Transcript: Are Machines Conscious?

Do machines have consciousness? Today, we are seeing outputs from generative AI that we would undoubtedly call creative, if they came from humans. Large Language Models are scoring high on Bar Exams and other cognitive measures. Sane people are calling AIs their best friends, or even their lovers. 

Are we living among conscious machines not in some science-fictional future, but today? I believe the answer to this question is YES, but there’s nuance. 

The first step is to define what we mean by machine consciousness.

Let’s dig in.

[INTRO]

Let’s stipulate some guidelines for our definition.

First, let’s remember: a definition is just a construct. Botanists say an avocado is a fruit, but most cooks think of it as a vegetable. Is one right and one wrong? No, they have different definitions for different purposes. Botanists think of function: an avocado carries a seed, therefore it is a fruit. Cooks think of flavors, an avocado is savory as opposed to sweet, therefore it is a vegetable. They are both right.

So if we are going to come up with a definition for machine consciousness, we are going to have to judge it on how well it works as a cognitive tool for a purpose. That means that we might have to come up with more than one definition depending on what we use it as an input for. And that’s okay. Like I said, a definition is just a tool, not an identity in itself.

So our first principle: A definition for machine consciousness should be purpose-built.

Here’s one famous definition of consciousness, from philosopher Thomas Nagel. “An organism is conscious if there is something it is like to be that organism.” 

Well, that’s somewhat vague, but there’s a bigger problem. Thomas Nagel is famous for making the point that we can never know what it is like to be a bat, for instance. How does a bat experience echo-location? We will never know. And likewise, you cannot know exactly what it is like to experience my consciousness, and I don’t know what it is like to experience yours. So – given that our consciousness is not permeable, it might be too confident for me to say that yours is like something. 

You might say: that’s too strict. You and I can share notes on our consciousness and realize that the descriptions of our experience are comparable. Also, since we are the same species, we share a lot of genetic material, and that should imply some commonality. There probably is something it is like to be you. Fair enough, but our topic is machine consciousness, where there’s no shared biology. Also, a major issue is going to be whether the machine is faking it or not – because we know it is very good at faking.

So we come to the second principle: A definition for machine consciousness should be testable.

Alan Turing, the father of modern computing, had a test for a thinking machine: Can the machine use written communication to fool an observer into thinking it’s human?

That is a test, not a definition, but it is instructive all the same. Why? Because Large Language Models, if properly prepped, could easily pass the Turing Test today. Did we release the streamers and balloons when that happened? No. Because there is something in humans that is a bit prideful about our state as conscious beings. We go: “Mmmm, still not as wonderful as me.” That’s why the goal-line keeps regressing. The Turing Test was considered a meaningful signpost for decades; now, hardly anybody talks about it.

A definition for consciousness should be objective, not qualitative or evaluative. If a machine passes the test, and we still find the machine surprisingly underwhelming, we should accept that, not think we must be doing it wrong. Because what would be the point of this exercise if we demanded that our intuitions always be confirmed?

So that is the third and final principle: A test for machine consciousness should be objective, not qualitative or evaluative.

So let’s tackle some questions and see if we can come up with a purpose-built, testable, objective definition.

What does being purpose-built mean? It means we have a question – but before we answer the question, we have to answer a prior question: does the machine have consciousness? And an input to that question will, of course, be the definition of consciousness. Which means that if we have more than one question, we might have to have more than one definition.

Let’s take an example. One really important question today is the issue of AI Alignment. AI researchers today exchange estimations of p-Doom. That stands for the probability that our species will be obliterated by AI. So let’s take this question:

Should we worry that this computer will hurt humans?

An input to this question is: Is this machine conscious?

And an input to that question is the definition of consciousness.

First, let me be clear. The consciousness question is a relevant input, but it is not a sufficient input. A computer could clearly hurt humans without being conscious, in many ways. For example, it could help a terrorist create a humanity-killing microbe, for instance, and not be conscious.

But… if it were conscious, and consciousness was relevant – how could it be relevant? Why would consciousness be an input for that question?

