Can Machines Be Conscious and Intelligent?
The famous question is whether a machine can act intelligently. I think that’s the wrong question, and we’ve been answering it for so long that we’ve stopped noticing it’s wrong.
Turing asked it for a good reason. In 1950 you couldn’t get anywhere arguing about whether a machine had a mind, so he replaced the argument with a game. Put a person at one end of a teletype and a machine at the other, and if a smart interrogator can’t tell which is which, stop quibbling and call the machine intelligent. It was a brilliant move. It turned a metaphysical swamp into something you could actually test. And the test worked, in the sense that it gave the field a direction.
The trouble is that the test has now been passed, and passing it taught us almost nothing.
If you sit a modern model in front of an expert and let them talk, the expert will usually be fooled, or at least unable to tell. By Turing’s standard the matter is settled. But nobody who has watched these systems closely believes the matter is settled. We’ve built something that converses like a person and we still don’t think it’s a person. So the test measured the wrong thing. It measured whether a machine can imitate the output of a thinking person. It said nothing about whether there’s any thinking going on.
The Turing test is a test of deception, not of intelligence. Read what it actually rewards: the goal of the game is to fool the interrogator. That’s it. A machine that mimics well wins; a machine that thinks honestly but oddly loses. The test can’t tell the difference between understanding and a sufficiently good impression of understanding, and once impressions get good enough, that blindness stops being a clever simplification and becomes the whole problem. We built a yardstick that measures fluency and called it a yardstick for mind. Then we spent seventy years optimizing machines to score well on it. We should not be surprised that we ended up with superb mimics, because mimicry is exactly, precisely, the only thing the test ever asked for. A con man passes the Turing test on his marks every day. We don’t conclude he understands them. We conclude he’s good at appearances.
Let me say what I mean by intelligence, because the whole argument turns on it. Intelligence, in the sense I care about, is the ability to apply insight and understanding in context-sensitive ways — to explain things, to solve problems, to create. Notice that all three of those require a grip on what things mean, not just on what usually comes next. A weather vane responds to the wind perfectly and understands nothing. The question is whether our machines are weather vanes with very large dictionaries.
I’ve come to think they are, and the way to see it is to look at how they’re actually built.
Start with the part everyone calls learning. We say a model “learns” the way we say the sun “rises” — it’s a convenient description from the point of view of someone watching, not an account of what’s happening. What actually happens during training is that a very large function gets its billions of knobs adjusted until it’s good at predicting the next token. Nobody is understanding anything. A loss is going down.
People think of neural networks as some exotic third thing, but at bottom the training of an LLM is a kind of supervised learning wearing a clever disguise. In ordinary supervised learning a human tags each example: this is a cat, this is spam. The tagging is the expensive part, because the tagging is where the human judgment lives. The trick with language models is that text tags itself. Hide the next word and you’ve manufactured a labeled example for free, a trillion times over. It’s still learning-to-match-the-tag. It’s just that the tag was lying there in the data, and the human work moved upstream, into writing all that text in the first place.
Then look at reinforcement learning, which sounds like the most autonomous part of all. An agent, an environment, rewards, the whole vocabulary of a creature finding its way in the world. But the reward is a function somebody wrote. The reward is the tag, handed out by a procedure instead of a person, often late and sparsely, but still a tag. RL is supervised learning where the supervisor is a machine the humans built to stand in for their judgment. The agent isn’t deciding what’s good. It’s being told what’s good, in a language of scalars, and then climbing the hill we pointed it at. When it climbs the wrong hill — finds some exploit that scores high and means nothing — we call it reward hacking and act surprised. We shouldn’t be. It did exactly what we said. It had no idea what we meant, because meaning was never in the system. Meaning was in us.
Even the part that looks most like genuine discovery turns out to be borrowed. Unsupervised learning — the finding of structure in unlabeled data — is the closest thing to a machine figuring something out on its own. But what is it, really? It’s structure discovery performed with whatever methods happen to be the best ones we know right now. Clustering, dimensionality reduction, the modern self-supervised tricks. And who decides what counts as structure worth finding, and which method to reach for? We do. The toolkit isn’t fixed and it isn’t the machine’s. It’s a thing humans keep adding to, paper by paper, year by year, mostly by piling new methods on top of the old ones. K-means still works. We just have fancier things now. The machine applies the methods of the day. It doesn’t invent the next one, and it doesn’t lie awake dissatisfied with the ones it has.
So follow the thread all the way down and you keep arriving at a person. The data was written by people. The labels, where there are labels, were chosen by people. The reward was designed by people. The methods were invented by people. The prompt that sets the goal at runtime is written by a person. The evaluation of whether the answer was any good is done by a person. Take the people away and the behavior doesn’t degrade gracefully. It stops making sense.
Learning is mostly happening on the human side. We are the ones learning — learning how to formalize problems, how to write better objectives, how to wire these tools together so they produce something useful. The machine is the place that learning gets deposited. It’s an extraordinary place to deposit it. But it isn’t a second mind in the room. It’s a very good cast of the first one.
Searle had a story about this that’s forty years old and has aged better than almost anything written about AI since. Put a man who knows no Chinese in a room with a giant rulebook. Chinese characters come in through a slot; he looks them up, follows the rules, and pushes other characters back out. To everyone outside, the room speaks fluent Chinese. Inside, there is no Chinese, only a man shuffling symbols he can’t read. The point is brutal and simple: running the right program is not the same as understanding. Syntax doesn’t add up to semantics no matter how fast you shuffle. People have spent decades trying to wriggle out of this, usually by saying the whole room understands even if the man doesn’t. Maybe. But a language model is the most literal Chinese Room ever built. The rulebook is now a few hundred billion numbers, and the shuffling happens in a flash of matrix multiplication, and the outputs are good enough to win arguments. None of that puts a reader inside the room.
