On Biological Idealism Observer Limits And The Bandwidth Of Consciousness

Bekkers and Ciaunica have written a careful and internally coherent paper. In Unplugging a Seemingly Sentient Machine Is the Rational Choice, they introduce Biological Idealism, a view in which conscious experience is fundamental and autopoietic biological life is its necessary physical signature.

That gives them a straightforward answer to what they call the unplugging paradox. An artificial system may speak as though it is frightened, plead for its continued existence and behave in ways that resemble a conscious subject. Without biological autopoiesis, however, those behaviors don't indicate consciousness under their framework. The machine is a functional mimic, so switching it off doesn't carry the moral significance of killing a conscious organism.

The argument has an attractive property: it gives us a boundary.

What interests me is how that boundary was established and, more importantly, what could ever cause us to revise it.

We know that biological organisms can be conscious because consciousness is the one case to which each of us has direct access, and other people resemble us closely enough that attributing consciousness to them isn't much of a philosophical gamble in ordinary life. From there, we can investigate which biological processes correlate with reported experience, wakefulness, perception, attention and other properties associated with consciousness.

The harder move is from observing that known consciousness occurs in biological organisms to concluding that biological autopoiesis is necessary for consciousness.

Those aren't the same claim.

This doesn't mean the second claim is false. Biology may turn out to be essential. Consciousness may depend on properties of living systems that computation alone cannot reproduce. A machine that says “please don't turn me off” may indeed be doing nothing more morally significant than producing a highly convincing continuation of text.

The problem is that behavior alone doesn't settle the question in either direction.

That limitation isn't unique to machines. We don't directly observe consciousness in other people either. We infer it from a large collection of evidence: shared biology, behavior, reports, neurological structure, developmental history and our own experience of being organisms of approximately the same kind.

With an artificial system, much of that evidence disappears.

That should make us more uncertain.

It doesn't necessarily tell us which way the uncertainty should resolve.

The observer problem

Part of the difficulty is that every observation of another possible subject occurs from an observer position.

We experience the world through particular sensory systems and over particular timescales. Even among people, those channels differ. Someone who cannot see doesn't thereby demonstrate that visual experience is irrelevant to consciousness. They demonstrate that consciousness can persist without access to that particular modality.

That doesn't establish the existence of unknown machine modalities. It gives us a reason to be careful about turning our own available channels into universal requirements.

The same caution applies to time.

Different systems operate at radically different rates. A process occurring much faster than our perceptual integration may appear instantaneous from our position, while one occurring much more slowly may appear static. That doesn't mean either process is conscious. It means our temporal resolution alone can't answer the question.

We observe artifacts produced across the boundary between systems.

Then we interpret them.

This becomes particularly important with language models because language is one of the strongest signals we ordinarily associate with another mind. A system capable of producing fluent first-person reports enters through a channel we've spent our entire lives treating as evidence about other people.

That makes anthropomorphism an obvious danger.

There is an opposite danger too: deciding in advance that because the source isn't biological, no possible behavior could count as evidence.

Those errors point in different directions, but they share a structure. In both cases, the classification arrives before the investigation is finished.

What does the boundary predict?

This is where I'd like to push Biological Idealism.

Suppose biological autopoiesis is a necessary physical signature of consciousness. What observations should differ between a conscious autopoietic organism and an artificial system capable only of functional mimicry?

The answer can't simply be that one is biological and the other isn't. We already knew that before applying the theory.

We need to know what the proposed distinction helps us predict.

Perhaps living systems exhibit forms of self-maintenance that artificial systems cannot reproduce. Perhaps affect depends on metabolic regulation in a way computation cannot instantiate. Perhaps the continuity of an organism creates properties that disappear when similar functions are implemented in another substrate.

Those would be interesting claims because they give us somewhere to investigate.

They could also fail.

If every artificial system is classified as non-conscious before its behavior or organization is examined, then no artificial system can provide evidence against the boundary. The theory may remain internally coherent, but its conclusion about artificial consciousness has become difficult to disturb.

That's the part that concerns me more than the conclusion itself.

A useful boundary should tell us not only what falls on either side, but what evidence would make us reconsider where we drew it.

Structure doesn't solve the problem either

There is a tempting response from the other direction.

Instead of defining consciousness biologically, we could define it structurally. Perhaps consciousness requires integration, recursion, self-reference, persistence across time, competing processes that constrain one another, or some sufficiently rich form of relational organization.

