A machine is MC4 conscious if it has phenomenal experiences that are comparable to human conscious experiences. From an ethical point of view it is important to know whether we have created MC4 consciousness in a machine. MC4 consciousness research can also contribute to the development of general theories of human consciousness. This paper discusses four problems that have to be solved before we will be able to address MC4 machine consciousness in a systematic way: We need more clarity about the measurement of consciousness, we need better ways of describing the physical world and consciousness, and we need to reach agreement about the final form that a theory of consciousness should take. When these problems have been addressed we will be able to develop scientific theories of consciousness that can make accurate believable predictions about MC4 consciousness in machines.
It is often helpful to distinguish four types of machine consciousness [
These different types of machine consciousness can be combined. For example, a robot that displayed conscious external behavior (MC1) could be controlled by a model of consciousness (MC3) and have phenomenal experiences (MC4).
This paper will address the question of MC4 consciousness: artificial systems that have phenomenal experiences that are similar to our spatially and temporally distributed experiences of color, smell, taste, etc. This type of machine consciousness is ethically significant because MC4 conscious machines could suffer. Research on MC4 machine consciousness could also be a key factor in the discovery of general theories of consciousness that are not limited to the neurons inside the human biological machine.
MC4 machine consciousness can be solved when we have developed scientific theories of consciousness that can make accurate detailed predictions about consciousness in any physical system. This will only become possible after four preconditions have been met. First, we need to reach agreement about how consciousness can be measured and abandon the idea that consciousness can be measured through machines’ external behavior (Section 2). Next we need to move away from neural correlates of consciousness and find new ways of describing spatiotemporal physical patterns in the brain that could form the basis for generalizable theories of consciousness (Section 3). Third, we need less anthropomorphic ways of describing consciousness (Section 4). Finally we have to drop our desire for intuitively satisfying explanations of consciousness and search for mathematical relationships between formal descriptions of consciousness and formal descriptions of the physical world (Section 5). When these preconditions have been met we will be able to use human and animal experiments to discover mathematical theories of consciousness that can be generalized to make believable predictions about MC4 consciousness in artificial systems. 2
We cannot directly measure other people’s consciousness: we have to measure consciousness through external behavior (first-person reports). An inference from external behavior to consciousness only works when the external behavior is generated by a system that we believe is capable of consciousness – for example, when conscious external behavior is exhibited by normal humans and primates. The inference from external behavior to consciousness is less certain with brain-damaged patients, infants and cephalopods.
Some people believe that conscious external behavior can be used to infer the
presence of consciousness in artificial systems. If this was the case, the problem of MC4
machine consciousness would be solved. Any machine that exhibited conscious
external behavior would be judged to be phenomenally conscious and we could easily
identify the correlates of consciousness in machines. The problem with this position is
that it is extremely easy to write a computer program that mimics human first-person
reports about consciousness. For example: “cout<<’I am conscious of a red apple.’;”
This behavior can be produced by a giant lookup table [
These problems prevent us from using external behavior to measure consciousness in artificial systems, regardless of how humanlike that behavior is. If we cannot measure consciousness in artificial systems, we cannot do MC4 consciousness experiments on artificial systems. Instead, we will have to discover the relationship between consciousness and the physical world in humans (and similar systems) and then generalize these results to animals and artificial systems. To make this generalization possible we have to rethink the way in which we study consciousness in humans. We need to describe the physical states of human brains in a way that will enable us to apply our theories of consciousness to non-biological systems. We also have to describe consciousness without the anthropocentrism of natural language. These challenges are covered in the next two sections. 3
The measurement problem forces us to develop our theories of consciousness through
experiments on humans and similar animals. Although there have been promising
results on the neural correlates of consciousness [
It might be thought that the generalization problem could be solved by linking
consciousness to functions, computations or information patterns in the brain. If
consciousness is linked to computations in the human brain, then a robot that executed
the same computations would also be conscious. A key problem with functional,
computational and informational theories of consciousness is that functions,
computations and information are not objective properties of the physical world. Information
appears when you apply a human-defined interface to the physical world [
An objective scientific approach to consciousness has to look for connections between spatiotemporal physical patterns and consciousness. Much of the recent research on consciousness has focused on the relationship between neural activity patterns and conscious states. However, this will not lead to a theory that can be generalized to artificial systems. Neurons are defined in a specific biological context – the brains of animals – we have no formal definition that would enable us to unambiguously identify neurons outside of this context. Suppose we genetically engineer a sequence of 100 hybrids between neurons and liver cells: the first cell is 100 % neuron; the middle cell is 50% neuron; the last cell is 100% liver. We have no systematic way of classifying intermediate cells in this sequence. Or suppose we synthesize an approximate neuron from basic biological components – we have no idea whether synthetic or silicon neurons could form correlates of consciousness. A generalizable theory of consciousness has to be based on precisely defined spatiotemporal structures that can be unambiguously identified in any physical system. We will have to define neurons more precisely if we want to generalize the results from the neural correlates of consciousness, or we could base our theories of consciousness on other properties of the physical world, such as electromagnetic waves. 4
The majority of work on consciousness has been based on contrastive analyses of conscious and unconscious brains. This gives us valuable information about the neural activity patterns that are necessary for any conscious state to occur – neural activity patterns that are common to all conscious states. There has been less work on the detailed relationships between the contents of consciousness and physical states.
