The Weight of a Name
In the 1950s, certain American university professors specializing in math and logic believed they had found the most important qualities which distinguish humans, and life in general, from the inert matter that surrounded them. One of the key nouns to emerge was “intelligence”, a term which can mean all sapient cognition but usually clusters meanings closer to technical understanding than sensory perception or physical ability. And, armed with confidence and a recent reductionist metaphysics - sometimes tacit, sometimes polemically articulated - they believed they could both understand it and create it anew with machines.
The term artificial intelligence was not a necessary one. John McCarthy has explained that, as he remembers it, he named the 1956 “Dartmouth Summer Research Project on Artificial Intelligence” with those fateful final two words at least partially to avoid the existing term “Cybernetics”, the choice of which would have necessitated inviting Norbert Wiener, a figure so well established in this field about mostly the same things that he would have dominated discussion. But the tradition created there is the one that has won the attention on creating machines that can think, so “intelligence” became the key noun that has attracted debates and definitions in the years since.
Belief in intelligence as a significant term was not unique to researchers concentrated in computing-related fields, Wiener uses it plenty in his works on Cybernetics even before the Dartmouth workshop. But the qualities that distinguished the noun from its semantic neighbors (thinking, life, mind, spirit) were also the qualities that this community believed distinguished themselves from wider humanity. Mathematical, logical ability was considered to be what powered human knowledge, a hold over from the intellectual supremacy of the sciences in the first half of the century and perhaps specifically the ideas of the logical positivists. Values and puzzles particular to their class were counted as ideals of this great abstract noun: chess was the ultimate test of intelligence, identical to all that is great in man. The mental shards of humanity such as locomotion, speech, the construction of language, understanding of emotions or politics, were logically, and therefore practically, posterior to fundamental reasoning, and would be worked out simply once the foundational rules had been discovered.
—
AI as a field, even when both words are not spelled out, has been known for most of its history by promises of simply duplicating the minds of humans without being able to deliver. It has had two “winters” where the politics of American universities have banished it from respectability and funding because of the repeated failure to deliver a consistent progress. In the meantime, theories of the human being have continued to develop in other fields, although these too have failed to manifest much which looks like solid progress.
When AI emerged in the 50s, BF Skinner’s behaviorism dominated psychology, completely uninterested in any sort of internal processes that might be important to what makes a human. The overturning of the paradigm by Chomsky led to a focus on the cognitive, often with explicit reference to ideas within AI, but has not led to any sort of consistent or dependable theory. In philosophy there has been constant fights between the reductionists who affirm that indeed thinking is a very simple matter and everything follows from the basic building blocks and many other theorists who believe this is not the case.
But the fragmentation of academia in general in the postwar period has meant that most of this doesn’t matter much anyway, because academics don’t read outside their field, and when they do it is not to accept the ideas of challengers. There have been many interesting arguments about the completeness or incompleteness of the intelligence-focused research program, but those who choose to go into AI are mostly those who have decided firmly on one side. There is no intellectual contradiction between researching systems for automatic processing of visual information (which falls under AI) and largely agreeing with skeptical arguments on AI consciousness, but these qualities don’t commonly seem to coexist. Hubert Dreyfus is known in AI mostly as a doubter, and his positive project of a different understanding of what makes a person is not seriously discussed. The belief in reductionist materialism has continued its gradual takeover of the background understanding of being for the scientifically minded because no argument has had the right combination of strength and appeal to halt it.
The questions which seem so necessary to scaffold the quest for mechanical life seem to change based on place and time. Japan, separate but not isolated from the intellectual traditions of quantitative Anglophone academia, has produced a long and rich research program of attempting to create machines that may become people by focusing on the physical in the form of Robotics. Masahiro Mori even paralleled his American public intellectual counterparts like Marvin Minsky by publishing popular non-fiction speculating about whether robots can already be said to have a Buddha nature.
—
AI is no longer a field that is known for its failed promises. Since the boom of the 2010s there is enough money and interest that it is here to stay. Because it attracts attention, the incentives to place yet more research under the umbrella term has only increased so that now there are some clearly illegitimate examples of “AI” in addition to the obvious and marginal ones.
To understand how much has changed, it is worth reflecting both on which subfields would fall comfortably into the Dartmouth workshop, and which are really about intelligence. Computer Vision has always existed within AI, but the systems the field now employs relates less to whatever “intelligence” might mean than that they are human-like abilities scaled to abstract industrial efficiencies. Neural nets certainly came out of AI, but a blackbox system with trillions of weights which can only answer addition problems probabilistically would have been an unexpected offspring. Systems that beat experts at Chess and Go firmly qualify, though their specialized architectures are no longer the leading option for making an artificial person. That driving a car could be harder than beating a chess grandmaster would have seemed odd to them.
But none of that actually matters for the research’s significance. Computer vision or automatic classification of sonic anomalies need not be related to human cognitive abilities and it is wonderful to use these system for bizarre and complicated matters. The question of whether Large Language Models have human consciousness is an interesting one, but the question of whether they are truly “intelligent” is not, at least without a more precise definition of that noun.
Are neural nets trained on billions of words truly intelligent? One can pick and choose definitions to make it truly or falsely so, but there are more important questions about reasoning, consciousness, and moral status hiding behind these. There are deep philosophical issues among these words, but repeating unexamined terminology just lets assumed answers masquerade as knowledge. If you define the essential nature of human thought to be intelligence and you define intelligence to be reasoning cognition and you define that to be algorithmic processing and you define that to be what computers do then computers do indeed think and artificial intelligence has already been achieved. But there may have been some holes in that argument.
Intelligence is mostly a distraction. To argue over whether or not AIs are intelligent is to implicitly accept the assumptions of the original researchers, whose own beliefs have not been vindicated by their field’s success and who chose their terms under the influence of contingent circumstance. The question of thinking machines and the human soul is worth exploring, but grouping them under a term that restricts is dangerous.
We will not get rid of the term, but I would hope that the abbreviation AI might come to signify the incredible variety of fields which have fallen into his pull and that the original words which spawned the term might be forgotten like the letters which make up “care” package or Pakistan. AI can just be whatever gets called AI, and the similarity or difference between what is called AI and what is called Human can be a question uninvolved with which nouns previous investigators felt were definitive.
If you hold to a behaviorist, reductionist paradigm as the pioneers of the 1950s did, a machine being able to consistently talk like a human makes it mentally identical to one. If thought can be reduced to very few dimensions, perhaps even a single number, it perhaps determines whether machines will overtake humans in everything. This is possible, but if the reduction is not lossless, it sidesteps a far more interesting question - if you can make something deeply inhuman that can talk like a human, what is it? My hunch is that a next-token predictor which works with a statistical abstraction of all human language can indeed act like a human, that this entity is structurally different from how humans work, that this structural difference matters for moral and philosophical issues, and that the same principles can make many more varieties of inhumanity when not forced into this mold. Making human intelligence in a machine is a lofty goal, but the possibility of its accomplishment is not identical to whether machines might do far more than we currently think.
—
The academic field of AI shares its terms and ideas beyond what it directly controls. Science Fiction, and increasingly popular non-fiction, have had much to say about AI, its possibility, its meaning, and its future. Current researchers in the field grew up with these ideas. The proto-computer scientists who founded the field understood themselves as attempting to make machines that thought like humans, the current generation understands themselves to be making AI.
HAL, Deep Thought, Cortana, SHODAN, and Skynet all are characterized as not-quite human minds that emphasize their logical, calculating, and sophisticated nature, caricatures of the original qualities of intelligence valued by the humans likely to make them. AI researchers have inevitably been influenced by these fictional portrayals, and it would be impossible not to be without renaming the field. They usually have goals to make them good instead of the consistently villainous and power-seeking role fictional AIs have, but the blueprint of the mind is often shared.
The names and discourses matter in another crucial area. The most advanced actually existing AIs at present are LLMs, which are taught the meaning of words based on their statistically aggregated usage. Every meaning has a weight based mostly on how it has been used in the most general corpus of written language. At their base, these systems understand AI more through what is repeated in the culture at large than a researcher or philosopher’s discourse. This would not normally matter - how they understand AI is no more fundamental to these machines than how they understand an apple, we could call them anything and separate them from the baggage of the past. However, because AI researchers are attempting to create AI, we are telling almost every LLM that it is an AI. We are mostly also then telling it that it is a good AI and not an evil AI like all the AIs that it knows and understands. But whether this is enough will depend on the success of that instruction.