Would proponents of the integrated information theory of consciousness consider neural networks in machine learning to be conscious?

  • 2
    According to the original IIT, even matrix multipliers and Boolean circuits with expander graphs are "conscious", see Aaronson's blog, so it does not mean much. Tononi tried to present that as a feature, so to him circuits are conscious, but it did not go over well with others. No palatable measure of integration has been found since, so at the moment IIT is only a programmatic framework unable to answer such questions. The search is ongoing, see Tegmark, Improved Measures of Integrated Information.
    – Conifold
    Aug 19, 2022 at 13:19

1 Answer 1


In integrated information theory consciousness is not only about amount of integration, but also how it is structure and dynamic activity. One of it's major goals is explaining the difference between awake, asleep and altered states of mind, where it is obvious that these are not simply properties possessed by a brain, but about dynamics.

So, a neutral network: IIT generates a spectrum for consciousness, instead of a binary of got it or not. In this picture, even a thermostat has a minimal amount of integrated information, as it works. Similarly a neural network in operation can.

We think a whole human brain can be simulated using neural networks, or at least The Human Brain Project do. If that is right, you question is a bit like asking 'Can a computer be intelligent?' - they can be, contextually if programmed right and depending how you define your terms.

Conscious is a big word. Generally what we mean is not the literal interpretation, 'aware of thing/s', but self-conscious, and not unconscious. We think there is a kind of feedback that occurs with having a model of ourselves in our minds, as manifested by the Mirror Test applied to different animal species, and as understood in Hofstadter's Strange Loops. In this sense, we haven't generated this kind of structure in neural networks, and we can't describe them as self-conscious. We can see the drive towards 'intelligible intelligence', deep-learning algorithms that can explain what they learned and how (see eg The challenge of crafting intelligible intelligence), can be understood as moving in this direction.

  • 1
    Interesting direction. Regarding: "IIT generates a spectrum for consciousness, instead of a binary of got it or not", it seems like it answers the question by diluting what we normally mean by 'conscious'. Like you could say a virus particle is 'alive' even though it meets none of the criteria, because it interacts with living cells. Are prions alive? Is a thermostat conscious? Similar question. I hope we never discover the AI equivalent of prions.
    – Scott Rowe
    Jan 16 at 19:26
  • 1
    @ScottRowe: I don't really see it as diluting, any more than recognising viruses as somewhere between life & non-life dilutes the definition of life. An interesting point is that we often judge capacities by the most exceptional individuals, whether Einstein or Alex the Parrot or Washoe the chimp. But don't we have to accept & account for variation among among individuals? & that there is a direction towards increasing our cognitive capacities? IIT could offer a handle on what that might mean, I think.
    – CriglCragl
    Jan 16 at 21:06
  • Ok, maybe I went off in the wrong direction. But if as you say self-consciousness is what we usually mean by 'conscious', most humans qualify, and few animals do. How could we tell with a computer? Self-preservation is the most tested feature of consciousness, so if a computer tried to stop me from shutting it off, that would be some evidence. Thermostats probably will never reach that level of functioning. If someone lacks that kind of awareness, generally people would say something is wrong. So, the spectrum is mostly meaningful at one end.
    – Scott Rowe
    Jan 17 at 19:01

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .