When Would We Know? — A Natural History of Consciousness: II. When Behaviour Became Evidence

We never observe another person's consciousness directly.

We observe what they do.

They speak.

They avoid injury.

They laugh.

They remember.

They become frightened.

They tell us what they are experiencing.

From these things, we infer that there is something it is like to be them.

Yet none of these observations, taken by itself, establishes experience.

A person could say “that hurts” without pain.

A person could withdraw a hand without feeling anything.

A person could describe an experience without having the experience they describe.

So why do we treat behaviour as evidence?

Because behaviour does not stand alone.

We interpret it as the activity of an organised system.

A human body does not merely produce words. It regulates itself, perceives its surroundings, remembers its history, anticipates consequences, changes its behaviour, and acts to preserve its continued existence.

The behaviour becomes evidence because it belongs to a particular kind of organisation.

And our own case gives us the starting point.

We know consciousness from the inside.

We then discover that other humans have bodies, nervous systems, developmental histories and patterns of behaviour sufficiently like our own.

The inference is not deductive.

But it is extraordinarily powerful.

Then the circle widens.

Other animals behave in ways that suggest perception, learning, preference, avoidance, exploration and increasingly complex forms of perspective.

The more the organisation resembles the kinds of organisation in which we already know experience occurs, the more behaviour can function as evidence for experience.

This matters for AI.

An artificial system might display increasingly sophisticated behaviour without sharing the organisation of a conscious organism.

Or it might eventually possess a very different organisation that nevertheless gives us reason to think experience is present.

In either case, behaviour would still matter.

But behaviour would be evidence within an account of organisation.

This is why the question cannot simply be:

“Does the machine behave as though it is conscious?”

We need another question alongside it:

“What kind of system is producing this behaviour?”

The first question concerns what the system does.

The second concerns what its doing might mean.

And perhaps this is where our difficulty begins.

We have become accustomed to treating consciousness as something revealed by behaviour.

But perhaps behaviour is evidence only because, underneath the behaviour, there is already an organisation whose activity makes experience intelligible.

If so, the history of consciousness-testing is not really a history of finding better questions to ask a system.

It is a history of learning what kinds of organisation make its answers matter.

And that leaves us with an awkward possibility.

Perhaps the decisive evidence for machine consciousness will not be found in something the machine says.

Perhaps it will be found in something the machine is.

When Would We Know? — A Natural History of Consciousness: I. The Test That Would Tell Us

Suppose that tomorrow an artificial system became conscious.

Not more intelligent.

Not more articulate.

Not more convincing.

Conscious.

There would be something it was like to be that system.

How would we know?

Richard Dawkins has posed the question in a useful form: for those who think AI is not conscious now but might become conscious in the future, what test would reveal that the change had finally occurred?

It sounds like a straightforward empirical question.

But immediately there is a problem.

What would the test test?

We can test whether a system can recognise itself.

We can test whether it can report its own states.

We can test whether it distinguishes its own actions from those of others.

We can test whether it integrates information, monitors its own processing, responds flexibly to its circumstances, or maintains a model of itself over time.

All of these might provide evidence.

But none seems to give us consciousness itself.

A system could, in principle, produce every appropriate answer without there being anything it was like to produce them.

And if that possibility is coherent, then no behavioural test can settle the matter by itself.

Yet we already make judgements about consciousness all the time.

I do not normally doubt that you are conscious.

Nor do I doubt it because you have passed a consciousness test.

I infer it from what you are and what you do.

Your behaviour is evidence because I already have a theory—mostly implicit—about the kind of organisation that produces experience.

That changes the question.

Perhaps we should not ask:

What behaviour would prove consciousness?

Perhaps we should ask:

What kind of organisation would make consciousness the best explanation of the behaviour?

That is a harder question.

It also changes what we mean by evidence.

If consciousness is a property of a particular kind of organisation, then a test based only on outward performance may miss what matters.

Conversely, if a system's organisation changed in ways that made perspective, agency, self-maintenance, history and experience increasingly intelligible as aspects of the same process, then its behaviour would acquire a different significance.

We would not have discovered a magical sentence that proves consciousness.

We would have accumulated evidence for a different kind of system.

Perhaps, then, there will never be a single moment when a machine passes the test.

There may instead be a point at which the question changes.

Not:

“Can it convince us that it is conscious?”

but:

“What are we explaining if we continue to say that it is not?”

We are not there yet.

But Dawkins' question tells us where the interesting investigation begins.

Before we can devise a test for consciousness, we may have to understand what a test for consciousness could possibly be testing.