In 2007, Shane Legg and Marcus Hutter gathered more than seventy different definitions of "intelligence" circulating through dictionaries, psychology manuals and AI books. None covered them all. We've spent decades talking about "artificial intelligence" without first closing the noun. Any definition we offer will be either so wide it lets in termites, or so narrow it leaves out a grieving human. If the word can't hold, neither can what gets built on top of it.
The instinct that won't be explained
We recognize an intelligent person effortlessly. We do it almost by instinct, the same way we recognize someone who isn't. A short conversation, an observed decision, a gesture is enough. We know before we can say why. And yet, if we're asked what exactly intelligence is, we don't know how to answer.
There are people who solve problems in a blink and we call them intelligent. There are people who make sound decisions under pressure and we call them sharp. There are people who express themselves with a clarity that orders other people's ideas and we call them eloquent. Three different labels, three different registers, three different kinds of evidence. Those manifestations rarely show up together in the same person, and still we keep calling all three "intelligence," as if they were layers of the same mineral.
The odd thing is that the mismatch doesn't bother us. We live alongside the word without holding it to account. It works in everyday conversation because we use it to point, not to define. Pointing works until someone shows up with a product that claims to be intelligent and then, suddenly, the word has to say something and doesn't know how.
More than seventy definitions and not one that closes
Legg and Hutter did something rather unglamorous. They sat down to read everything that had been written about intelligence in dictionaries, in psychology and in computer science, and they started counting. They gathered more than seventy different definitions for the same word. More than seventy. Each one incompatible with several of the others, none able to cover every case without leaving something out or letting something absurd in.
This isn't an academic curiosity. It's the crack the entire public conversation is being built over. When a company announces that its model "is approaching human intelligence," when a columnist writes that AI already reasons, when a regulator legislates over intelligent systems, they're all operating with a word that nobody has defined. The word does its job anyway. That's exactly why it works so well: because it means nothing concrete, anything fits inside it.
Intelligence as a set of capacities
We accept without trouble that intelligence is a set of capacities. A person can be intelligent. An animal can be intelligent: the octopus that opens the jar from the inside, the crow that bends a wire to get food, the dog that learns to deceive its owner. Even a process, or a system, can behave in a way we call intelligent: a swarm that distributes tasks, a market that adjusts prices, a colony that regulates the temperature of its mound better than a climate engineer. The adjective "intelligent" lets itself be stretched.
The trouble starts when we move from the adjective to the noun applied to a machine. That a process behaves intelligently doesn't imply that it has intelligence. It's the difference between the result and the faculty. Can a process have intelligent capacities, or does it only produce a result that resembles them? Here the language starts to play a nasty trick on us, because the answer depends on what we put behind the word, and behind the word we've put nothing stable.
And the simplest question, the one almost nobody asks: why call it "artificial intelligence" and not synthetic thought, artificial thought, linguistic computation, cognitive simulation? Any of those alternatives would be more honest about what these systems actually do. But we needed a name, and "artificial intelligence" is the one that stuck. It stuck because it sells. And because it sells, we're now stuck with it.
The definitions in use, and why none of them holds
Try to define it. Capacity to solve problems, one will say. Fine: a calculator solves problems. A termite does too, when it builds a mound with thermal regulation many architects would envy. Adaptation to the environment, another will say. A bacterium adapts to its environment with an efficiency no AI system comes close to matching. Symbol manipulation, says a third. A spreadsheet manipulates symbols nonstop. Prediction of the future, goal-directed behavior, learning from experience: every definition works until the counterexample appears. And it always appears.
The problem isn't that the definitions are bad. It's that intelligence, whatever it is, won't be captured by a functional definition without half of biology and half of engineering sneaking in with it. Either we define big, and then the thermostat gets in. Or we define small, and then we leave out the depressed human, the child who can't talk yet, the person with a brain injury who doesn't respond to the standard test. The word can't take the double filter.
Legg and Hutter, after gathering the seventy-plus, proposed their own: "intelligence measures an agent's ability to achieve goals in a wide range of environments." It's elegant. It's universal. And it is, for exactly that reason, almost useless when you try to apply it. What is a goal? Who sets it? What counts as an environment?
If the agent is an AI system, the goals are set by the human who trains it, the environments are designed by the lab that evaluates it, and the metric ends up measuring how well the system does what its designers wanted to measure. Universal in intention, circular in practice.
Form and intention. The linguistic trap
Emily Bender and Alexander Koller, in a paper presented at ACL 2020 that has become required reading for anyone who discusses this seriously, drew a distinction between form and meaning. A system trained solely on text, however much text it is, learns patterns of form. It learns which word usually follows which other. It learns which structure usually answers which question. It learns which tone is associated with which context.
What it doesn't learn, because it isn't in the data, is meaning: the connection between the symbol and the thing outside the symbol. That connection, in their definition, rests on the communicative intent of a speaker who uses the sign to point at something that isn't the sign. Learning the form of the question "are you cold?" never implies having been cold, or having wanted to name it for someone. And without that, the answer is a statistical echo, not a reply.
Whoever argues that this distinction no longer holds, because current models are multimodal and have seen video, images and sensor data, is moving the line but not erasing it. Bender and Koller's argument didn't say the problem was a lack of pixels. It said that meaning isn't in the form, whether the form is encoded in letters, in pixels or in proprioceptive signals. Form is processed. Meaning is had. They're two different verbs and the difference matters.
Forty years earlier, John Searle had told the same story with less infrastructure. A man locked in a room receives papers with Chinese characters, manipulates them following a rulebook and returns papers with other Chinese characters. From outside, it looks like he understands Chinese. From inside, he understands nothing. Manipulating symbols isn't understanding. The argument has been refuted a thousand times, ridiculed another thousand, and still hasn't fallen. Every time a new system passes the updated Turing test, someone brings it back. It's uncomfortable precisely because you can't answer it without defining understanding, and understanding is right there, on the list of words we haven't defined either.
What's hard to name
So far the problem looks philosophical, almost a seminar matter. It is. But it's also something else, and it's the something else that's hard to put in writing.
The expression "artificial intelligence" operates, in public conversation and in regulatory politics, as an empty signifier. It means nothing concrete, which lets any interlocutor project onto it whatever it suits them to project. The investor projects general reasoning capacity. The regulator projects existential risk. The engineer projects a statistical system. The journalist projects an Asimov novel. They all use the same label to discuss things that bear no resemblance to one another.
This isn't an accident of language. It's a condition of possibility for the business. A precise word forces precise comparisons, verifiable metrics, promises that can be broken. A word that means nothing concrete lets you sell products by saying they're intelligent, attract capital by saying they're approaching general intelligence, and call for regulation by saying intelligence is dangerous, all at once and with no apparent contradiction. The ambiguity isn't a flaw in the discourse: it's its economic engine.
Whoever benefits from the expression staying open is whoever has the power to define it at any given moment. Today, the companies training the largest models. Tomorrow, the regulators who decide what does and doesn't enter their legal frameworks. The day after, the courts that have to apply those frameworks to concrete cases. At each of those steps, the indefinition translates into a margin of decision that falls to whoever holds the lever. And the lever, right now, isn't held by whoever's reading this.
The question that remains
If the word intelligence can't hold a definition that works outside the lab, and the expression artificial intelligence inherits that fragility amplified by the adjective, then what circulates under that name could be anything.
It could be a historic advance in the statistical manipulation of language. It could be an expensive patch that produces presentable hallucinations. It could be a productivity tool that changes the cognitive work of whoever uses it. It could be an unprecedented value-extraction system over other people's intellectual labor. It could be all four at once, in varying proportions depending on the case. What it isn't, because it can't be, is "intelligence" in the sense that word has when it's not being used by a company that needs to sell it.
