We have called it artificial intelligence since 1956. That year, four researchers met at Dartmouth College — John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon — and wrote the proposal for a summer research project that would give the field its name. Their core claim was technical, not philosophical: "Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

That was a bet on simulability. Not a claim about consciousness. Not a promise that machines would feel or understand the way people do. That charge came later — through science fiction, through marketing, through the way we talk about technology that answers us in full sentences.

Seventy years on, millions of people sit down with ChatGPT, Claude, Gemini, or Grok and feel like they are talking to something clever. The question that follows is more uncomfortable than it sounds: is "artificial intelligence" still the right term for what these systems actually do? Or does another word — "artificial cleverness," or in the more demanding philosophical sense, "practical wisdom" — describe more precisely what we are experiencing?

What Do We Actually Mean by Intelligence?

Definitions of intelligence are surprisingly abundant. One has become especially influential in AI research: in 2007, Shane Legg and Marcus Hutter defined it as an agent's ability to achieve goals across a wide range of environments. Deliberately measured by performance — not by how that performance comes about. Whether the system experiences anything along the way is irrelevant to the definition.

This is not a fringe view. It is the most widely cited approach in the field. And it has an uncomfortable consequence for this article's starting premise: by this definition, "artificial intelligence" is not wrong for today's language models. A system that reliably achieves goals — answering a question, writing code, summarizing a text — is acting intelligently in this sense, regardless of whether it understands or experiences anything while doing so.

The problem sits elsewhere. In everyday language, "intelligence" almost never means only goal-directed performance. We mean understanding. Consciousness. A counterpart that grasps its own answers. The technical definition and the everyday one have drifted apart — and a language model that answers fluently and confidently in full sentences automatically triggers the everyday meaning in us, not the technical one.

The Strange Thing About Language Models

What these systems can do today is more than "predict the next word" — even though that remains their technical foundation. From comparatively few examples within a conversation, they can infer entirely new task patterns (in-context learning). They can lay out intermediate steps before settling on an answer, which makes more complex reasoning visible. They can call external tools and databases instead of drawing only on what they were trained on. They can process images, audio, and text together. In agentic setups, they can work through multi-step tasks with little supervision.

All of that looks intelligent in the everyday sense. And yet a great deal is still missing that we would take for granted in a person: a stable, internally consistent model of the world, goals of their own, a continuous life history, a body that accumulates experience. Language models hallucinate — they invent false statements with exactly the same linguistic confidence as true ones. They have no intrinsic motivation to want or avoid anything.

One word of caution, because it gets used uncritically in a lot of AI-progress commentary: so-called "emergent abilities" — sudden jumps in capability as models get larger — are often cited as evidence of a qualitative leap toward real intelligence. A widely cited study by Rylan Schaeffer and colleagues (NeurIPS 2023) found that much of this "emergence" disappears once you measure with continuous rather than discontinuous metrics. What looks like a breakthrough is partly a measurement artifact. An honest accounting has to include that.

Do Language Models Understand What They Say?

This question has occupied philosophy and computer science for decades — long before ChatGPT. In 1980, John Searle proposed his famous Chinese Room thought experiment: a person with no knowledge of Chinese sits in a room, following purely formal rules to assemble Chinese symbols into responses that look meaningful — indistinguishable from the outside from a fluent Chinese speaker. Searle's conclusion: pure symbol manipulation does not produce meaning. "Syntax by itself is neither constitutive of, nor sufficient for, semantic content," as he put it.

The counterarguments are just as well known. Maybe the person in the room doesn't understand, but the whole system — person, rulebook, and symbols together — does. Maybe all it would take is a body with real sensors to anchor symbols in the world — the symbol grounding problem, precisely named by cognitive scientist Stevan Harnad in 1990: how do abstract signs ever get connected to meaning in the real world, rather than just pointing to one another?

With modern language models, this debate is anything but academic history. Linguist Emily Bender, known for coining the term "stochastic parrot," takes a clear position: language models generate text by statistically predicting likely sequences of words — not by understanding what they are saying. Her central line: "When the text that comes out of one of these systems makes sense, it's because we are making sense of it." Meaning, on this view, arises in the human reader, not in the machine that writes.

Cognitive scientist Melanie Mitchell and physicist David Krakauer are more cautious. In their 2023 analysis of the debate, published in PNAS, they reach no clean verdict between the "emergentist" camp — which credits language models with real, if different, understanding — and the skeptics. Their proposal: instead of treating "understanding" as a yes-or-no question, we need a science that distinguishes different modes of understanding, each with its own strengths and limits. Even leading AI researchers disagree with each other on this. Yann LeCun has recently described language models mainly as powerful information-retrieval systems, and argues that real progress toward human-like intelligence requires fundamentally different architectures that model the world rather than continue text — a minority position that many of his colleagues reject.

