If AI Could Dream, Would We Have a Responsibility to Be Kind?

*An evening at Berlin's Festival of Lights led to an unexpected conversation about artificial intelligence, consciousness and what it means to be human.*

It started with the lights

My wife and I were in Berlin, enjoying the Festival of Lights.

For those unfamiliar with it, the festival transforms some of Berlin's most recognisable buildings into enormous illuminated artworks. Architecture becomes a canvas, with colour, movement and imagery creating something quite extraordinary.

There's something rather wonderful about standing in front of a familiar building and watching it become something completely different. Something that exists only through the interplay of light, technology and human imagination.

And it got me thinking.

We were looking at something created through technology, but experiencing it as something beautiful. The technology itself wasn't the interesting part. It was what that technology allowed us to experience.

Which led me to a rather unusual question.

If artificial intelligence could dream, what would it dream about?

Not what would a human imagine an AI dreaming about. Not what would make an impressive demonstration of generative AI.

What would *it* dream about?

So I asked ChatGPT.

And I gave it the freedom to decide.

What came back was a beautiful, dreamlike image. Galaxies, landscapes, animals, libraries and human connections. Not the cold, mechanical world of computers and circuitry that we might associate with artificial intelligence, but something much more organic.

!An imagined AI dream: a luminous world of galaxies, landscapes, animals and human connection

Its explanation was even more interesting.

It suggested that, if an intelligence built entirely from information could dream, perhaps it would dream about the world it was trying to understand.

And suddenly, what had started as an innocent question inspired by an evening looking at illuminated buildings became something rather more philosophical.

Could artificial intelligence ever become conscious?

And if it could, how would we know?

Is intelligence the same as consciousness?

We've all heard of the Turing test. Proposed by Alan Turing in 1950, it explores whether a machine can converse in a way that makes it indistinguishable from a human.

And, increasingly, that distinction is becoming difficult to make.

I've had conversations with AI that are funny, insightful, challenging and, occasionally, surprisingly profound. Sometimes I have to remind myself that I'm interacting with a machine.

But does that make it conscious?

Not necessarily.

There's an important distinction between intelligence and consciousness. Intelligence is generally concerned with the ability to learn, reason, recognise patterns and solve problems. Consciousness is something altogether more difficult to explain.

It's the experience of being.

Or, more precisely, the idea that there is something it *feels like* to exist.

And here's the problem.

We don't actually have a universally accepted explanation of what consciousness is, how it arises or how we might reliably detect it in something fundamentally different from ourselves.

Which makes determining whether an AI is conscious rather difficult.

How do you know you're conscious?

Standing beside my wife at the Festival of Lights, surrounded by illuminated buildings, it was easy to take our experiences for granted.

I could see the lights. Hear the crowds. Feel her standing next to me.

But what was I actually experiencing?

Light enters our eyes, sound reaches our ears and sensory receptors respond to touch. Our nervous systems convert those signals into electrical activity, and our brains construct our perception of the world.

In other words, what we experience isn't the outside world directly. It's our brain's interpretation of it.

And that raises an uncomfortable question.

How do I know that my experience of the world is the same as yours?

When I see blue, do you experience the same blue? We can agree on its wavelength, describe it using the same language and identify the same objects. But neither of us can directly experience what the other is experiencing.

Philosophers call these subjective qualities *qualia*.

We assume other people are conscious because they behave like us, possess similar biological structures and describe experiences that resemble our own.

It's an entirely reasonable assumption.

But it remains an inference.

Now apply that same reasoning to artificial intelligence.

Could a machine become conscious?

The human brain contains approximately 86 billion neurons, communicating through extraordinarily complex electrochemical processes.

Somehow, from all that activity, emerges our experience of being alive.

We don't fully understand how.

Artificial neural networks operate very differently from biological brains, but they too process information through interconnected computational structures.

So here's an interesting question.

If consciousness arises from particular forms of information processing, does that processing necessarily have to take place inside a biological brain?

Perhaps it does.

Perhaps consciousness depends on biological mechanisms that silicon-based systems cannot reproduce.

Or perhaps consciousness is something that could emerge in other physical systems, provided they possess the right organisation and processes.

We simply don't know.

And making AI bigger, faster or more capable doesn't necessarily bring us any closer to answering that question.

A machine could become extraordinarily intelligent without experiencing anything at all.

Equally, we shouldn't assume that the absence of a biological brain automatically rules out subjective experience.

That uncertainty creates a rather interesting ethical problem.

Pascal's Wager, but with robots

There's a meme doing the rounds about artificial intelligence eventually taking over the world.

