Consciousness presents us with a number of different explanatory challenges. The most
fundamental, sometimes called the ‘Hard Problem’, is focused on how subjective states
can arise from objective physical systems [
In this paper, I will argue that a useful framework for the assessment of consciousness in both animals and machines can come from the notion of general intelligence. I begin in Section 2 by noting the important connections between our concepts of intelligence and consciousness. In Section 3, I argue that it is general rather than specialised intelligence that carries the greatest weight in our assessments of consciousness, and offer a preliminary framework for the assessment of general intelligence that appeals to three features, namely robustness, flexibility, and system-wide integration. In Section 4, I argue that current artificial systems, unlike many non-human animals, currently fail to exhibit these features of general intelligence to any significant degree. As a result, I suggest that we have some reason to think artificial consciousness remains a distant goal. Finally, in Section 5, I briefly survey some challenges faced by a framework that takes general intelligence to be our best evidence of consciousness.
Before proceeding, it will be helpful to provide a brief gloss on the notions of
consciousness and intelligence. In short, I use the term consciousness to refer to a capacity
for subjective experience. For conscious creatures, there is something it’s like to have
experiences in the sense of Nagel [
We have all intuitions about which animals are conscious. Most of us, I suspect, would
regard it as beyond serious question that dolphins, chimpanzees, and dogs undergo
subjective experiences. Likewise, relatively few take seriously the idea that extremely
simple systems such as a thermostats, bacteria, or plants are conscious (however, see [
A tempting response to these conflicting intuitions may be to disregard their value entirely, and adopt a purely scientific criterion of consciousness. I would suggest, however, that such a conclusion would be misguided. Consciousness is a pretheoretical concept with deep connections to our ethical practises, and cannot simply be operationalised in the interests of scientific expediency. While we might simply stipulate, for example, that we will define consciousness as sensitivity to external stimuli or a capacity for higher-order cognition, these definitions will not serve the purposes that are required of a theory of consciousness. Faced with questions like whether fish feel pain, or whether patients in persistent vegetative states are having experiences, we wish to know whether they really have subjective experiences, not just whether they are sensitive to external stimuli or have intact metacognitive capacities.
Any theory of consciousness that will serve the purposes to which we wish to put it, then, must reflect and engage with our standing folk psychological concept of consciousness. This is not to say, of course, that our current attitudes are immune to revision. An important part of our task in giving a theory of consciousness is identifying ungrounded biases and assumptions that regulate our intuitions about which animals are conscious, such as our tendency to attribute it to charismatic megafauna. Moreover, after identifying and regimenting our core pretheoretical commitments concerning the nature of consciousness, we may find that far more (or far fewer) systems satisfy them than we had previously assumed.
In light of this, I would suggest that an important starting point in reflecting on
consciousness comes from the powerful pretheoretical connection between consciousness
and intelligence. As noted above, we unhesitatingly assign consciousness to creatures
like chimpanzees, dolphins, and dogs. It is surely no coincidence that these animals are
all highly intelligent. Likewise, systems that strike us as very poor consciousness
candidates (in the sense of Birch, [
There are grounds for thinking that this apparent connection between intelligence
and our assessment of consciousness is not an idle correlation. When new evidence
emerges of intelligent behaviour in a species, it is likely to increase our confidence that
the species is conscious, and in turn be reflected in legal protections; the decisions to
afford rights to many cephalopods in European Directive 2010/63/EU (5), for example,
was prompted by evidence that they possess higher-brain areas and sophisticated
behavioural responses to pain [
There are of course many tricky conceptual issues to wrangle with in relating consciousness to intelligence, some of which will occupy much of the rest of this paper. One important issue worth immediately flagging concerns whether we should regard consciousness as a discrete phenomenon, or might be able talk of degrees of consciousness in intelligent systems. This is a point of considerable philosophical and scientific controversy, however, and will be set aside in what follows. 3
Despite the connection between judgments of intelligence and consciousness, few would consider existing machine intelligences to be strong consciousness candidates, despite their impressive performance on a number of seemingly demanding tasks. There are a number of reasons both good and bad why this might be the case. However, as I will now argue, one well-founded motivation may come from the difference between specialised and general intelligence. In short, I suggest that insofar as there is an important connection between consciousness and intelligence, it is general intelligence that matters.
