, meaning that it can link together chains of associations, and trace conditioning, which is learning an association when there is a time gap between stimuli. Having the capacity for unlimited associated learning, therefore, implies that an organism or system can integrate information across modalities and over time, and that can evaluate stimuli and change these evaluations in response to new information.

This capacity is also, therefore, linked to many of our indicators. Integration of information from different modalities and flexible learning are emphasised by GWT. PRM also suggests similar capacities by requiring a general-purpose belief-formation and decision-making system, which receives input from any sensory modality subject to metacognitive monitoring. Evaluation and evaluative learning are closely connected with agency. AI systems that combine agency with indicators implying integration and flexibility may be particularly good candidates for consciousness, because they will share notable cognitive capacities with the animals which are, according to UAL, most likely to be conscious. Furthermore, having an architecture of the kind described by either GWT or PRM is perhaps particularly good evidence for consciousness if it facilitates flexible learning.

However, it is possible that the capacity for unlimited associative learning could be achieved in AI systems in different ways—using architectures that are both unlike those described by theories like GWT, and unlike those belonging to animals that share this capacity. These systems might have this capacity while lacking some of the hallmarks of consciousness identified by Ginsburg and Jablonka. This would undermine the argument for consciousness in such systems, which is a reason to doubt whether UAL itself is a good indicator for consciousness in artificial systems. Another reason is that the status of the UAL hypothesis is very different from that of the other theories we have considered: rather than claiming to identify the mechanism underlying conscious experience, it claims to identify a behavioural marker for consciousness in living organisms. For these reasons, we do not include the capacity for UAL in our list of indicators.

#### 2.4.5 Agency and Embodiment

Current AI systems often relate to their environments in very different ways from humans and other animals, and it can be argued that these differences are evidence against consciousness in such systems. For example, consider the well-known image classifier AlexNet (Krizhevsky et al. 2012), a relatively small and simple DNN trained by supervised learning. The beings which we usually take to be conscious are very different from AlexNet: they are agents which pursue goals and make choices; they are alive and have bodies; and they continually interact with the environment in a way that involves storing and integrating information over short periods of time. AlexNet, in contrast, has the function of classifying images but does not take actions or pursue any goal. It is physically realised in the sense that the weights and other features which define it are stored in physical memory devices, but it does not have a body. And it processes inputs that are separated in time and independent of each other and its outputs in feedforward passes that do not change its state. In this and the following section, we discuss whether further indicators of consciousness can be identified among these “big-picture” differences between humans and some AI systems.

As we will see in this section, scientists and philosophers have argued that various properties related to agency and embodiment are necessary for consciousness. We survey some of these arguments, identifying possible indicators, then discuss whether they can be formulated in ways that are consistent with computational functionalism. We close the subsection by adding two more indicators to our list.

## 2.4.5(a) Agency

One argument for the claim that agency is necessary for consciousness is that this is implied by many scientific theories. Most of the theories of consciousness we have discussed make some reference to agency.

PRM is particularly clear that agency is a necessary condition for consciousness. PRM claims that the subsystem that discriminates between sensory signals and noise must output to a “general belief-formation and rational decision-making system” (Lau & Michel 2021). One of the foundational ideas on which the theory is built is that conscious perceptual experiences have an “assertoric force” which may be explained either as a certain kind of persistence of the signal as an input to this decision-making system or as a strong disposition to form corresponding beliefs. What distinguishes beliefs from other representations, on most philosophical accounts, is in part their use in decisions about how to act (Stich 1978, Dretske 1988, Schwitzgebel 2021). We have already formulated an indicator that expresses this aspect of PRM, but the point remains that agency is a prerequisite for the functions cited in this theory.

The midbrain theory explicitly requires agency, since it claims that the function of the midbrain is to integrate information for action selection, and GWT also emphasises agency, without being quite so clear that it is necessary. Dehaene and Naccache (2001) claim that representation in the global workspace is needed for information to be available for intentional action, in keeping with their presentation of GWT as a theory of access consciousness, and, therefore, of the availability of information for rational agency. The UAL hypothesis claims that “agency and embodiment” and a “flexible value system” (Birch et al. 2020) are hallmarks of consciousness.

There are also independent arguments in philosophy that agency is necessary for consciousness (Evans 1982, Hurley 1998, Clark & Kiverstein 2008). Hurley (1998) claims that consciousness requires intentional agency, which she spells out as agency in which:

…[the system’s] actions depend holistically on relationships between what it perceives and intends, or between what it believes and desires. Relations between stimuli and responses are not invariant but reflect the rational relations between what it perceives and intends and various possibilities of mistake or misrepresentation. (p. 137)

Hurley’s argument for this claim is related to the sensorimotor theory of consciousness and theories of embodied and enactive cognition, which we explore further below. A key idea is that consciousness requires a perspective or point of view on the environment, and embodied agency can give rise to such a perspective. Hurley also argues that part of what it is to be conscious is to have access to the contents of one’s conscious experiences and that this requires the ability to act intentionally in the light of these contents. However, other philosophers argue that it is at least conceptually possible that there could be conscious experience in entities that are entirely incapable of action (Strawson 1994, Bayne et al. 2020).

There are three possible indicators of consciousness suggested by these considerations (in addition to indicator HOT-3, arising from PRM). These are: being an agent of any kind; having flexible goals or values, as suggested by advocates of UAL; and being an intentional agent, as suggested by Hurley. The latter two are strictly stronger conditions than the first, but that does not in itself imply that any should be excluded. It could be that agents in general are significantly stronger candidates for consciousness than non-agents, and intentional agents (say) are also significantly stronger candidates than other agents, in which case agency and intentional agency would both be useful indicators.

In their classic AI textbook, Russell and Norvig write that “an agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through activators” (2010, p. 34). This is a very liberal definition of agency. AlexNet meets these conditions—in fact, all AI systems meet them—and so do many simple artifacts, such as thermostats. So it is too liberal a notion of agency for our purposes.

A more substantive notion of agency can be defined by adding three conditions to Russell and Norvig’s account. First, it is a plausible condition for agency that the system’s outputs affect its subsequent inputs. Without this, a system can only respond to inputs individually, as opposed to interacting with an environment. AlexNet does not meet this condition because, in general, the labels it produces as output do not affect which images it is subsequently given as input. This relates to the second condition, which is that agents pursue goals. We typically think of this as involving ongoing interaction with an environment, in which outputs are selected because they will bring the system closer to the goal—that is, they will change the environment state, affecting their own input, so that the goal can be more readily achieved by future outputs.