AI Is Not Thinking. We Just Say It Is

Why the words we use to talk about AI matter

Photo by Thijs Kremers on Unsplash

If you are like me, you use AI almost every day. Whether it is for my research or as my default search engine, AI has become a part of my daily life. One of the stranger parts of using it now is that you can choose whether you want the “thinking” model or not.

But I’ll be honest with you. I do not know what the difference is.

I know a “thinking” model is supposed to be better at more complex questions and take longer to produce an answer than a “regular” model. In that sense, the label is useful. I know what button to click when I want help with something complicated. Yet the label still confuses me.

As a philosopher, the idea of calling what AI does ‘thinking’ naturally raises a lot of questions outside of practice. People ask whether AI is conscious or can think for itself. In fact, there has been an explosion in the discourse around AI in this regard.

The problem is that we may already have been primed to answer a certain way. Discussions around AI often use words like ‘thinking’, ‘hallucinating’, or ‘learning’. Thus, when we are asked if AI knows what it is doing or if it is conscious, then we often point back to these words as proof.

Here, I want to argue that we do not believe AI is thinking because we discovered or created a mind. We believe it because the use of thinking language around AI has trained us to see one there when in fact AI is not thinking — at least, not in the ordinary human sense of the word.

I believe what has happened instead is that a useful metaphor hardened into a literal meaning. In other words, what began by saying AI acts as if it thinks because it helped us understand a complicated technology turned into the fact that AI thinks, with the words “as if” simply being dropped.

This move, accidental or not, has deep-reaching implications and is behind many confused ideas regarding technology. Once we call AI a thinker, other ideas quickly follow. If AI thinks, maybe it understands. If it understands, maybe it has agency. And if it has agency, maybe it has consciousness.

The Metaphors That Made Computers Friendly

Computers have always relied heavily on metaphoric language. We have desktops, folders, clouds, and trash all because they help put highly technical concepts into terms all of us can quickly grasp. These metaphors worked because they made something unfamiliar feel familiar. They allowed ordinary people to use complicated machines without needing to understand code, circuits, or pathway allocation.

Take the ‘cloud,’ for example. Your photos are not sitting in the sky. They are stored in data centers, on servers, using cables, electricity, cooling systems, and physical infrastructure. The term ‘cloud,’ on the other hand, is easier to understand and easier to sell as a product to users. I may not understand the computer science behind how cloud storage makes sure I don’t lose important documents or pictures of my daughter. But I can understand how the ‘cloud’ allows me to retrieve pictures even if they have been deleted from my phone.

In fact, a lot of teaching and explaining revolves around the use of comparison to connect something not understood to something that is. As a teacher, I regularly compare learning a language to going to the gym. Building physical muscle is easier for students to wrap their heads around than something abstract like language ability. Comparing the two helps them see that what I am asking them to do is like telling them to go to the gym four days a week.

With AI, the metaphors are a special case because they borrow from mental life. “Thinking,” “learning,” “reasoning,” “understanding,” and “hallucinating” are not just interface metaphors. They are person-words in that these words are used only to describe the actions of people and nothing else — until recently.

When “As If” Disappears

The most important word in all of this is probably “thinking.” Other AI words matter, but thinking is the gateway. Once we accept it, much of the rest follows. Thinking is not just one mental activity among others. It is the word we use for the inner life that seems to stand behind all other mental activity such as reasoning, understanding, judgment, and intention.

This is why the word is so tempting. AI often does things that look like thinking from the outside. It answers difficult questions, summarizes complex documents, writes code, explains mistakes, and solves problems in steps. All things we humans have been doing since before computers.

OpenAI’s o1 announcement, for example, described a model that was trained “how to think productively” and said its performance improved with more “time spent thinking.” It also compared the process to “how a human may think for a long time before responding to a difficult question.”

This language was used in order to make the world of tokens, weights, inference, reinforcement learning, and probability easy to understand. Furthermore, no one wants to give a university computer science lecture every time they have to describe what a chatbot is doing. “Thinking” is shorter. It gives the user a practical sense of what kind of tool they are using.

But this shortcut has consequences.

At first, “thinking” may be understood as shorthand. The model is not literally thinking, but it is doing something like thinking that the comparison helps. This is mainly the case because the system was built to replicate the output of the human mind. The problem begins when the shorthand becomes the concept itself. The longer phrase “As if thinking” becomes simply “thinking.”

This is a kind of linguistic priming.

If I am told a system is calculating, I see calculation. If I am told it is generating, I see generation. If I am told it is thinking, I begin to look for thought. Because AI is designed to produce human-like language, the evidence is easy to find. The flow of electricity through circuits becomes a mental life in reality, not just metaphor. Yet, the system has not changed just the description. And once the description changes, so does what we think we are seeing. It is as if we are being led by the nose to the answer and forgetting that fact.