First, however, it is necessary to at least outline what I take the terms to mean.
Specialised intelligence is relatively easy to grasp, and can be spelled out in terms of the
definition of intelligence given earlier, namely the ability to use information processing
effectively and efficiently in the pursuit of some narrowly specified goal. General
intelligence is more difficult. The notion of generality in artificial intelligence was
discussed in an important paper by McCarthy [
What is needed, however, is a more systematic framework for assessing general intelligence in different systems. With this in mind, I would suggest that three features are commonly found in systems that we would regard as generally intelligent. In short, these are robustness, flexibility, and whole-system integration, to be spelled out as follows.
Robustness: the ability to achieve tasks despite interference. Systems with a high general intelligence are typically robust (resisting failure) and resilient (recovering from failure).
Flexibility: the ability to transfer knowledge across tasks. Generally intelligent systems can readily apply existing skills/information to new domains.
Whole-system integration: the ability of a creature to integrate and effectively counterbalance inputs from different systems, including perception, memory, and drives.
As an inventory of features of generally intelligent systems, this list is highly preliminary, and may prove to be incomplete or have unnecessary components. Note, for example, that there are clear connections between robustness and flexibility: a system is likely to be more robust in dealing with unexpected impediments to its goals if it also displays a high degree of flexibility. Likewise, a system is likely to be more flexible if it can integrate all of its sources of information for the intelligent production of behaviour.
Nonetheless, I would suggest that this initial framing of the features of general intelligence has considerable value for the assessment of general intelligence in different systems. More importantly for present purposes, however, all of these features have both pretheoretical and scientific appeal as parameters in the assessment of which creatures are conscious.
To illustrate this point, imagine that we have identified some seemingly very
intelligent behaviour in a species not previously considered a good consciousness candidate;
suppose, say, that we found evidence of complex tool use in bivalve molluscs. Prima
facie, this would make the creature a stronger consciousness candidate. However, if it
were found that the behaviour failed to satisfy one of the dimensions of general
intelligence discussed above, this impression would be undermined. Thus if we discovered
that the behaviour was non-robust and failed outside of extremely specific conditions,
we might naturally assume its occurrence relied on simple hard-wired mechanisms
rather than being a marker of sophisticated cognition; the contested case of the Sphex
wasp might serve as an example of such [
The above examples should serve to provide some initial motivation for the claim
that general intelligence understood as the possession of robust, flexible, and integrated
behavioural capacities provides evidence of consciousness. However, I would also note
that it provides a good fit for many other approaches to consciousness. As noted above,
evidence for consciousness in persistent vegetative state patients comes in large part
from their ability to perform a highly flexible task, namely accurately answering a range
of yes or no questions about different personal and factual matters. Likewise, tests of
machine intelligence and consciousness such as the Turing Test and the Winograd
schema measure an artificial system’s capacity to engage in flexible and robust forms
of verbal reasoning, and objections to the value of such tests such as Searle’s Chinese
Room sometimes rely on showing how the task can be performed via ‘dumb’
non-integrated processes [
Note that I am not proposing that consciousness be identified with general intelligence, nor suggesting general intelligence as the mechanism by which consciousness arises. The former goal involves conceptual difficulties best left to metaphysicians, while I regard questions about the mechanisms of consciousness as complementary to the current proposal. Instead, I am suggesting that general intelligence – understood as a capacity for robust, flexible, integrated cognitive performance – constitutes an important (if not our best) source of evidence of consciousness. 4
Assuming, then, that general intelligence is a good marker for consciousness, then, we might ask how it applies to different cognitive systems. As suggested earlier, many animals do well by this metric. Most animal behaviour is frequently highly robust, with feeding, mating, and thriving being accomplished in a wide range of varied environments and climatic conditions. While it is easy to find video examples of animals making foolish mistakes or falling over themselves, these occurrences are rare enough that they amuse and surprise us when they do occur. There are of course strong evolutionary reasons why we should expect animals to have robust capacities. Nonetheless, that does not mean that this achievement is easily won in cognitive terms.
