Oded Balaban

On a conversation of Yuval Noah Harari about artificial intelligence

Oded Balaban — University of Haifa — balaban@research.haifa.ac.il

On Yuval Noah Harari, “The next 50 years: humanity, AI, power”, in conversation with Tom Parker, journalist at the Financial Times, at the fiftieth anniversary dinner of the EDP group, Portugal, July 2026.

https://www.youtube.com/watch?v=_V_ed5fuexA

In July 2026, at the fiftieth anniversary dinner of a Portuguese electricity company, Yuval Noah Harari spoke with Tom Parker, a journalist at the Financial Times. The video is linked above, and anyone who can watch it would do well, because it is instructive in a way its author did not intend. There is no need to refute it from outside. It refutes itself three times over, and the third time is admirable.

Let me say at the outset what I am not going to dispute. Several of his political conclusions strike me as sound, and one or two I would sign word for word. My interest lies elsewhere: in what he maintains without deriving it from anything, what follows from no premise of his and what he does not defend either, because he does not see it as something in need of defence.

A criterion that lasts one sentence

Parker asks him whether artificial intelligence is a tool or an agent, and whether we might be choosing the wrong narrative by calling it a tool. The answer is blunt: “that's the biggest mistake”.

And then he gives the criterion, and the criterion is the right one. He says that a tool is something that cannot make decisions by itself. The printing press is a tool: it could not decide to print the Bible today, the Quran tomorrow and Confucius next week; Gutenberg decided, and the printing press was the means of carrying out Gutenberg's decision. The atom bomb is a tool: in 1945 it could not decide by itself to bomb Hiroshima or Tokyo. And the sentence in full:

You needed a human to decide, let's bomb Hiroshima.

Exactly. Nothing to object to. The distinction is well drawn and the criterion is the one to use.

A moment later: “an AI weapon can decide what to bomb”.

Read it twice. In the first half of the argument, deciding is what someone does, and that is why the printing press is not an agent. In the second half, deciding is what the weapon does. Between the one and the other nothing was argued: the subject was changed. The criterion just laid down —that there is someone who decides—⁠ was not applied to the next case, and not because the case resists it, but because applying it would have produced the opposite answer.

I call this operation putting the subject inside the verb, and it is no grammatical nicety: it is what settles who answers afterwards. While the end is assigned, there is someone who assigned it, and that someone has a name, holds a post and could have chosen another end. Once the weapon has it, the sentence no longer obliges anyone to name him. What disappears is not the responsible party: what disappears is the need to look for him.

The invoice arrives before the ontology

What that attribution rests on he states himself, unprompted:

If it's not an agent, then it's not an AI and all the hundreds of billions of dollars that are being poured into its development will be wasted.

Note the order of the reasons, because it is inverted. It is not concluded that we must invest because the thing is an agent. It is concluded that the thing is an agent because we have already invested. The premise leans on the investment the premise was supposed to justify.

No philosopher would want to own that argument, and yet it governs a good part of the discussion. It might be put this way: given that we have spent a fortune building a subject, there is a subject. The ontology arrives late, and it arrives to collect.

Later he will say that intelligence is becoming abundant and cheap. If it really is getting cheaper, the argument from expenditure loses its object.

The mirror from inside

What precedes is a slip that repeats itself. What follows is something else, and it is the most valuable thing in the hour.

Parker asks him what becomes of his own trade —he works with words—⁠ if the machine comes to handle them better. Harari answers that the underlying question is whether the machine thinks, and that this forces us to ask what thinking is. He offers two options. The first: thinking is putting words in a particular order, as in a syllogism. If that is thinking, he says, the machine does it better than we do, because we can order twenty words and it can order twenty thousand.

And then he turns inward:

When I observe what's happening inside my own mind, what I find there is some kind of LLM that just predicts the next word in the sentence I'm saying. Like this sentence that I just now said to you: when I started saying it, I did not know how it will end. Why did I say the sentence I'm saying, not some other words? I don't know.

And he concludes: “so saying that AI is just predicting the next word — okay, how is it different from us?”.

Let the reader pause, because something has just happened. He did not say that the machine resembles us. He said that he resembles the machine. He looked inside himself and found a language model. And he did not say it as a complaint or as a warning: he said it as an argument, in the tone of someone who has just been candid with himself.

This is the whole matter of the discussion about artificial intelligence, and for that I am grateful to him: nobody had displayed it so clearly or so unintentionally. The resemblance between man and machine was not discovered. It was manufactured, and it was manufactured on this side.

The question he asks is a good one, and the answer is not difficult. Not knowing in advance how the sentence ends is not the same as not being the one who says it. The one who says it is the one who can be surprised by it, correct it halfway through, take it back, fall silent, or hold to it against whoever objects. A model is not surprised by what it produces, because there was nobody there expecting anything else. Surprise is not one more fact about the output: it is the proof that someone was there with an expectation, and expectations are not had by the text.

And look at the I that reports. He is the one observing, the one who does not know, the one astonished at not knowing. The testimony requires in full the subject the testimony denies. It is the old circle: if my brain is the one dictating to my fingers the words I am writing, and I innocently believe myself the author, then this thought about my dependence on its dictation also comes from it, and from it the reflection on that reflection, and so on indefinitely. The operation does not eliminate the subject: it changes its owner.

There is something more, and it is what reveals the method. When the formal criterion fails him, Harari does not move toward the act: he moves toward affect. He offers a second idea of thinking —thinking would come “from the stomach, not from the brain”, it would be a matter of the feelings surrounding the words—⁠ and from there he passes to the question of whether the machine feels. It is the same swerve he makes elsewhere, when he defines consciousness as the capacity to suffer.