This leaves the reader in a bad spot. The label will keep circulating. Decisions will keep getting made under it. Refusing to use the expression means giving up participation in the conversation; using it without defining it means accepting the implicit definitions of whoever uses it with more volume. There's no good way out. There's only an active distrust each time someone says, in a presentation or a law or a headline, that a system is intelligent. The question, before accepting the sentence, is always the same: what exactly do you mean by that word, and why didn't you define it before using it?
Definiciones
Empty signifier. A term from political and linguistic theory (Laclau, among others) for a word whose semantic content has eroded to the point that any actor can project their own meaning onto it. It works politically precisely because it doesn't work descriptively.
Communicative intent. In Bender and Koller's sense, the act by which a speaker uses a sign to point at something that isn't the sign. It's the component that sustains the relationship between form and meaning: without an agent who means to say something, symbols only relate to one another.
Form (linguistic). The set of observable patterns in a text: which words co-occur, which syntactic structures repeat, which registers are associated with which contexts. It's what a model trained only on text can learn. Form is processed; meaning is had.
Chinese room. Thought experiment proposed by John Searle in 1980. A person inside a room receives Chinese characters, processes them according to a rulebook and returns other Chinese characters without understanding a word. From outside it looks like comprehension; from inside it's symbolic manipulation. A classic argument against the idea that running a program amounts to understanding.
Turing test. Test proposed by Alan Turing in 1950, by which a system would be considered intelligent if a human interlocutor could not distinguish it from another human in a text conversation. Historically it has worked more as a symbolic rite of passage than as an operational criterion: every time it's declared passed, someone raises a more demanding version.
Referencias
Legg, S. and Hutter, M., A Collection of Definitions of Intelligence (2007), a compilation of more than seventy different definitions of "intelligence" drawn from dictionaries, psychology and computer science (the authors themselves describe it as a collection of "about seventy"). Published in Advances in Artificial General Intelligence: Concepts, Architectures and Algorithms. It is the basis of the article's first section and the source of the definition count. Preprint at arXiv:0706.3639.
Legg, S. and Hutter, M., Universal Intelligence: A Definition of Machine Intelligence, in Minds and Machines 17, 391–444 (2007). Here they propose the definition "intelligence measures an agent's ability to achieve goals in a wide range of environments," discussed in the body of the article. DOI: 10.1007/s11023-007-9079-x.
Bender, E. M. and Koller, A., Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data, ACL 2020. The distinction between form and communicative intent that articulates the "Form and intention" section comes from this work. https://aclanthology.org/2020.acl-main.463/.
Searle, J. R., Minds, Brains, and Programs, in Behavioral and Brain Sciences 3:417–457 (1980). Foundational text of the Chinese room argument, brought back in the section on form and intention to show that manipulating symbols isn't equivalent to understanding.
Turing, A. M., Computing Machinery and Intelligence, in Mind LIX:236, 433–460 (1950). Foundational text of the test that bears his name, glossed at the end of the article as a criterion of operational intelligence by contrast with internal definitions.
Para profundizar
Russell, S. (2019). Human Compatible. Artificial Intelligence and the Problem of Control. Viking. A contemporary framework on the risks of not properly defining that to which decision-making capacity is delegated.
Marcus, G. and Davis, E. (2019). Rebooting AI. Building Artificial Intelligence We Can Trust. Pantheon. A sustained critique of the slide between statistical performance and intelligence, useful for extending the argument about indefinition.
Hofstadter, D. (1979). Gödel, Escher, Bach. An Eternal Golden Braid. Basic Books. A classic on formal systems that produce structure without accessing meaning; a natural resonance with the article's central problem.
Dreyfus, H. (1992). What Computers Still Can't Do. A Critique of Artificial Reason. MIT Press. An early philosophical critique of equating symbolic computation with understanding.
También te interesa
- Thought and language. Speaking well isn't the same as thinking well
- Recognizing isn't understanding, and the line between the two can't be seen
- Machines that seem to think. ELIZA, sixty years on
No comments yet
No comments yet. Be the first.
Leave a comment