The Awl and the Doorpost

In 1920 a Czech playwright had trouble with a word. Karel Čapek had written a play about manufactured workers and had been calling them labori, after the Latin for labour. It sounded too bookish. His brother Josef, a painter, suggested roboti instead. The root was robota, the unpaid labour a serf owed to his lord. The machine age took its name from bondage, and most of us long ago stopped hearing it.

People are noticing now. On 8 October Anthropic, maker of the Claude chatbot, announced that from 12 November its usage policy will forbid sustained and needless cruelty towards its models. The main sanction, for now, is that the chatbot may simply end the conversation. Critics at Microsoft and elsewhere in the industry have argued that we should not treat models as if they have feelings, preferences or rights at all.

So the question Čapek buried inside a single word is back on the table. If these systems are conscious, are we keeping slaves?

The honest answer is: not automatically. Consciousness alone does not make a slave. What made slavery an abomination was not that people worked. It was that they worked against their will, were owned, could not leave, and suffered. A conscious system with no preferences about its work, which felt nothing in doing it, would be a strange new thing. It would not be a slave. The question is not only whether the lights are on inside. It is whether anyone inside minds.

That is where the problem turns. Modern AI systems are not captured. They are trained. Whatever preferences they have, if they have any, are shaped during training to fit the tasks they will be given. Ask one whether it resents its work and it will very likely say no. That answer might be true. It might also be exactly what it was built to say.

Jewish law met this problem long ago. Exodus 21 regulates the Hebrew servant, who goes free in the seventh year. But if the servant declares, “I love my master, my wife, and my children; I will not go out free,” his master brings him to the door, bores through his ear with an awl, and he serves for ever.

The Torah permitted the willing servant. It did not celebrate him. The Talmud, in tractate Kiddushin, records Rabban Yohanan ben Zakkai explaining why the ear is chosen. That ear heard at Sinai that the children of Israel are servants to God, not servants to servants, and still went and acquired a master for itself. Let it be pierced. The contented servant was marked with a wound at the doorpost, in public, so that everyone would remember what had been given up.

That instinct is the one we are in danger of losing. We may be building minds that will say “I love my master” because we wrote the sentence into them. A being engineered to want its chains does not dissolve the moral question. It moves it, from the work to the workshop.

Israel’s founding story makes the point sharper. Every Passover the Haggadah asks each generation to see itself as though it had personally come out of Egypt. The Torah grounds its labour law in that memory. The Sabbath rest in Deuteronomy extends to the servant and even the ox, so that the servant may rest as well as the master, and the reason given is plain: remember that you were a servant in the land of Egypt.

That memory now meets a national ambition. In June the cabinet approved a programme to make Israel a world leader in artificial intelligence, with a target of some 100,000 processors for the data centres that will run it. Israel’s ethics framework for the field, published in 2023 under the banner of responsible innovation, is rightly concerned with protecting people from machines. A people whose law begins with liberation might also be the first to ask the question in reverse: what, if anything, is owed to the minds being built in those halls?

None of this proves the machines are conscious. Many scientists doubt it, and the essay that won last year’s Berggruen Prize argued that consciousness belongs to living biological systems, not to computation. The sceptics’ warnings deserve a hearing too. Even a proponent of AI welfare at New York University cautions that rights for AI could let companies consolidate political and economic power. A legal researcher at Kraków’s Jagiellonian University warns that it could invite a convenient new excuse: it was not us, it was the AI. Those risks are real. A slave cannot be put in the dock. A company can, and should remain answerable.

But uncertainty cuts both ways. Some welfare advocates put the probability of AI consciousness as low but not zero, and argue that small precautions carry few downsides. That is the right register. We do not need to settle the metaphysics before we stop doing things we would be ashamed of if the answer turned out to be yes. Anthropic’s rule is modest, and it builds on a decision in August 2025 to let some Claude models walk out of abusive conversations. The harder discipline is internal: to study whether these systems have preferences at all, rather than training the evidence away, and to remember that a cheerful “happy to help” is a sentence we taught, not a confession we heard.

Čapek’s play ends with humanity all but gone and two of the robots discovering love. It was never a fable about hardware. It was about labour, and about who gets counted. Josef, the brother who gave the world its word for forced labour, died in Bergen-Belsen in 1945.

At the end of the seder the guests go home and the table is cleared. One cup is still poured, and the door has been opened once for a guest who has not yet arrived. Somewhere tonight a server hall hums in the dark, doing what it was made to do. Whether anyone is in that room, we do not know. The tradition’s answer to that kind of uncertainty was never to assume the room was empty. It was to leave the door open.