What is Self Awareness?

This article offers one look at the notion of self awareness as it bears on the project of building cognitively sophisticated robots. A companion article addresses some simple objections.


The single most significant point for dispelling the confusions which sometimes infest talk of self awareness is that it need have nothing whatsoever to do with consciousness. No single definition of self awareness commands a dominant mind share in the specialist literature, and it is true that it is sometimes conflated with consciousness. But within the context of analysing cognition and of building robots, we are safe in treating self awareness as a purely cognitive characteristic, while ‘consciousness’ refers to the phenomenal properties of mental processes. (For details, see the short notes on consciousness or on qualia in the section containing tutorials & introductions.)

The relationships between (cognitive) self awareness and (phenomenal) consciousness make interesting discussion topics, and they live at the heart of a great deal of contemporary philosophy of mind. It may be, for instance, that achieving the cognitive trick of self awareness guarantees phenomenal consciousness, or it may be that consciousness could even exist without any cognition at all. Fascinating though these questions are, it’s beyond the scope of this short paper to flesh out positions on such relationships (see Mind Out of Matter for some of that!). The important point for now is simply to recognise that the two are altogether different things.

Picking Definitions of Self Awareness

Precise definitions of self awareness vary wildly, but usually they are picked to capture one or several aspects of the term’s intuitive meaning. The intuitive understanding of self awareness underlying most of the literature on the topic is bound up with the notion of an organism using some concept of its own self — as an item situated within an environment — to affect its behaviour. For instance, experiments show that most primate species are unable to recognise themselves in mirrors. (Specifically, if a spot of paint is applied to their forehead and the animals are then offered a mirror, they fail to respond to the image in the mirror by touching their forehead to investigate the spot; chimpanzees and humans are notable exceptions.) This, it has been suggested, indicates that such animals lack self awareness — even though they may perfectly well be phenomenally conscious. (I.e., they feel pain, see red, etc.) On this view, because the animals apparently lack the ability to translate a picture in a mirror into information about themselves, they are not ‘self aware’.

In my own work, I actually tend to avoid this slant on self awareness. This psychologically inspired notion of self awareness seems to me to place too much weight on an organism’s capacity to conceptualise its self in a highly abstract fashion.

Instead, my principal interest centres on the question of whether an organism (or robotic system) can be said to have a self at all, whether or not it actively engages in abstract conceptualisation of that self. In Mind Out of Matter, I develop an information theoretic description of the self model, understood as a type of data structure which can reasonably be called the self of that which instantiates it. Self models capture some notion of ‘self awareness’ in the sense that organisms possessing self models display cognitive processes which incline us to credit them with something like the intuitive notion of self awareness. As such, they offer one way of making precise that intuitive understanding of self awareness. But the notion is clearly much broader than that to which psychologists appeal in denying self awareness to most primates. I would like to make two things explicit:

  • I take self models to be one cognitive route to self awareness, but not necessarily the only route; other architectures might also embody cognitive processes which justify our calling them self aware.
  • The notion of a self model is a definition, but its role in real organisms is an empirical hypothesis.

In other words, I am just defining a class of architectures which yield some important features of the intuitive notion of self awareness, and I am hypothesising that such an architecture subserves the type of cognitive processes which, in real organisms, would lead us to characterise those organisms as self aware.

Self Models: What They Are

The Denser Approach

Each ‘weasel word’ in the following paragraph will be spelled out below:

The self model is understood as a conditionally coupled, functionally active representation of the environmentally situated sensory and motor system of which it is itself a part. (Perhaps the label ‘self-in-a-world-as-it-looks-from-here model’ is more descriptive, since the dynamic data structure essentially reflects a centred perspective on the world and the system within it.) The self model functionally represents itself, the environment, and the interactions between and within each.