Conscious states are typically described in natural language. But generalizable
theories of consciousness cannot be based on natural language because it is vague, highly
compressed and context dependent. It is likely that artificial systems will have
radically different experiences that will be impossible to describe in human language.
Natural language descriptions of consciousness are also likely to be highly misleading
because artificial systems might have different representations of the temporal and
spatial properties of objects [
In some experiments conscious contents are specified using the stimuli that
produced the conscious experiences. For example, in the work on brain reading with
fMRI the decoded conscious experiences are presented as videos, which the subject
compares with their own conscious experiences [
To address this problem we need to find new less anthropocentric ways of
describing consciousness that will enable us to generalize the experimental results on human
consciousness to artificial systems.1
1 One solution to this problem has been put forward by Balduzzi and Tononi, who suggested
how states of consciousness could be described using high dimensional mathematical
structures [
There is a lack of general agreement about the final form that theories of consciousness should take. A generalizable theory of consciousness should meet the following criteria: 1. Can generate testable predictions. 2. Is applicable to any physical system. 3. Compact (Occam’s razor). 4. Based on objective properties of the physical world.
Functional, computational and informational theories of consciousness meet criteria 1-3. However, as explained in Section 3, they fail to meet Criteria 4 because they are not based on objective properties of the physical world. Theories about the neural correlates of consciousness meet Criteria 4, but they are often not compact or generalizable, and they have a weak ability to generate testable predictions.
Many people are looking for an intuitively satisfying explanation of the
relationship between consciousness and the physical world. They want to make an
imaginative transition between a physical state (for example a mental image of a physical
brain) and a conscious state (for example, the color red). We are never going to get
this type of theory because we can only imagine conscious experiences, not the
invisible physical world as it is in itself without any conscious properties. In principle we
might be able to discover a theory that will enable us to make an imaginative
transition from a conscious experience of a brain to another conscious experience. But this
would only become intuitively convincing once we had learnt which brain patterns
are linked to conscious experiences [
The most plausible type of theory that meets all of the criteria is a mathematical
relationship between a formal description of the physical world (see Section 3) and a
formal description of consciousness (see Section 4). Compact mathematical theories
are the gold standard in many sciences, they can generate testable predictions and
they can be applied to any physical system. The most impressive mathematical theory
of consciousness that has been developed so far is Tononi’s information integration
theory [
To develop fine grained mathematical theories of consciousness we need high resolution data from the brain. However, the vast amount of data that could potentially be recorded from a brain could not be comprehended by a single human brain. This suggests that humans are not likely to be capable of discovering mathematical theories of consciousness. We could address this problem by using artificial intelligence to search for mathematical relationships between consciousness and the physical world.
This paper has discussed four problems that have to be addressed before we can develop a scientific solution to MC4 machine consciousness. We need to reach agreement about how consciousness can be measured and abandon the idea that consciousness can be inferred from a machine’s external behavior. Scientific work on consciousness needs to move away from neural correlates and start to look for more accurately defined spatiotemporal structures in the human brain that will lead to generalizable theories of consciousness. We need to develop ways of describing consciousness that avoid the anthropocentrism and context dependence of natural language. We have to drop our desire for intuitively satisfying explanations of consciousness and search for mathematical relationships between formal descriptions of consciousness and formal descriptions of the physical world. If these challenges can be addressed, we will have made significant progress towards tackling MC4 machine consciousness in a scientifically plausible way. 7