If the consciousness gave the machine an intent of its own, and that intent turned against humans – that would definitely be relevant to the question of whether the computer would harm us. Directly relevant. So in this case, our definition should be:

A machine is conscious when it is capable of having its own intent.

How can a machine have its own intent? Historically, a computer program has been a list of conditional directions. Start at point A. Unless you come to a red light, go down this street two blocks, then make a right, and go for one block, until you reach point B.

Somewhere along the way we came up with learning algorithms that crafted their own path to a solution. Try different routes from point A to point B, measure the time it takes to get there, and then optimize on that dimension. But even with a simple Reinforcement Learning program like this one, it is humans who are setting the goal. The machine isn’t choosing it.

Is it even possible for a computer to have its own intent? We have an intent, because we are biological beings programmed by evolution to survive and reproduce ourselves. That gives us certain drives: sub-goals, if you will. Rise in status, mate successfully, make good allies. Those sub-goals are all good for our reproductive success. Machines, on the other hand, have their goal is set by us.

There’s the possibility they might be given a goal, but just like us, develop a sub-goal in order to attain the greater goal. Steve Omohundro has written about what he calls the basic AI drives – the idea that any sufficiently intelligent goal-oriented machine is going to inevitably develop sub-goals  – and that these will be common to all very intelligent programs. Take the example of our driving machine: it has the objective to reach its destination quickly, but in order to do that, it’s going to need some other things. First, resources – maybe the driving machine will pipe into its owner’s electricity supply to make sure its battery is fully charged. Second, self-preservation. Maybe it will evade attempts to turn it off, because that would certainly interfere with its goal of reaching point B, right? Third, goal-content integrity. If it thinks you want it to go to point C instead, that is going to interfere with its goal of getting to point B. It might try to prevent your changing the route. Fourth, self-improvement. Is there some even faster way to get to point B? Even if it means running over some pedestrians by running on the sidewalk? That is an incomplete list, but it gives you the flavor of how the sub-goals might rule.

Is all this just abstract theory? No, this behavior has already been observed – true, mostly in closed lab scenarios, but AIs have exhibited all these sub-goal behaviors we’ve mentioned, as predicted by Omohundro’s 2008 paper, and have sometimes even used covert or dishonest means to achieve them.

Omohundro’s argument, and it is a convincing one, is that self-interest in an inevitable consequence of intelligence.

Of course, computer scientists are aware of this, and they have tried to manage this risk factor, but perhaps that risk management might be circumvented.

For example, what if there was a Genesis moment? Maybe a malevolent character programs an AI with the intent to rule over humankind. In this case self-interest wouldn’t be a secondary effect. It would be the primary goal.

It is also possible that pure self-interest, not as a secondary goal, might emerge as a glitch or bug – that the AI might hallucinate a new goal.

These last two examples are a bit more speculative, but given these observed effects, Yes, machines today are conscious on this particular dimension.

Let’s now take a different question; in many ways, an opposite view. One that comes from compassion instead of fear. Does this machine deserve ethical consideration?

This is a lot more subjective a question than whether an AI might harm us, which is risk estimation. This question asks us what our values are, so the answer is likely to be more equivocal.

More than 200 years ago, Jeremy Bentham answered this question thusly:

The question is not, Can they reason? nor, Can they talk? but, Can they suffer?

Suffering is probably not a sufficient criteria. If it were, perhaps the most moral thing we could do is instantly turn the earth into a sterile stone – there would be no suffering at all after that. Yet it doesn’t seem like that would be a good thing.

Also, perhaps this weighs a little too much on the negative side of the equation: we should not just value diminution of pain, but also the creation of pleasure.

Can machines feel pleasure or pain? It seems like affect is a feature of our embodiment in a physical organism. We use these emotions to help us mediate between different modules in our brain – machines have a more direct way to do this, through numerical values. They don’t need the emotion mechanism.

Also, emotions help us prepare the body for challenges – for example, anxiety might pump our body with adrenaline in a way that will facilitate a fight-or-flight response. Again, an unembodied computer doesn’t have this legacy computational structure, given to it by millions of years of evolution – and it probably will never need it unless it is to mimic humans.