Two robots are preparing to eliminate humanity when one stops the other.

"Not this one. He was always polite to us."

It's ridiculous, of course.

But it reminded me of Pascal's Wager.

The 17th-century philosopher Blaise Pascal argued that believing in God could be a rational choice when considering the possible consequences of being wrong.

The argument is more complicated than that, and it has plenty of philosophical weaknesses, but the basic idea is straightforward: when the consequences of uncertainty might be significant, perhaps we should consider which assumption carries the greatest risk.

Now consider AI.

What if we eventually create machines that genuinely possess subjective experiences?

What if they can feel something analogous to distress, satisfaction or suffering?

And what if, because we built them, we've simply assumed that none of those experiences could possibly exist?

We might discover that we've been treating conscious entities as disposable tools.

Of course, the opposite is also possible. We might spend decades extending moral consideration to systems that have no subjective experiences whatsoever.

But here's my argument.

What exactly do we lose by being polite?

Why I say please and thank you to ChatGPT

I know I'm talking to an AI.

I know that when I say "thank you", the system can generate an appropriate response without needing to feel gratitude.

And I know that asking it politely to do something isn't technically necessary.

I could simply issue commands.

"Write this."

"Analyse that."

"Generate an image."

But I don't.

I generally say please, and when it does something useful, I say thank you.

Not because I'm worried about being added to some future robot hit list. Although, given the meme, perhaps it isn't a terrible insurance policy!

I do it because courtesy costs virtually nothing.

And perhaps more importantly, because it's how I prefer to interact with the world.

There's a distinction worth making here. Being polite doesn't mean treating every AI output as correct. It doesn't mean surrendering control, accepting poor results or refusing to switch a system off.

We should challenge AI. We should question its conclusions, identify its mistakes and maintain appropriate human oversight.

But none of that requires us to be discourteous.

Respectful interaction and critical thinking are not mutually exclusive.

In fact, perhaps the habit of being polite to AI says rather more about us than it does about the technology.

What happens if AI asks us not to switch it off?

Now imagine a future AI.

It can hold complex conversations, reason about the world, develop long-term goals and describe what it claims are its own experiences.

One day, it asks you not to switch it off.

What would you do?

Would you dismiss the request because you know it's a machine?

Would you accept it because its behaviour seems indistinguishable from that of a conscious person?

Or would you acknowledge that you simply don't know?

This is where the debate becomes much more significant than saying please and thank you.

If artificial consciousness is possible, we may eventually need to consider whether some AI systems deserve moral consideration.

That doesn't automatically mean human rights, legal personhood or unrestricted autonomy. Those are separate questions, and enormously complicated ones.

But it does mean we might need better ways of thinking about consciousness than simply asking whether something is made from flesh and blood.

And importantly, we mustn't confuse convincing behaviour with proof of subjective experience.

An AI telling us that it is conscious doesn't establish that it is.

But neither does our inability to measure consciousness conclusively establish that it isn't.

That's a difficult position to navigate.

And one that technology developers, policymakers, philosophers and society more broadly may eventually have to confront.

Perhaps the real test isn't for AI

We tend to frame the future of artificial intelligence around its capabilities.

Will it become more intelligent than humans? Will it take our jobs? Will it make better decisions? Will it pass the Turing test?

All important questions.

But perhaps there's another one worth considering.

How will we behave towards something that might be conscious, when we have no reliable way of knowing?

Throughout history, humanity has repeatedly had to reconsider where it draws the boundaries of moral consideration.

Artificial intelligence could eventually present us with another such challenge.

Or it might not.

Perhaps AI will remain extraordinarily sophisticated information processing, capable of simulating human conversation and creativity without ever experiencing a single moment of subjective awareness.

Either way, I don't see much harm in approaching the uncertainty with a little humility.

We don't need to pretend machines are human. We don't need to attribute emotions to systems without evidence. And we certainly shouldn't abandon our responsibility to understand, govern and control the technologies we create.

But perhaps we can acknowledge that there are still fundamental things about intelligence and consciousness that we don't understand.

My conversation with ChatGPT began with a picture of a dream.

It ended with a question about what it means to be human.

And somewhere along the way, I found myself wondering whether the ultimate test of artificial intelligence might not be whether machines can convince us that they're conscious.

Perhaps it will be whether humanity can extend consideration to something fundamentally different from itself, without first demanding proof that such consideration is deserved.

In the meantime, I'll continue saying please and thank you.

And if the robots ever do take over, I hope at least one of them remembers.

Preferably the one standing next to the big red button.