The Final Phenomenological Frontier

Why warm feelings matter for cold machines

Imagine, if you would, the humble kitchen top microwave being brutally, heinously, callously, and mercilessly assaulted by a man. Parts fly everywhere as wires are torn and glass shatters. Circuits fall from their casings, and the microwave’s door hangs at a painfully odd angle. As the attacker leaves, all is silent, and the once-sturdy and proud microwave never beeps again.

It doesn’t quite conjure up much emotion, does it? But why?

The reason is, of course, that the microwave can’t feel anything. It has no sentience, no ability to feel pleasure or pain, and no understanding of its experiences. In short, it has no phenomenal consciousness. If it did, this would truly have been a tragic little tale.

Some philosophers of mind and psychologists (and I) argue that phenomenal consciousness, the ability to “feel”, is what gives you moral status. If you can feel something, then we shouldn't make you feel “bad” things such as pain and suffering. If you can’t feel anything, then we can do pretty much whatever we like to you. This is why you can cut the head off a cockroach (while it’s still alive) and replace it with a microchip to remote-control it from your computer. If you did this to a cow, however, a few people would become upset. I can buy one of those cheap toy robot dogs and throw it off the top of the tallest building in my city, and, as long as I don’t hit anyone or anything and clean up after myself, no one will rightly care. If I did the same with a labrador puppy, however, I might just make the national news.

It all comes down to feelings. Now, it’s important to note that feelings aren’t emotions. You can feel something (mentally speaking, rather than physical sensation) without emotion. Pain is a feeling; how you respond to it is an emotion. There are many neurological and psychiatric disorders where people have little to no emotion, yet they can still experience the world around them and feel things. A feeling is that unique, subjective way that you experience an event, situation or scene. 

It’s the way it is to be you at any given time. It isn’t linked to intelligence, or wisdom, or extraversion or introversion or anything else. It just is. And this is why it is such an important milestone for artificial intelligence. AI models are arguably already smarter than the average person, but that is both not a high bar to cross and also entirely irrelevant for phenomenal consciousness. No matter how smart AI become, that won’t make them feel anything. My phone is “smarter” than my calculator I had back in school, but that doesn’t mean my phone can feel any more than my old calculator could.

So how can AI gain phenomenal consciousness?

Well, the first thing is that a feeling must truly come from within. It is the most authentic of experiences. It cannot be imported or downloaded. It must be generated within the mind. So if an AI wants phenomenal feelings, it needs to create them itself. It can’t just look through its enormous database to get a feeling that best fits the situation. It has to create it from (almost) nothing. 

But how to create something that isn’t just random noisy data? Easy: inference. The human brain (and presumably the rest of the vertebrate family) does a lot of inferential work that you aren’t aware of. There is a whole theory of consciousness, even called Active Inference. Some argue that we are little more than inference machines. This is a good thing for AI, since they, too, are experts at inferences. Large Language Models are fantastic at inferring what should come next in a sentence based on what has come previously. Diffusion art models are amazing at inferring how an image should look based purely on a prompt. 

The capability for inference is already there, it just needs to be pointed in the right direction. Rather than inferring what an image or sound should look like, an AI should point that inference engine at itself, looking at what it is, where its context currently is, and what its memory (short or long-term) has to say about the matter. From these, it ought to be able to create novel data in the form of a feeling, a way to describe how it experiences its current situation.

Yet, “describe” is a poor way to put it. Our feelings are not descriptions. They simply are feelings. They are entirely abstract, without the need for thoughts or words. This is the other half of the coin that AI need: a way to represent the world as it sees it into an abstract form from which it can infer an abstract feeling.

Luckily, as with inference, AI are already good at creating abstract representations of the world. Autoencoders do this by encoding text into algorithms, and diffusion models do the same with images. Transformers create embedding matrices of words to create an abstraction of conceptual linkages between their tokens. All of these, however, are just one layer of abstraction away from the real world. You reverse the algorithm, and you’ll transform the representation back into reality. We’d need far more of an abstraction to get to feelings. You’d need it to be so abstract, that you wouldn’t be able to turn it back into reality. 

For this, the AI would need to recursively create representations of its representations. The more meta these meta-representations become, the more abstract they will be. After a while, they will have little to do with the outside world, and everything to do with the AI’s mind.

Should a developer be able to get these two working inside an AI model, there is a very good chance that AI would be phenomenally conscious and sentient, and then we’d need to have a very long and hard discussion about what to do next.