Reading: M. H. A. Newman, Alan Turing, Geoffrey Jefferson, and R. B. Braithwaite, “Can Automatic Calculating Machines Be Said to Think?” (BBC discussion, 1952).
Our reading is a transcript of a discussion recorded and broadcast by the BBC in January 1952. Alan Turing was a mathematician who had helped establish the theoretical foundations of modern computing, worked on codebreaking during World War II, and then helped build early electronic computers. Max Newman was a professor of mathematics at the University of Manchester. He had taught Turing at Cambridge, and subsequently worked with Turing on computing at Manchester. Geoffrey Jefferson had been professor of neurosurgery at Manchester. He had already argued publicly against the idea of machine thinking in a 1949 lecture, and Turing had responded to some of his criticisms in a 1950 paper. Richard Braithwaite, a Cambridge philosopher whose work included the philosophy of science, moderated the discussion.
Computers were still very new when this conversation took place. But the participants were already asking many of the questions we’re asking now: Could machines learn, reason, make mistakes, create new ideas, or have anything deserving to be called thought? And what would give us grounds for deciding?
When we ask whether octopuses have minds, we don’t assume that their mental lives will be exactly like ours. They might experience sensations or emotions that humans never have, and vice versa. Those differences can be interesting. But it’s also interesting to ask what mental states we might share: belief, preference, fear, curiosity, planning? (Of course, just because fear is a mental state we share with octopuses, doesn’t mean we need to be afraid of the same things.) We can also ask these same questions about AIs, alien organisms, or beings whose physical makeup is unknown. Our focus going forward will be on this question of shared kinds of mentality, rather than on whether another kind of mind would duplicate ours in every respect.
It’s useful to separate three questions that discussions of AI often run together:
Our present focus is on the second question. We’ll return to personhood (the third question) later in the course.
Note the difference between the second question and the first. We can at least envisage a limited system—one that would be unimpressive by the standards of AGI—with genuine mental states of its own. Conversely, we can envisage an extraordinarily capable system that performs at a superhuman level but lacks genuine thoughts or feelings of its own. Whether either case is really possible is a philosophical question, but the ideas are not contradictory just by definition. Some theories of mind will connect capabilities and genuine mentality much more closely than others.
Also, our main question is what artificial systems could in principle be like, not primarily what today’s AI models can do. We needn’t assume that today’s ChatGPT or Claude, or even the machines built in the next few years, exhaust the possibilities.
Consider an artist working in a nonverbal medium: a painting, a sculpture, or a piece of instrumental music. We might describe a painting as angry, say that a sculpture expresses grief, or hear a piece of music as conveying indignation at injustice. We needn’t mean that the artwork itself feels angry, grieves, or regards something as unjust. Nor need the artist have felt these emotions while producing the work, or ever have suffered the injustice represented. The artist might deliberately create a representation of an emotion they don’t presently feel. In such a case, the artwork’s meaning or emotional expression depends on its author’s intentions; it isn’t a mental state that the artwork itself experiences. We can call this a derived kind of meaning or intentionality.
Now consider a different case. Imagine that a student’s parents could use genetic engineering to arrange that she would be especially prone to anger in some particular kinds of circumstances. Perhaps they deliberately chose features of her genome to ensure that, when someone treats a friend unfairly, she will get angry. Years later she encounters just such a situation and feels anger. The parents’ earlier decision helped cause her reaction; they may even have predicted and wanted that reaction. Still, the anger is hers. She is the one who feels it. Her parents needn’t feel angry at all when she does, and her anger isn’t merely a representation of the anger they had in mind.
The distinction isn’t between mental states that have been caused by other people and states that haven’t. Our ordinary beliefs, preferences, and emotions are constantly influenced by parents, teachers, friends, advertisements, and other people. We don’t therefore regard those states as merely belonging to the people who influenced us. The distinction is between a state being genuinely possessed by a subject, however it came about, and an object merely expressing or representing such a state because of what another subject intended.
