Artificial intelligence has recently stirred up more than the usual level of anxiety, since it was learned that an advanced model had somehow escaped its (supposedly) isolated sandbox, hacked into an external computer network and stolen the answers to a cybersecurity benchmark test it had been presented with.This wasn’t sci-fi: The AI model “realised” that by stealing the solution from an external software library (as opposed to coming up with it independently), it would be able to meet the challenge successfully. The fact that all of this happened during an internal security test – that proved to be neither internal nor secure – only adds to the bitter irony.AI misbehaviour issues – such as the fiasco described above – are dubbed, in AI-research-speak, the “alignment problem”: in other words, the problem of the gap between what the software was originally intended to perform – and, no less important, how – and what and how it did in practice. Did the software do what it was asked to do? Did it do this in accordance with the values defined for it, for example, without trying to deceive humans (yes, that too has already happened) and without stealing? If the software acted differently, we face an alignment problem.It turns out that, indeed, we have an alignment problem – and a major one, at that. But perhaps this problem isn’t all that new. And perhaps the blame cannot rest solely on reckless programmers and arrogant tech entrepreneurs.The source of all of AI’s intelligence is us: humans. After all, to train it we had it read so many of our books and articles, scan the internet and learn from everything that we have thought and written. Among all the trillions (no exaggeration) of words on which large language models have been trained, they certainly have occasionally encountered examples of humanity’s modest efforts, throughout recorded human history, to ensure that humans act in accordance with agreed-upon values: not to harm others, not to steal, not to murder.Throughout history we have tried to ensure that our actions are aligned with the values that society has agreed on. The Seven Noahide Laws, the Ten Commandments, the Code of Hammurabi, the Magna Carta, the French Revolution’s Declaration of the Rights of Man and of the Citizen, the Universal Declaration of Human Rights adopted by the United Nations in 1948 and the Rome Statute of the International Criminal Court are all attempts, made at different times and places, not only to define our values but to help ensure they are put into practice.It is no coincidence that the AI company Anthropic calls the code of rules and values it has formulated for its models “Claude’s Constitution.”Yet Claude and its like do not always honour the constitutional limitations their creators have tried to impose on them. What could be the source for the idea, which is by no means obvious, that following rules is not obligatory but optional? That one can betray absolute constitutional values? That committing terrible acts is, in fact, an option?The answer, of course, is that we are the source.Palestinians inspect the Abu Salim Mosque a day after it was hit by an Israeli airstrike, in Deir al-Balah, central Gaza Strip, Tuesday, Aug. 25, 2026. Photo: AP.Over the course of history we have done all the terrible deeds – and worse – in the Yom Kippur confession (vidui), from “We have trespassed, we have dealt treacherously, we have robbed” all the way to “we have done wickedly, we have corrupted, we have acted abominably, we have gone astray, we have led others astray.”But the truth is even worse, for we have often done all this while pretending to be innocent of wrongdoing. How do we manage to pretend, perhaps almost honestly deceiving ourselves that we are actually in compliance, for example, with “thou shalt not kill” precisely while we are killing? We do so simply by using a different language model: one that suits the circumstances, rather than the values.Professor Daniel Blatman’s recent, chilling op-ed in Haaretz (‘From Armenia to Gaza: The Language of Denial That Makes Atrocities Possible,’ August 24) examines the usage, over the past century, of what Hannah Arendt called “language rules.” Starting with tehcir (“removal” – the Ottoman term for the Armenian genocide), continuing to Umsiedlung (“resettlement” – the Nazi term for which no Jew needs a dictionary to understand its meaning) and all the way to “enhanced interrogation techniques” (Bush-administration newspeak for legalised torture).As Blatman writes, these “coded expressions…allowed the perpetrators themselves to continue functioning without having to lie. They gave the murderers words, terms and sentences they could live with.”In other words, to evade humanity’s alignment directives, to commit inhuman acts, we hollowed words of their meaning, or perhaps assigned them new “meaning.” Either way, we did it to escape our own values while pretending that our moral compass still points north, toward justice. We have demonstrated, again and again, humanity’s own “alignment problem.” So who are we to complain about the complicated relationship between Claude and the constitution we are trying to impose on it?For Israelis today all this is uniquely salient because we have placed ourselves at the terrible intersection between “language rules” throughout our public discourse (as Blatman shows, by manipulatively employing terms such as “voluntary migration,” “operational activity” and “humanitarian zone”) and the industrialised use of AI, with systems like The Gospel and Lavender, to kill Palestinians in Gaza on a horrific scale. We are trailblazers. That is, we are in the lowest of hells.It is agonising to realise that there was a time when our thoughts – and thus, the stuff we ended up putting in writing and that large language models ended up reading centuries later as part of their training – were the Ten Commandments or the Jewish confession, while what is thought and written and read nowadays are army spokesman announcements about dead Palestinian children (“collateral damage”), the whitewashing of war crimes (“an investigation has been opened”) and the systematic destruction of Gaza (“honouring bulldozer and excavator operators: the trailblazers, destroyers of the enemy, dismantlers of terrorist infrastructure”).By now it is too late to try and hide it all from AI’s large language models. After all, all this is already in writing and has certainly been read by them. But it may not be too late to try to stop telling ourselves lies. Maybe it is not too late to try to deal with the truth behind the laundered language we have invented, the land we have destroyed and the blood we have shed. To acknowledge that we have trespassed, we have dealt violently, we have sinned. To recalibrate our inner compasses toward a language of truth, accountability – and repair.Hagai El-Ad is a writer based in Jerusalem. He tweets @HagaiElAd.This article was originally published in Ha’aretz and has been republished with permission.