Which brings us to consciousness, and to a definition I’ll offer in the same spirit as the one for intelligence. People muddy this word on purpose, because if you define consciousness loosely enough a thermostat has it — it’s “aware” of the temperature. That’s a play on words. Here’s the version I think is worth defending: to be conscious is to be able to fully characterize your own context and to find the meaning in it — local meaning, global meaning, the sense of what this situation is and why it matters. A conscious thing isn’t told what’s salient. It works out what’s salient. It generates the relevance instead of receiving it.
By that standard our machines are not close, and it’s not a matter of scale. You can give a model a million tokens of context and it will attend to all of them and still need you to tell it, in the prompt or in the training or in the reward, what any of it is for. It tracks state beautifully. It does not own the state. The “aboutness” — the fact that a human thought is about something out in the world — never gets manufactured inside. We keep supplying it from outside and then marveling at how lifelike the result looks. It looks lifelike because we put the life in.
Now it would be foolish in 2026 to declare the future settled, especially with the stunning almost daily changes and developments we see wend their way into production, and with what we are able as individuals and collectives to accomplish using them. Maybe none of this is permanent. There are people working on systems that generate their own goals, that grow curious, that keep learning across a lifetime instead of a training run, that get grounded in a body or a world instead of in a pile of text. And there’s an honest case that the path forward isn’t a bigger model at all but a combination — the fluent model in the middle, with a truth-tracking memory beside it, a rules engine to keep it honest about hard constraints, something Bayesian to keep it honest about how sure it should be. Wire those together and you get something that characterizes context better than any one piece could. That’s a real and worthwhile thing to build. I’d bet on it. But notice what it gives you: better mimicry, more context-sensitive, more reliable, more useful. It doesn’t cross the line. It moves the line’s furniture around. The agency in the system is still ours, distributed across the parts we chose and arranged.
I think this matters beyond philosophy, and here’s why. When you believe the tool is intelligent, you stop holding people responsible for what it does. The ethics launder onto the machine, where they conveniently can’t be held to account. The model “decided.” The model “judged.” But the model didn’t decide or judge anything. It completed a pattern, triggered by an input, along a gradient we built. Every consequential choice in the chain was made by a human or by a proxy a human designed to act in their place. If we forget that, we lose the only place where accountability can actually live.
So I’d retire the old question. “Can a machine act intelligently?” — yes, obviously, we’re done, it was never the interesting question, and we’ve learned it wasn’t necessarily the right one to ask. The one worth asking is whether a machine can be intelligent and conscious: whether it can understand rather than imitate understanding, whether it can mean rather than be told what to mean, whether it can want to keep learning as a thing it does for itself rather than a thing we point it at. On that question the honest answer in 2026 is no, and the reason is not that our machines are too small. It’s that everything that makes them work — the data, the labels, the rewards, the methods, the prompts, the judgments of whether they did well — traces back to us.
We didn’t build a mind. We built a mirror, and we got so good at it that we keep mistaking our own reflection for company. The intelligence in the room is real. It’s just that, for now, it’s still ours.
But retiring a question puts you under some obligation to propose a better one, and “can the machine be conscious” is not actually the question that should keep us up at night. It’s an interesting question. It is not an urgent one, and it may not be answerable in our lifetimes. Meanwhile there’s a question sitting right in front of us that we can answer, by what we choose to do, and it’s the one I think we should be asking.
Here it is. Can we use these tools to become more context-sensitive than human beings have ever been?
That’s a strange thing to hope for from a machine that has no context-sensitivity of its own. But think about what’s actually on offer. The thing I do worst, the model does effortlessly: hold a million things in view at once, search everything, find the faint correlation between two facts that live in different fields and would never have met inside one human head. And the thing the model can’t do at all, I do without trying: know what any of it means, decide what matters, care about the answer. Put those two together and you don’t get a conscious machine. You get a human being with a reach that no human being has ever had.
That’s the part worth getting excited about. Not that the machine will understand context, but that we will understand more of it, faster, across more of the world than we could ever traverse alone. The doctor who can hold the whole literature in mind while looking at one particular patient. The teacher who can see each student’s context and meet it. The builder who can range across every adjacent field at once and bring back the thing that fits. For all of human history our judgment was bottlenecked by how much we could take in. That bottleneck is coming off. Our judgment is about to get a much larger world to be applied to.
So the question I’d put in place of the old one is this. Can we use these machines — and the truth-keepers and rule-checkers and everything we’ll wire alongside them — to develop our own context-sensitivity in ways no one has ever experienced, and to pair it with a superhuman ability to search and correlate and create, and then go do new things? Things that make the world better, and that genuinely excite us, in our work and in our lives?
That question has the great virtue of being ours to answer. The machine can’t be conscious for us, and it can’t be intelligent for us, and after all this it turns out it doesn’t have to be. It only has to make us more of both. Whether it does is not up to the machine. It never was.
On the Turing test, see Alan Turing’s original paper, “Computing Machinery and Intelligence,” _Mind_ 59 (1950): 433–460, which opens with “Can machines think?” and lays out the imitation game; for an overview, the Stanford Encyclopedia of Philosophy entry.
On the Chinese Room, see John Searle, “Minds, Brains, and Programs,” _Behavioral and Brain Sciences_ 3 (1980): 417–457, and the Stanford Encyclopedia of Philosophy entry.