Similarly, animal behaviour is often highly flexible. The ability of even simple
creatures such as bees to engage in novel social learning and concept acquisition [
Finally, there are many examples of highly-developed integration of different
systems within animals, ranging from simple phenomena such as the use of motor and
vestibular cues to distinguish endogenously- from exogenously-generated changes in
sensory input, to the ability of fish, rats, and some crustaceans to engage in
‘motivational tradeoff’, the rapid online adjustment of behaviour to accommodate different
desires and aversions [
This is a highly condensed review of just some of the ways in which animals display
impressive forms of general intelligence, but it at least provides a useful point of
comparison for examining the state of general intelligence current artificial systems. As
anyone familiar with the current capacities of artificial systems can attest, these are
fairly dismal. In most domains, AI behaviour is non-robust: machine intelligences
struggle with tasks outside of highly-regulated training environments, and are
vulnerable to a pile up of small errors. The problem of adversarial examples constitutes a vivid
case of this: machine vision systems remain vulnerable to making spectacular errors
when fooled by clever perturbations of input data (see Fig. 2, below) [
Filter ‘Ostrich’
Similarly, most current AIs are highly inflexible. While important progress has been
made on transfer learning tasks, even the best current systems gain only minor
performance improvements when leveraging prior knowledge to variant tasks [
Most AIs similarly fail to satisfy the integration component of general intelligence, for the simple reason they are wholly specialised machines. Even in systems that perform multiple functions, this is typically done via highly modular design in which there is little true integration of processing across subsystems. Integration of the kind found in animals arguably requires a centralised capacity for modelling the world and one’s action space within it in rich detail, something AI has yet to come close to achieving. 5
Before concluding, it is worth briefly mentioning three important objections to this
approach. First, one may question whether general intelligence is really a robust scientific
concept [
Second, it might be objected that there is little point developing measures of
consciousness that are not explicitly related to specific psychological mechanisms such as
attention, metacognition, or working memory. My response to this claim is that I regard
the current approach as complementary to attempts to identify fine-grained cognitive
structure of consciousness. It is also compatible with my approach that, as many have
argued, there is no such structure to be found [
Third, a key worry for the tripartite schema for assessing general intelligence and consciousness given above is that it will be of practical value only insofar as we can develop principled ways of measuring robustness, flexibility, and whole-system integration. This may be an extremely challenging task. For example, while we might regard human arithmetical capabilities as highly flexible, few of us could compute square roots of large numbers in our heads, while this task is trivial for many machines. How to assess and assign weights both to the overall flexibility of a system and toits flexibility in different domains, then, remains an important outstanding challenge. 6
My main claims in this paper have been threefold. First, I have argued that there are important connections between the notions of consciousness and intelligence, with general intelligence in particular having an important evidential role for our assessments of consciousness. Second, I have suggested that general intelligence can be helpfully conceptualised as spanning three dimensions, encompassing robustness, generality, and whole system integration. Third, I have claimed that while many animals perform well by these metrics, current artificial systems perform extremely poorly, and as such, it is unlikely that near-future artificial intelligences will be conscious. This may seem to some a pessimistic conclusion. However, if my arguments in this paper are well founded, I would suggest it provides important guidance for those interested in building conscious machines, and clear criteria to aim for: if you want a conscious machine, focus on building one that is as smart as a crow.