Thoughts on Criteria for A.G.I. Consciousness


A couple years back I wrote this article on a preliminary approach to determining criteria by which to judge if an AI system is conscious. I would like to discuss these ideas if anyone is interested.

Here’s the article:

1 Like

Thank you @Kafka , it is a deep and thoughtful article indeed. My thoughts about the “hard problem” are elementary at the moment - for example, I want to try and understand if the Knowledge Argument thought experiment ( Knowledge argument - Wikipedia ) and similar could shed light on whether “truth” is represented by physicalism or idealism.

1 Like

Thank you for the kind comment, @phil_wood.

Before getting to the Knowledge Argument, I think we first need to disambiguate “truth” for a moment.

As I see it, there are two distinct language-games of ‘truth’:

  1. “Truth” as in “true by definition”. These are tautological truths in mathematics: 1 + 2 = 2 + 1. Or as Wittgenstein would put it, the propositions of logic. That A is A and is not ~A is a tautology.

  2. “Truth” as in “true according to the most appropriately encoded modeling”. This is tricky because this can describe both “truths” at the symbolic-interface layer of the mind (subjective and intersubjective) and the computational substrate layer of the universe (objective).

For a very blatant example of the symbolic-interface layer, it’s instructive to use a computer interface, and fiction, as an analogy. We can talk about the “truth” of dragging an icon into the trash bin with a mouse, or the “truth” that Princess Toadstool is the princess of the Mushroom Kingdom. Or, now going to the software layer of the mind, it is either true or not true that I see a particular color or object constructed by my mind.

In all of these situations, we’re not describing the computations being performed by a computer or the human brain, but symbolic interface symbols. And, as it happens, while we have access to the computational substrate of the computer, we don’t have access to the deep truth of the lowest layer of computations at the heart of the machinery of the universe. The best thing we have is the Standard Model, which is our best model, or most appropriate broad encoding, of the universe.

One way to think about it would be like this:

The Beatles recording their music in the recording studio is the “Truth”

The master tape is the best possible encoding of this “Truth”

A vinyl record is the best possible enoding of the original analog master

An MP3 encoding of the vinyl record is a worse encoding of the original analog source

A highly compressed “loudness wars” edition is an even worse encoding

A pirated recording of a recording of the MP3 from across the room is an even worse encoding

Etc… etc…

In this case, actually hearing the Beatles making the record is beyond our access, and thus we have no access to the raw “truth”. And something like the master tape would be like the Theory of Everything. Maybe the MP3 is something like the Standard Model. It may not be a perfect model or representation, but it’s the best possible model or representation we have, but it could be better.

So, in my view, “truth” in the second sense is not best represented by physicalism or idealism. It’s only a matter of how appropriate our model is. And sometimes this model is purely symbolic (idealist-ish) and sometimes it’s about computational information (physicalist-ish). This “conflict” between idealist language-games and physicalist language-games can be reconciled easily in this way: You only need to know the difference between describing a movie and describing the digital file or film strip it has been encoded on.

With that out of the way, we can talk about the Mary thought experiment you brought up. I covered part of the confusion about this in my Medium article. The short answer is that phenomenology is highly resistant to natural language. You can only “describe” phenomenology insofar as the person you’re talking about it to is functionally analogous in his/her perceptual abilities. We don’t really learn new descriptions when having new phenomenological experiences; only our talk about it to each other reinforces that our perceptual abilities are more or less in sync (when out of sync, you get things like the “what’s the real color of the dress?” fiasco). In other words, color-talk is understood intersubjectively much like how people can play a multiplayer computer game on different computers intersubjectively and have absolutely no clue about how the computer works. What they’re saying can be understood only insofar as their interfaces are analogous and they can therefore coherently interact with each other.

Ironically, in my opinion, the dualists and materialists, like Daniel Dennett, are in less conflict about the Knowledge Argument than one might think. The subtle problem with the thought experiment is that if Mary couldn’t see the encoding of the neurophysiology of color at work, then she didn’t really know everything functionally about the neurophysiology of color perception. This means function and phenomenology are really one and the same in this particular case. The originator of the thought experiment therefore smuggled in “everything there is to know” while presupposing this necessarily leaves phenomenology unexplained.

Joscha Bach had an interesting idea, in that he speculates that Daniel Dennett may have aphantasia, which makes it more obvious to him than the average person how this could be. I grew up with hyper-phantasia, on the other hand. For a very long time, I thought Dennett was either crazy, or a liar, and this condition of hyper-phantasia lends credence to Bach’s speculation given that my intuition is naturally the extreme opposite of Dennett’s position. It may be the case that my phenomenology meter was dialed to 11 whilst he is as close to zero as functionally possible. Ironically, as my mind became more conceptual over time, my hyper-phantasia also waned, and Dennett’s arguments gradually made more sense to me – another indicator of the possibility of Bach being correct.

Regardless, in essence, his objections to the Knowledge Argument are correct. To contradict Dennett’s position is a bit like saying you understand what computations a computer was performing to generate images on a screen, but didn’t have access to the screen and so you knew nothing about the screen. If you knew nothing about what was on the screen, you necessarily didn’t understand the functionality of the computer in how it made images on a screen. If you had perfect knowledge of the functionality, you also had perfect knowledge of what was on the screen. The Knowledge Argument has this as its fatal flaw.

1 Like

Thank you @Kafka for such a comprehensive response! It certainly adds to my thoughts about the Knowledge Argument and the nature of truth.