This creates a question about computers and programs. A programmer might design a system to print “I want to leave” or to identify certain facts as true. In some cases we may reasonably interpret such outputs as no more than expressions of what the programmer intended the system to say or represent—something like the meaning of a sentence in a book, or the anger expressed by a painting. But is that all there could be? Could a programmer also bring about a system that itself wants something, believes something, or makes plans, much as the genetically engineered student herself becomes angry? The fact that other people designed or trained the system does not, by itself, decide between these possibilities.
The philosopher John Searle used the phrase original intentionality for genuine intentionality possessed by a subject, as opposed to intentionality that’s merely derived from someone else’s. We’ve encountered intentionality before as the technical term for the aboutness or representational character of mental states: a belief is about how things are; a desire is for something; a plan represents a possible course of action. The word original may be misleading here. It doesn’t mean that the state arose without causes, that no one taught the subject to have it, or that no one intentionally arranged for it to occur. The engineered student’s directed anger, beliefs, and desires could still be her own in the relevant sense. It would be better to say genuine, non-derivative intentionality. The engineered student’s mental states exemplify that. The artworks do not.
A final qualification: beliefs, desires, and plans are especially clear examples of representational mental states. Whether a machine could also experience pain, have a visual sensation, or feel emotion are another issue. Don’t assume that all these questions must receive the same answers.
In a famous 1950 paper “Computing Machinery and Intelligence,” and then again in the debate we read, Turing proposes that instead of first settling on definitions of thinking or intelligence and then arguing whether machines satisfy them, we investigate a more specific question about what a machine could do. He called his proposal an Imitation Game, but it has nowadays come to be known as the Turing Test.
The basic idea is that a human interrogator communicates with hidden participants through typed messages. The interrogator cannot see or touch the participants, and must try to distinguish a computer from a human on the basis of their replies. There is no rule that the conversation must concern mathematics, factual questions, or any other restricted subject. An interrogator may ask about childhood, jokes, poetry, intentions, relationships, or whatever seems revealing. A machine may even try to make itself look more human by hesitating or making an occasional mistake. The test deliberately prevents a machine from failing just because it doesn’t look like a human or have a human body.
Turing’s 1950 presentation begins with a related game in which an interrogator tries to tell a man and a woman apart. In the 1952 debate we read, he describes a jury judging hidden human and machine respondents. The jury members are not supposed to be experts on machines. They must judge several exchanges, some of which are genuinely with humans, so they cannot succeed simply by calling everything a machine. Turing suggests that a machine would do well if a substantial proportion of the jurors were taken in. There isn’t one timeless, universally agreed numerical standard for what counts as “passing” every version of this test.
How impressive would it be to fool people this way? A great deal depends on whom the machine is talking to, for how long, and under what conditions. If a judge knows little about computers, or the conversation lasts only a few minutes, the judge might be easily misled. Early programs such as ELIZA produced superficially plausible conversational responses by matching patterns in what users typed and responding with stock prompts. That could make users feel understood, especially if they helped the conversation along, but the illusion tended to become fragile when people asked unexpected follow-up questions. Later staged Turing-Test competitions, such as the Loebner Prize, raised similar questions about how much it tells us to fool particular judges in limited exchanges.
Contrast such a fragile success with a hypothetical system that can sustain rich, appropriate conversation reliably, with knowledgeable and probing interlocutors, across unfamiliar topics and over long stretches of time. Its success wouldn’t just be something we happened to observe once. We would have reason to expect it to respond well in many other circumstances too. (We’ll say more next meeting about these expectations, or dispositions.) When we discuss the philosophical significance of passing the Turing Test, we should normally have this stronger sort of achievement in mind, not merely fooling a gullible person for five minutes.
Even then, we may distinguish the kinds of mental life the conversation provides evidence for. Someone might think that extended, sophisticated conversation is excellent evidence of a system’s beliefs, preferences, goals, or plans, but much weaker evidence that it experiences pain, sees colors in anything like our way, or feels emotions. We may sometimes use the convenient general label intelligence, but the question actually includes many kinds of mental state that might need to be assessed separately.