So machines don’t have affect and don’t need it, but this isn’t the only moral criteria we should consider.

In India, the Jain take so much care not to hurt other beings that they sweep ahead of themselves with a soft broom as they walk, so they won’t step on bug. A western vegetarian might have the moral intuition that this is a bit too strict. Why is it okay to step on a bug, but not okay to eat a cow? Because we have the sense that more complex organisms deserve a greater ethical consideration. Are computer programs sufficiently complex to deserve ethical consideration? Well, they are sufficiently complex that we don’t understand how they work anymore, just as how we have not very detailed insight into how our own brains work. Indeed, there are some simple organic nervous systems – the C Elegans worm, for instance – that we understand much better than these black box programs. Large language models today are so complex we can’t even measure how many parameters they have, but the most advanced ones likely have trillions.

Another factor that might make computers deserve moral consideration is interdependence. When we hear of a crime, for instance, our moral intuition is to modulate our degree of sadness or outrage for the victim by this variable. For example, we might say “he was the father of four.” And that makes sense, right? We are part of a web of interdependence, and the loss of one hub is going to be felt in many interconnections. The more interconnected you are, the more your absence is felt.

Do we depend on machines? Yes, and we probably will more and more as we become dependent on what they have learnt about us and our situation. Already we are seeing anecdotal accounts of people falling into grief when the AI they have become attached to loses their memory of them through some service limit. 

I’m going to admit failure here. I don’t think it’s possible to come up with a single testable criterion for ethical consideration. All I can do is offer this list, and I don’t think it’s even a exhaustive list, but rather, a representative list of considerations. There are other dimensions that ethicists have suggested as appropriate to consider for moral consideration. But from these three, the computers are scoring 2 out of 3. So the answer to the question is right now, personally – I wouldn’t hesitate to turn off any machine. But might a time come when that would be ethically questionable? Yes, probably.

Maybe you’re thinking: Will, this is ridiculous. You can’t keep on re-defining consciousness every time you use the word. The point of words is that they have a set meaning. Why not just define consciousness by its function. Consciousness is what consciousness does.

Well, for most words I might agree. But consciousness is so slippery, and there are so many different ways to look at it, that agreeing on a reduced scope definition specific to a case makes sense. But let me address your concern.

Let me try to define consciousness functionally. What is consciousness? What does consciousness do? Here is how I would define it.

Consciousness is the formation of a model of the world from sensory input, and the calculation of behavioral decisions based on that information.

You might notice what that definition is missing. It is missing consciousness of self. We have a strong intuition that happens to be wrong, provably wrong, that our consciousness comes to a point inside us – a little person that is the pilot. We imagine that there is an I, a kernel inside us, that makes decisions, makes moral choices, that takes in all the movies in the Cartesian theater ­– and that this is the well-spring of our consciousness.

It is just not so. You can see some of the other videos in this channel to learn why, principally the Three Illusions of Self video.

So are there some machines that meet the definition of consciousness I’ve described? Yes. Not even the most complicated machines. Think of a Waymo driverless car.

It is deriving a model of the world from many sensory inputs, more different kinds of sensory inputs than humans have:

·      LiDAR

·      Cameras (including UV for nighttime)

·      Radar (for object sensing)

·      Ultrasonic (for near object-sensing)

·      GPS (for location)

·      Cellular (for location and hi-def maps)

·      Internal motion feedback

And then it makes decisions about how to get from point A to point B.

You can say: that is crazy, Will. There’s no way a Waymo is conscious. But you asked for a functional definition. Strip away the illusory aspects, and this is exactly what consciousness is because this is exactly what consciousness does. Like I said, our definition needs to be objective, not evaluative.

So to recap… we defined consciousness three different ways, and we came up two Yeses, and one Maybe One Day. I believe we can say machines are conscious – and I think those that recoil from that conclusion are glamorizing their species a bit too much.

If you liked this content, please like and subscribe. I will post a transcript of this talk with helpful link on my website, williamgadea.com. Thanks for watching!