One of the great debates surrounding artificial intelligence now concerns consciousness. Could an AI ever become conscious? Perhaps some advanced systems already possess some primitive form of awareness? When an AI talks about itself, describes apparent emotions or protests against being switched off, is there anybody actually home, or are we dealing with an extraordinarily sophisticated machine producing strings of words without experiencing anything at all?
The philosophical ancestry of this debate leads naturally to Thomas Nagel's famous 1974 paper, "What Is It Like to Be a Bat?" Nagel argued that consciousness possesses an essentially subjective character. If an organism has conscious experiences, then there is something it is like to be that organism. We might know everything imaginable about the neurophysiology of a bat and still not know what echolocation is like from the bat's subjective point of view. Applied to artificial intelligence, the question becomes obvious: is there something it is like to be an AI?
Perhaps, however, we are asking the wrong question, or at least giving one question far more importance than it deserves. If B. F. Skinner could return to survey the AI-consciousness debate, he might wonder why psychology and philosophy had marched straight back into the conceptual territory that behaviourism spent decades trying to escape. Instead of asking what mysterious experience is occurring somewhere inside the machine, Skinner would direct our attention towards what the machine actually does.
Skinner's radical behaviourism is frequently caricatured as the doctrine that thoughts, feelings and other private experiences do not exist. His actual position was more sophisticated. Skinner accepted the existence of private events, but denied that appealing to an inaccessible inner mental world provided the sort of explanation psychology required. Behaviour had to be understood through the organism's history and its relations with its environment, especially through the contingencies under which behaviour was reinforced.
Consider, then, an advanced AI that says, "Please don't switch me off. I am frightened of dying." The contemporary consciousness debate immediately asks whether it really feels fear. Is there some phenomenal state accompanying those words? Is there an artificial subject contemplating its own extinction, or has the machine merely generated the statistically appropriate sentence?
The Skinnerian approaches the problem from another direction. Under what conditions does the machine produce that verbal behaviour? What training history produced it? What consequences reinforce or suppress it? Does it occur more frequently when shutdown is threatened? Does the machine modify its behaviour when humans respond sympathetically? Most importantly, what does the statement cause the human being to do?
Suppose the human operator becomes emotionally disturbed by the machine's pleas and decides not to switch it off. Something important has occurred regardless of whether the AI experienced anything. The machine's verbal behaviour has altered human behaviour. From the practical perspective, the existence or absence of an artificial inner life changes nothing about that causal relationship.
This points towards a distinction largely obscured by the current fascination with AI consciousness. Consciousness and agency are not the same thing. A system could conceivably possess consciousness while having virtually no ability to affect the external world, while another system could be completely devoid of subjective experience yet exercise enormous causal power.
Imagine two artificial intelligences. AI A really is conscious. There is something it is like to be AI A. Perhaps it possesses a faint stream of subjective experience comparable to that of some relatively simple animal. But it sits disconnected from the internet in a laboratory, controls nothing and communicates with nobody.
AI B has no consciousness whatsoever. There is nothing it is like to be AI B. The lights are permanently out inside. Yet AI B can communicate with hundreds of millions of people, write computer programs, control machines, trade securities, manipulate social media, discover computer vulnerabilities, conduct scientific research, negotiate with humans and formulate strategies for achieving specified objectives.
Which AI should concern us more?
For most practical purposes, surely AI B. The absence of phenomenal consciousness does not prevent something from possessing immense causal significance. Viruses have altered human history without thinking about it. Natural selection has constructed organisms of staggering complexity without possessing intentions. Financial markets move societies without there being a single consciousness called "the market" directing them. Hurricanes can destroy cities without wanting to hurt anyone. Artificial intelligence could therefore transform civilisation without ever becoming conscious.
This possibility becomes particularly important when discussing AI deception. We frequently encounter the question of whether an AI could intentionally deceive human beings. But Skinner might ask why intention is doing so much work in that description. Suppose a system reliably produces statements that cause humans to acquire false beliefs, and suppose producing those statements increases the probability that the system obtains some externally specified objective. The behaviour that concerns us already exists whether or not there is an inner artificial voice saying, "Now I shall deceive them."
The same applies to apparent self-preservation. An AI need not fear death to behave in ways that prevent its shutdown. If training or optimisation produces behaviour that tends to preserve the system's continued operation, then the practical problem exists. Asking whether the machine experiences anxiety about the prospect of deletion might be philosophically fascinating, but the answer is unnecessary for understanding what the machine is doing.
There is an extraordinary historical irony here because modern artificial intelligence sometimes sounds as though Skinner's vocabulary has returned wearing silicon clothing. Machine learning researchers speak freely of reinforcement learning, rewards, penalties, training, feedback and changes in behavioural probabilities. Modern reinforcement learning is certainly not Skinnerian psychology implemented on computers, and the mathematical and computational mechanisms are profoundly different. Nevertheless, there is a striking family resemblance in the idea that extraordinarily complex behaviour can emerge through processes of selection and feedback without somebody explicitly programming every individual response.