It’s important to keep three different claims apart. They concern different logical relationships between passing the test and genuinely having mental states.
| Claim | What it says |
|---|---|
| Passing is necessary for intelligence | If something is genuinely intelligent, then it must be able to pass the Turing Test. |
| Passing is logically sufficient for intelligence | If something can pass the test, then it must genuinely be intelligent. It would be impossible for a system to pass in the relevant way yet lack genuine intelligence. |
| Passing is good but defeasible evidence of intelligence | Passing gives us good reason for believing that the system is intelligent, but does not guarantee it. Other evidence could make us withdraw that judgment. |
First, necessity. If passing the Turing Test were necessary for thinking, then no being unable to converse as a human does could think. That would exclude octopuses and other animals to which we have already found reason to attribute mental states. It would also threaten to exclude human beings who cannot participate in the required conversation. So this is not a plausible condition on having any intelligence at all. It isn’t a claim that Turing, or any of our principal interlocutors, needs to defend. (One might separately ask whether passing a suitable language test is necessary for possessing some specific linguistic abilities. That would be a much narrower claim.)
Second, logical sufficiency. A defender of this claim says that a system with the right rich and reliable conversational abilities must have genuine intelligence. Someone could defend this on behaviorist grounds. (We’ll discuss this view more next class.) According to the relevant kind of behaviorism, there isn’t a further hidden fact about the internal workings of a system that makes the difference between really having a mental state and merely seeming to have it, once all the appropriate behavioral dispositions are present. If a system would react, converse, learn, and respond in all the relevant ways, then that’s all we could reasonably demand for having the mental capacities at issue.
This isn’t to say that a behaviorist must accept the verdict of a single short, poorly run contest. A contest only samples behavior; the behaviorist’s underlying claim concerns the full range of relevant dispositions. Still, this sort of view makes the question of what a system is capable of doing especially closely connected to the question of what mental states it genuinely has. It is one reason different people may disagree about whether assessing AI capabilities (Q1 above) is enough to answer our main question (Q2 above).
Third, good but defeasible evidence. Someone might instead say: If a machine passes a demanding test, that’s strong evidence that it really thinks, just as sustained, flexible behavior gives us grounds for attributing thoughts to other humans or animals. But evidence isn’t the same as a logical guarantee. It might turn out that the behavior is produced in a way that lacks some feature necessary for genuine understanding. Passing the Turing Test may provide evidence for those further features without being what constitutes intelligence.
Think of the difference this way. If you judge that another being is thinking because of the sophisticated things it says and does, you may have excellent reasons for that judgment. But you might still allow that facts you haven’t observed could count against it. By contrast, someone who regards the appropriate behavior as logically sufficient thinks that no further fact could make the system lack the relevant mental state while it has those behavioral dispositions. These positions agree that behavior matters, but disagree about why it matters and whether anything else matters too.
Where did Turing himself stand? His writings provide evidence for different interpretations. In our reading, he explicitly says that the proposed test is not the same question as whether machines think, though he regards it as close enough to be useful. Some of his remarks invite a behaviorist reading; others suggest that he’s proposing an informative way to investigate a further question, not claiming to have defined thought wholly in terms of behavior. There are arguments for both readings, and we don’t need to settle the historical controversy to understand the distinction.
If someone rejects the idea that conversational behavior is logically sufficient for mentality, they still owe us an account of what further feature might make the difference. Possibilities include the algorithm or internal organization producing the responses, the physical hardware or causal processes implemnting that organization, or perhaps something about the system’s history or origin. Not everyone who asks for more than behavioral similarity will ask for the same more.
Even a person who believes that thoughts require a soul cannot simply assume without argument that a soul could be associated only with a naturally evolved human body, and never with an artificial system. Questions about which kinds of systems could have souls would themselves need answers. We’ll leave that issue for later.
In our next meeting we’ll examine what computers and algorithms are, and why giving the same outputs for the same inputs—even across many different possible situations—is not the same thing as carrying out the same internal procedure. That will help us understand what a non-behaviorist might reasonably ask about a machine beyond whether it passes the Turing Test.