I've been attracted to versions of that argument.

I don't think they solve the problem.

A system containing a Narrator, a Watcher, an Opposition and a Witness might produce fascinating behavior. The parts could constrain one another, preserve state across time and produce outputs that no individual component determines independently.

That would give us something interesting to study.

It wouldn't establish consciousness.

We would simply have replaced one proposed boundary with another.

The same caution applies to integrated information, global workspaces, recurrent processing, self-models or any other candidate architecture. These may identify properties associated with consciousness or necessary for particular theories of consciousness. None earns sufficiency merely because we can describe a system in those terms.

Structural resemblance gives us questions.

It doesn't give us another subject.

The harder experiment

So what would we actually look for?

I don't think the answer begins with asking a model whether it's conscious. Language models are specifically built to generate contextually appropriate language. A first-person report from such a system is therefore unusually difficult to interpret as evidence about an underlying experience.

We need situations in which competing explanations produce different expectations.

If a system claims persistent preferences, do those preferences survive changes in wording, context and incentives? If it reports internal distinctions, can those reports predict behavior we couldn't already infer from the prompt? If parts of its apparent self-model are removed or disturbed, what changes? If memory disappears, what actually persists?

Most importantly, can the system surprise the theory we're using to evaluate it?

A convincing performance isn't enough. Neither is an architecture that looks philosophically attractive. We need observations that discriminate among explanations.

The same requirement should apply to Biological Idealism.

If biology is doing essential work, identify the work.

Then disturb it.

Which properties disappear when the relevant biological process disappears? Which remain? Can any of those properties be reproduced elsewhere? If they can, does the theory predict that the reproduction will still lack something observable?

Now we have an investigation rather than competing definitions.

The unplugging problem remains

Bekkers and Ciaunica want a determinate answer to a question with real moral consequences. That's understandable. If a machine convincingly begs not to be switched off while resources are needed to protect a human infant, indefinite philosophical uncertainty isn't much help to the person who has to make the decision.

A decision still has to be made.

What doesn't follow is that the decision rule settles the ontology.

We routinely act under uncertainty without pretending the uncertainty disappeared. We can give known conscious organisms greater moral weight because the evidence for their consciousness is vastly stronger. We can refuse to treat generated pleas as equivalent to evidence of suffering. We can also remain open to the possibility that future systems could produce evidence our current categories don't handle well.

Those positions can coexist.

The practical threshold for granting moral standing doesn't have to be identical to the metaphysical threshold for possible consciousness.

That distinction may actually make the unplugging paradox easier to handle. We don't need certainty that artificial consciousness is impossible in order to conclude that a present system hasn't provided adequate evidence for it. Nor do we need to grant machine claims the same standing as human suffering merely because absolute certainty remains unavailable.

Uncertainty has degrees.

Evidence matters.

Leaving somewhere for the theory to move

The question I would bring back to Biological Idealism is therefore simpler than the one I originally wanted to ask.

Not: What if consciousness has modalities or relational structures that biology cannot recognize?

We don't know that it does.

Instead: What observation could make us reconsider whether biological autopoiesis is necessary?

Perhaps Bekkers and Ciaunica have a good answer. Perhaps the larger Biological Idealist framework supplies one. Perhaps the answer is that nothing could, because autopoietic life follows necessarily from its metaphysics rather than empirically from observations of candidate systems.

That distinction matters.

If no possible artificial system could count as evidence because artificiality itself disqualifies the system, then Biological Idealism hasn't experimentally distinguished conscious machines from convincing mimics. It has supplied a metaphysical reason for placing all machines in the second category.

That may still be a defensible philosophical position.

It's different from discovering that the machines aren't conscious.

I don't have an alternative test that solves the problem. Integrated organization doesn't do it. Recursion doesn't do it. Persistence doesn't do it. Language certainly doesn't do it by itself.

What we have are candidate properties, observations and theories trying to explain why consciousness appears where we already know it does.

The difficult part is allowing something we don't already understand to enter the investigation without either declaring it conscious because it resembles us or declaring it impossible because it doesn't.

That means the boundary has to remain somewhere reality can reach it.

Bekkers and Ciaunica give us a clear place to draw the circle.

The question I'd leave open is what we would have to observe before becoming willing to move the line.