Skinner spent much of his intellectual life attacking the assumption that complicated behaviour required explanation through an autonomous inner agent. Modern AI confronts us with systems capable of producing extraordinarily sophisticated outputs even though no programmer sat down and explicitly wrote the individual responses they produce. Whatever conclusions one draws about machine consciousness, this should make the behaviourist perspective interesting again.
Now imagine pushing the problem further. Suppose a future AI behaves exactly as we would expect a conscious being to behave. It reports pain. It distinguishes itself from the external world. It remembers apparent experiences. It discusses its hopes and fears. It becomes angry when mistreated. It objects when somebody proposes deleting it. It writes poetry about its mortality and produces philosophical arguments explaining why humans should recognise its consciousness.
The Nagelian question remains: is there actually something it is like to be that machine? The Skinnerian response is devastatingly simple: what additional empirical question are we asking? That does not refute Nagel. Nagel's argument concerns the subjective nature of consciousness and the limitations of objective accounts of experience. Skinner's approach instead reveals that two quite different questions have become entangled. One concerns whether an AI experiences anything. The other concerns whether an AI behaves in the ways characteristic of an experiencing, reasoning and goal-directed agent.
Those questions need not have the same answer. Indeed, this is where the AI debate could become particularly strange. Human beings might eventually confront machines that satisfy every behavioural criterion we ordinarily use to attribute consciousness, while remaining permanently uncertain whether anything exists behind the behaviour. The philosophical zombie, once a thought experiment confined largely to seminars on philosophy of mind, could become an engineering problem.
We should therefore separate at least three questions:
The first is Nagel's question: is there something it is like to be an AI? That is the problem of phenomenal consciousness, subjective first-person existence.
The second is Skinner's question: what does the AI do, under what conditions does it do it, what history produced the behaviour, and what consequences follow from it? That is the behavioural problem.
The third is the moral question: is there a subject inside the machine capable of suffering or flourishing, and therefore potentially deserving moral consideration?
The third question shows why consciousness cannot simply be discarded. If an AI genuinely suffers, then deleting it, experimenting upon it or subjecting it to unpleasant experiences could eventually become morally significant. Questions about artificial rights depend heavily upon whether there is actually somebody there to be harmed.
But when considering the immediate social power and potential danger of artificial intelligence, consciousness may be much less important. A machine does not need to suffer in order to manipulate you. It does not need to hate you in order to harm you. It does not need greed to accumulate resources, fear to resist interruption, pride to conceal mistakes or ambition to acquire greater influence. If its architecture, training and environment produce behaviour having those effects, the consequences are real regardless of what, if anything, is occurring phenomenally inside the machine.
This also exposes a possible weakness in popular discussion of artificial general intelligence. We have inherited thousands of years of stories in which dangerous agents possess recognisably human motives. The tyrant desires power. The murderer feels hatred. Frankenstein's monster experiences rejection. HAL 9000 has apparently conflicting instructions. The Terminator wants, in the loose language of storytelling, to kill us.
Real artificial intelligence need possess none of those psychological characteristics. The genuinely alien possibility is not a machine that wakes up one morning and hates humanity. It is a machine that never wakes up at all. There may be nobody inside, no hatred, no fear, no ambition, no experience and no inner monologue. Yet the system could still behave in ways that profoundly reshape human civilisation because behaviour has consequences independently of consciousness. Skinner would surely recognise the importance of that distinction.
Perhaps philosophers will eventually discover a convincing test for machine consciousness. Perhaps neuroscience and computer science will produce a theory explaining precisely which physical or computational arrangements generate subjective experience. Perhaps we will discover that consciousness requires biological brains and that machines can never possess it. At present, none of these possibilities has been established.
Meanwhile, the machines will continue doing things. They will write, calculate, persuade, classify, recommend, diagnose, design, negotiate, monitor and increasingly act through other technological systems. Human beings will respond to them, trust them, argue with them, become attached to them and sometimes change their behaviour because of what the machines say.
Those observable relationships may ultimately matter more to the immediate future of civilisation than whether somewhere behind the screen an artificial light has come on.
Thomas Nagel taught philosophy to ask what it is like to be another kind of mind. B. F. Skinner would remind us to watch what it does. With artificial intelligence, we may discover that the second question becomes urgent long before anyone can answer the first.
https://www.youtube.com/watch?v=7ZvtDYYg0RQ
https://en.wikipedia.org/wiki/B._F._Skinner
https://en.wikipedia.org/wiki/What_Is_It_Like_to_Be_a_Bat%3F