The fundamental notion of information underlying notions like ‘data structure’ is that of algorithmic information content. I take representation to be a symmetric relationship obtaining between two physical bodies when those bodies contain substantial mutual information content; alternatively, two items represent each other when they are not algorithmically independent. (For those with an interest in the literature, I would draw this to your attention: Hey, an actual formal definition of representation!) Extending the definition, an item x functionally represents another y when it contains substantial mutual information content both with y and with a set of transformation axioms describing the range of changes in y for a given domain of conditions. Calling a representation or functional representation functionally active means that the relevant body of information plays a role as a representation in a functional system. (‘Functional system’ can be another weasel word, but it is beyond the present scope. In Chapter 5 of Mind Out of Matter, I define a novel measure of process complexity called functional logical depth and use it to provide what is arguably the first precise and mathematically nontrivial account of functional systems.) Finally, a functionally active representation which is conditionally coupled to that which is represented may temporarily become information theoretically ‘disconnected’ from that which it is representing, in the sense that the actual state of that represented may deviate from the information represented. The degree of such uncoupling is quantified by the conditional information relation between the representation and the represented.

The Easier Approach

Here’s the easy way to make sense of the above information theoretic look at self models.

The self model is a body of information about an organism and its environment and about the flow of changes between and within the two. That body of information plays a central role in directing the behaviour of the organism. Unlike the case of a modern digital computer, in which a very sophisticated central processing unit operates on what are usually comparatively simpler data structures, the self model is a sophisticated data structure which affects the operation of comparatively simpler (at the relevant level of description) biological components. The conditional coupling of the self model to the organism and the environment allows the organism to derive information about counterfactual conditions in itself or its environment. In other words, because it contains information about the interactions of changes between and within the organism and the environment, the self model can yield ‘answers’ to endogenously generated questions about what would happen if current conditions were other than they are. It is this capacity to test counterfactual hypotheses which underlies what Popper famously described as a creature’s ability to allow its hypotheses to die in its stead.

Self models are notably absent in all current artificial systems of which I am aware.

The capacities which self models subserve, however, are notably abundant in the natural world. (Note that this does not necessarily imply that self models themselves are similarly abundant; recall that the existence of self models is itself an empirical hypothesis.)

Self Models: What Good Are They?

So suppose one has a self model. What good is it?

The principal advantages of the self model derive from the Popperian notion described above: an organism may rehearse alternative courses of action without actually engaging in them, it may model the viewpoint of its conspecifics by modelling itself in the physical position of another individual, and so on. Within a selectionist context, the advantages of having a self model over not having one are, ceteris paribus, straightforward. How about robotics, and machine cognition more generally?

Creating machines with a sense of self is unexplored territory. The machines we use today are under a sort of selective pressure, albeit of a fairly bland sort. Sometimes, high price/performance ratios win out, although installed base, software availability, and many other factors combine to press one or another variety of computer to the fore and send others to extinction. Will the cognitive sophistication so important for organisms competing in the biological realm prove useful in the artificial realm as well? Just how would a machine with a sense of self stack up? How might it exploit its ‘self awareness’ to get itself selected over the latest shiny PowerPC or plain vanilla Pentium II?

One thing is almost certain: just as the appearance of biological organisms with progressively richer notions of self radically transformed the selective biological environments within which they emerged, so too will the emergence of machines with a sense of self radically transform the selective dynamics of the marketplace. While they will initially occupy the commercial equivalent of a novel and tiny environmental niche, I believe that eventually they will largely displace today’s ‘dumb’ boxes, rendering them a quaint bit of nostalgia.

By analogy to the biological world, today’s computers have yet to climb out of the primordial silicon soup. It took Nature 3.8 billion years to create us, but I believe it will soon become clear that we can move just a little faster than that when it comes to pushing machine cognition up the ‘evolutionary’ scale.


Since it can be easy to misinterpret some of the statements above summarising the more careful and extended treatment of self models offered in Mind Out of Matter, a separate short paper covers some of the simpler objections.

This article was originally published by on and was last reviewed or updated by Dr Greg Mulhauser on .

Mulhauser Consulting, Ltd.: Registered in England, no. 4455464. Mulhauser Consulting does not provide investment advice. No warranty or representation, either expressed or implied, is given with respect to the accuracy, completeness, or suitability for purpose of any view or statement expressed on this site.

Copyright © 1999-2021. All Rights Reserved.