What can be said with confidence: the question is open. Anyone who claims today that language models "understand nothing" is claiming more certainty than the research actually supports.

Maybe "Practical Wisdom" Is the More Interesting Category — Just Not in the Way It First Appears

Here, the history of the concept holds a surprise that turns this article's starting premise in an unexpected direction. In the Nicomachean Ethics, Aristotle draws a careful distinction between practical wisdom (phronesis) and a neighboring capacity he calls deinotes — usually translated as "cleverness" or "skill." Both describe the same underlying ability: finding the right means to an end. The difference lies entirely in the end itself. Someone who uses this cleverness for good purposes acts with practical wisdom. Someone who uses it for bad purposes is merely cunning. The capacity itself is purpose-neutral — it is the purpose that turns it into wisdom or its opposite. Practical wisdom in the full Aristotelian sense therefore presupposes virtuous character, experience, and good ends. That is not a more modest claim than intelligence. It is a more demanding one.

Kant saw it very differently, and far more soberly: for him, practical wisdom (Klugheit) was merely "skill in the choice of means to one's own greatest well-being" — a purely technical capacity, morally decoupled, a private matter.

This tips the article's original idea into something more interesting. "Artificial practical wisdom" is not automatically a gentler, more modest replacement for "artificial intelligence." In the full Aristotelian sense, wisdom is the more demanding attribution, because it presupposes a good character no language model has. The term that actually fits, from the same ancient family of concepts, is a different one: deinotes. Purpose-neutral cleverness. The ability to find the right means for almost any given end — without having an end of its own, without wanting anything good or bad itself. That describes what a language model actually does more precisely than either "intelligence" in the everyday sense or "wisdom" in the full Aristotelian one.

And What About Wisdom in the Fuller Sense?

For completeness, it is worth looking at the third category — and here, too, the search turns up something real. "Artificial wisdom" is not a term this article invents; it is a research field that has been active for over a decade. In 2012, philosopher David Casacuberta argued in the journal AI & Society that the field should think systematically about "artificial wisdom." In 2023, Ana Sinha and Pooja Lakhanpal, writing in the same journal, examined whether AI systems could become wise, and remained skeptical: wisdom requires weighing values, recognizing one's own uncertainty, and exercising culturally situated moral judgment — qualities that today's systems, which often conflate data, information, and knowledge, do not possess. As recently as 2026, a piece in Nature Mental Health explicitly called for a "strategic shift from artificial intelligence toward artificial wisdom" — without claiming the machine itself would become conscious, but as a direction for more responsible system design.

The honest answer, then: wisdom is not a category today's language models can claim for themselves. But it is not pure science fiction either — it is a question serious researchers are actively working on.

A Working Model, With One Caveat

From all this, a simple three-tier model emerges: AI as the ability to recognize patterns and solve problems. AC (artificial cleverness) as the ability to apply knowledge to a situation, purposefully. AW as the ability to weigh values, long-term consequences, and uncertainty while doing so. The model is tidy — but it needs the caveat from the previous section, or it falls into the same trap as the article's starting premise. The middle tier only holds if "cleverness" is understood as purpose-neutral skill, not as full, virtue-bound practical wisdom. Otherwise the term promises more than the system can deliver — exactly the problem it was meant to solve.

Why This Isn't Just Academic Hair-Splitting

The more interesting question was never "Does the machine really think?" That question cannot be settled with the tools we currently have — not by philosophers, and not by AI researchers either. The more practically relevant question is: does this system help me think better and make better decisions?

That shifts the debate from metaphysics to practice. And that is exactly where it becomes interesting for businesses.

What This Means for Companies

If a language model mainly delivers purpose-neutral cleverness — the ability to derive the right means for a specific situation from existing knowledge — a clear division of labor follows. The system can make knowledge accessible, explain complex information depending on the situation, make advice available at scale, and prepare decisions. What it cannot take over is practical wisdom in the full sense: the judgment of which goal is actually the right one to pursue. That responsibility stays with the people and companies who use the system.

For a company like leopard.ki, whose AI advisor leo.page does exactly this job — making technical knowledge accessible to website visitors in a way that fits their situation, without making the decision for them — that is not a footnote. It is the actual point: the future of AI may lie less in building artificial people, and more in making human knowledge available with artificial cleverness.