Of course, someone might object that this depends on how we define thinking. If thinking simply means solving problems, manipulating symbols, generating useful outputs, or moving through steps toward an answer, then perhaps AI does think by such a definition.

However, that is not usually what people mean when they ask whether AI can think. They are not asking whether a system can produce an answer. They are asking whether there is something mind-like behind the answer. They are asking whether the performance points to understanding, intention, judgment, or experience.

The Philosophical Consequence of Grammar

This tendency to codify linguistic shorthand or mistakes is not limited to technology. Wittgenstein famously warned that many philosophical problems arise when language “goes on holiday.” In other words, many confusions begin when words are taken away from their ordinary use and made to do strange work elsewhere. We then treat the confusion created by language as if it were a discovery about reality.

Let’s use AI as an example.

The word “thinking” has its ordinary home in human life. Human thinking is not just output. It is connected to bodies, experience, memory, desire, responsibility, fear, hope, confusion, and the strange feeling of being someone in the world. When I say that I am thinking, I do not mean only that words are being produced or my neurons are firing. I mean that I am trying to understand something from within a life that is mine.

Thinking ‘goes on holiday’ when I take the sentence “I am thinking” and, through the rules of grammar, change the subject of the sentence to something like “X is thinking” where X can be anything that grammatically fits.

There are obvious cases where such an artificial sentence clearly fails, such as “the chair is thinking,” in that I am not going to mistake that for a true sentence. Yet in borderline cases the question becomes harder to answer. Just take sentences such as “the dog is thinking” or now “AI is thinking”.

AI, unlike chairs, presents to us some of the outward signs of thinking that we see in humans. Sometimes it imitates them very well, and without internal access we feel that, at least metaphorically, it is ok to utter the sentence “AI is thinking.” At the very least, such a sentence is comprehensible.

The mistake is that we take this case of anthropomorphism, this case of a linguistic grammar mistake, and attribute to it a technological and philosophical. Again, the issue is not the metaphoric language. Humans anthropomorphize everything. We yell at our cars, name our Roombas, and accuse our phones of acting weird. The mistake is treating anthropomorphic language as evidence of reality.

Why the Words Matter

At this point, someone might say this is just semantics. Maybe AI does not think in the human sense, but who cares? If the tool works, why worry so much about the word?

The answer is that usefulness is not the same as truth. A metaphor can be useful, clarifying, and even necessary while being false if taken literally. Calling online storage “the cloud” helps people understand a technical system, but no one concludes that their family photos are floating above them in the sky. The metaphor works because we know where it stops.

The trouble with AI language is that the stopping point is much less obvious. When we say that a model is “thinking,” “learning,” or “understanding,” we are borrowing from ourselves. We are using words that normally belong to minded, embodied, responsible beings and applying them to systems that produce human-like language without living a human life.

That does not make the metaphor useless. In many ordinary contexts, it works well enough. If I ask whether a model “understood” my prompt, I usually mean whether it responded in a relevant way. If I choose the “thinking” model, I usually mean that I want a slower and more careful answer. In everyday use, these shortcuts are convenient. They help us navigate the tool.

But the fact that a phrase is useful does not mean it is innocent. The danger comes when we forget that we are using a shortcut at all. “AI thinks” begins as a convenient way of describing a pattern of behavior, but it can quietly become a claim about what the system is. The grammar stays the same, while the meaning hardens. What began as metaphor starts to look like discovery.

This is where we risk tripping over our own language.

We ask whether AI has understanding, agency, or consciousness, but part of the answer may have already been smuggled into the question by the words we chose. If we describe a system as thinking for long enough, then asking whether it really thinks starts to feel natural, even inevitable. But perhaps that feeling tells us less about the machine than it does about the language we have allowed to surround it.

So the point is not that we need to ban mental language from AI altogether. That would be unrealistic and probably unnecessary. Human beings explain unfamiliar things by comparison. We anthropomorphize, simplify, and speak loosely because that is part of how language works. The point is that we should not let loose language become literal without noticing.

AI can produce answers that look like the products of thought. It can respond in ways that resemble understanding. It can imitate reasoning, explanation, and judgment with increasing fluency. But resemblance is not identity. A system may behave as if it thinks without thinking in the ordinary human sense. Those two words, “as if,” are not a minor qualification. They are the line between metaphor and mistake.

This is why the real question may not be whether AI can think. At least, not at first. The question is what we are willing to say about AI, and what our words make it easier for us to believe. Are we describing a tool, or are we slowly talking ourselves into seeing a mind? Are we naming what is there, or are we being led by the metaphors we forgot were metaphors?

I will probably keep using AI almost every day. I will probably keep choosing the “thinking” model when I have a complicated question. The label is useful, and I know what it means in practice. But I also want to remember that clicking on a word is not the same as discovering a mind.

Maybe that is where we should begin: not by asking whether AI is thinking, but by asking whether we are being careful enough with the words we use.