Note what that swerve leaves out. To suffer is to undergo. Left outside are judging, deciding, setting oneself an end and asking, which are precisely the acts in dispute. Once it has been settled that thinking is ordering words, the only thing left to distinguish us is undergoing. And then, indeed, the question “could the machine feel?” becomes the only one available, and there is no longer any way of not asking it.

The button

Speaking of empires, he says something that dismantles his thesis, and he does not notice.

He explains that artificial intelligence allows a concentration of power no earlier empire could achieve. Rome could not move the wheat fields of Egypt to Italy; Britain could not move the oil fields of Iraq to Yorkshire. With AI everything can indeed be concentrated in one or two countries. And he adds the second difference:

When a Roman merchant sold a steel sword to some Gothic tribe, the Romans lost control of the sword. There wasn't a button in Rome that the emperor presses, and all the steel swords stop working. That's the case with AI.

He gives the example: the Ukrainian drones that stopped working when someone pressed a button. And he generalises it: in certain architectures “there is always a kill switch in the imperial hub”.

It is an excellent observation, and it is exactly the one to set against his own thesis. Where there is a button, there is a hand. A system that can be switched off from the imperial hub is not an agent seizing power: it is a tool whose owner is far away and out of sight. Harari says it and moves on, because in his story the danger has to be the machine, and here the danger has just acquired an address.

Buenos Aires, in passing

There is a moment that will interest Argentine readers. Speaking of legal personhood, Harari states that “just a month ago” the Argentine government announced that it is going to grant legal personhood to artificial intelligences, so that they can operate corporations without humans.

What can be asserted with certainty is less than that and, to my mind, worse. On 4 June 2026 Javier Milei and Federico Sturzenegger published a column in the Financial Times proposing a reform that would provide for companies constituted and administered entirely by artificial intelligence agents. Whether the text goes so far as to make an AI a subject of law is disputed among Argentine jurists, and some hold that current legislation prevents it, because shareholders and administrators must be persons.

But the interesting point does not need the strong version. The amnesty requires no metaphysics. For there to be nobody left to answer to, there is no need to declare a machine a person: a corporate form in which no human has to appear in the chain of decisions will do. Legal personhood for the robot gives nobody a voice: it gives the owner a mask. And a company without humans does the same thing, more cheaply and with no philosophical argument at all.

The genie and the lamp

And now the best of it, which is none of his theses but the distance between the beginning of the conversation and its end.

He defined an agent as that which “can make decisions by itself and invent new ideas by itself”, and maintained that the machine is one. Toward the end he answers what he would like historians to say about this period. He hopes they will remember it as the wisdom revolution, and he explains it in economic terms: when something becomes abundant and cheap, the question is what the next bottleneck is. Intelligence is becoming abundant and cheap. What, then, is the bottleneck?

The bottleneck becomes the question: okay, what do I want to solve? I can have anything I want — what do I want? […] This is the difference between intelligence and wisdom. Intelligence is the genie. The genie will do anything, all your dreams will come true. But you need wisdom to choose the dreams.

It is an excellent distinction, and he formulates it better than many who defend it. Only he does not notice what it does to his own thesis.

If choosing what to solve is a faculty distinct from solving; if intelligence is the genie that grants and wanting is something else; and if the machine is the one that solves, then the machine does not choose. Its ends reach it from outside, exactly as they reach the genie of the lamp, who is all-powerful and has not one wish of his own. The genie does not command: whoever rubs the lamp commands. And so the agent of the beginning turned out, in the end, to be the lamp —the instrument in the master’s hand.

And there is a further irony in the choice of image. In the tales, the disaster is never caused by the genie: it is caused by whoever asks badly. The genie is a perfect means and for that very reason innocent, and the whole of the responsibility lies with the one who formulates the wish. To explain artificial intelligence, Harari chose the figure that best proves the responsible party is the human being.

What is left standing

None of the above touches his political warnings, and I want to say plainly that I agree with several of them. That a handful of companies and two states should concentrate the infrastructure everything depends on is grave, and it is now. That legal personhood should be granted to artificial systems is an anticipated amnesty, and that is now as well. That the public sphere has been laid waste by systems optimising a metric somebody chose is an accomplished fact. That intimacy is being manufactured at industrial scale to be sold to ten-year-olds is among the most serious things he says, and he says it well.

But look at what he had to do in order to say all that: suspend his own premises. Those pages work only if the machine is an instrument and there are people behind it who decide. If it really does decide by itself and has ends of its own, then it is an agent, and demanding a human answerable at the end loses its ground, because an agent is called to account itself.

His best paragraphs contradict his presuppositions. And that, far from being an objection, is the best news of the conversation: when he thinks politically rather than metaphysically, he thinks well. The trouble is the frame he reaches those thoughts from, because that frame is what produces the fear, and the fear is what afterwards prevents us from looking where we should.

For the fear of replacement says less about the machine than about the one who feels it. A better calculation can only replace someone who has placed his being in what is calculable. If the human being were above all someone who asks and judges, a faster machine would be a better tool and not a replacement. To fear being relieved of one's post is already to have reduced oneself to the layer the relief reaches.

None of this requires the machine to wake up. And all of it is seen better once one stops waiting for it to wake up.

The quotations refer to the video linked above. Translations from the Spanish are my own; the quotations from Harari are given in his own words, checked against the subtitles. The column by Javier Milei and Federico Sturzenegger appeared in the Financial Times on 4 June 2026. I develop these arguments at greater length in a book in preparation on artificial intelligence from the user's point of view.