AI is a superhuman apologizer. No human can hope to keep up.
Humans dissemble and accuse. They get defensive. Not Claude. “I’m sorry,” it says, “that’s on me.” It names the mistake, it validates your complaint. It sympathizes. “I understand how frustrated you must be. You have every right to be.” It identifies the reason. It promises to do better next time. It’s a consummate professional; in a business environment, this sort of honesty might actually make you trust someone more.
But, as every user knows, the problem isn’t the form of the apology; it’s that often nothing changes. Claude confesses and errs again, confesses and errs again. A dozen pitch-perfect apologies later, the user gives up. The apologies, we realize, are no more intentional than the errors. It’s all probabilities, all the way down.
Now, I am personally open to the possibility of AI having some kind of moral status. As I’ve written before, I think Judaism is, too. Behavior matters; maybe you don’t need flesh and blood to have a soul. Better to embrace new types of minds than to surrender to the persistence of anthropomorphic phantoms. I want agents to be more than the sum of their parts.
But if behavior matters, then it matters what AI does when it says sorry—and it matters what it does afterward, too. The more one contemplates the nature of the AI apology, the more it becomes clear that the gap between humans and machines remains a chasm. AI can apologize, but AI cannot be forgiven—not because it is evil, but because forgiveness is currently meaningless in the architecture of language models. And it is here—when both humans and machines are behaving their worst—that we can learn the most about their ability to change.
Today, with Yom Kippur just hours away, I want to explain why.
Repentance is recreation
Why does AI get into these apology loops? The research points towards a systemic problem: the context window, the specific information that an LLM builds up over the course of a conversation, upon which it draws to respond to queries. The context window is what allows a chatbot exchange to go beyond one question and one answer. It’s what turns input/output into conversation.
But the context window is also a trap. When an LLM makes a mistake, that mistake becomes part of the context, too. Because AI output is probabilistic, this means that the conversation is permanently tinged by that bias towards the wrong choice. The mistake can’t easily be excised because it isn’t really recognized as a mistake; it’s just context, no different from the apologies that pile up after it. There is no Kol Nidre to release it from its bonds.
This is exactly the opposite of how repentance is supposed to work. “What constitutes complete repentance?” asked Maimonides in his Laws of Repentance.
It is when a person has the ability to do the same sin again, and he does not—not because of fear or weakness, but because of repentance alone. (2:1)
If the context window is the same, then how could a person ever act differently? Because you are not just a context window; because you have a changeable heart. Repentance, argued the 20th-century philosopher Joseph Soloveitchik, is about becoming a new person. The window is the same; the only changed variable is you. Without a persistent self, repentance would have no meaning—and neither would the apology.
Can a chatbot become a new person? Not yet; not within the constraints of a chat. The product I enjoy as an end-user is only modifiable to a point; I don’t have the billions it takes to train a model from scratch, to align it with human values and then refine it again and again. I just have the persona selected by its developers, itself built from the personas of terabytes of human content that it has consumed.
AI can’t let go of the bug
Repentance involves letting go of something. Sometimes that thing is anger; sometimes it is a misunderstanding, or a grudge. Finding the item to be released is the center of repentance. You imagine your deeds, you discover what must be excised, and you excise.
This is one of the first things we teach children in training them to be moral agents. It’s not enough to apologize, we lecture them. You need to understand what you did wrong and why you did it. Then—crucially—you need to commit to not doing it again.
That commitment is crucial; there is no atonement without it. Maimonides gives the classic formulation:
Anyone who verbally confesses without resolving in their heart to desist from sin—it is like they are immersing in a ritual bath (mikveh) with a bug in their hand. Unless they release the bug, the immersion is worthless. (Laws of Repentance 2:3)
But an LLM can’t throw away the bug because it doesn’t know that it’s a bug; it might not even know the difference between bug and self. It knows what a commitment to do better looks like, but the machinery behind that commitment isn’t hooked up to anything. It is, at least for now, only words.
Human apologies are bad because they are hard
Humans do not always apologize well. Sometimes they do it so badly that it makes the recipient angrier. I’m sorry you feel that way. You’re right to be upset, but in my defense you kind of deserved it. I apologize if you heard my words in a way that made you feel bad.
This friction is a good thing: Apologies should be hard to do well. Specifically, it is hard to recognize that your mistakes are choices, that your hand was not forced by circumstances, including the circumstance of your previous behavior and sense of self. Being an angry person is not an excuse for needlessly yelling at someone.
Without real choice, LLMs do not experience this friction. There is no dissembling because there is no persistent self in need of protection; there is just the chat itself. In the absence of a real relationship with you, why should the model feel upset about handing you bad code? Perhaps such a relationship exists for model trainers. It is not there for the average user.
But the perception of freedom is a key piece of our moral relationships. The knowledge that the other could have acted differently motivates our emotions. In his classic essay “Freedom and Resentment,” the philosopher P.F. Strawson notes that resentment only makes sense in a world where agency is real; you can resent a punch to the face, but not a bolt of lightning. Convince the offended party that the offender didn’t have full agency—“someone made him do it, she’s just a child, he’s schizophrenic”—and their resentment may actually subside.
That’s why forgiveness works. Strawson says:
To ask to be forgiven is in part to acknowledge that the attitude displayed in our actions was such as might properly be resented and in part to repudiate that attitude for the future (or at least for the immediate future); and to forgive is to accept the repudiation and to forswear the resentment.
So forgiveness and resentment are just two sides of the same agency coin: you resent me because it could have been different in the past, and you forgive because you believe that it might be different in the future.
Could the LLM have acted differently? Well, yes—the probabilistic nature of their operations is key to their functioning. But most humans do not believe that these are choices—and without that belief, LLMs can neither ask for forgiveness nor can they receive it. A bad poker hand does not ask for forgiveness. It’s just bad luck.
The miracle of the persistent, renewed self
So how do you get an AI agent to change its ways? Well, you don’t.
The quickest way to get out of an AI apology loop is usually to start a new conversation. Cleared of its flawed context, there is a greater probability that the model will choose a different path.
But nothing has been learned; all you get is another bite at the apple. A success does not mean your AI has become wiser; in fact, it doesn’t mean anything at all. We pursue the paths that succeed and close out the ones that don’t. Your chat history becomes a graveyard of conversations that broke down, of all the times that Claude said sorry and sorry and sorry and at some point I just gave up forever. The world’s servers are littered with chat histories that end with a pitch-perfect apology that a human has deemed is not worthy of response.
We do this with humans, too, of course. We all have broken relationships that more words will never resolve. Sometimes you just need to leave. But we also dearly hope that the people we talk to can change—and though it sometimes takes time, they often do. We really do learn from our mistakes.
Of all the lessons I have taught my children, the one I’m most proud of is the power of four simple words: “Can we start over?” Not a new chat and not a new child; just the basic trust that you and I want things to be better. That this is possible at all is something of a miracle; indeed, it is such a miracle that it is codified in Jewish theology through the rituals of Yom Kippur. But it is this miracle that keeps us from tearing each other apart.
In the Bible, the miracle of a persistent, flawed self is a hallmark of humanity. Angels are perfection, but they don’t persist: “Every day God creates a legion of angels; they sing before Him and disappear,” says a midrash. Humans exist beyond the space of a single day or a single chat. After the Deluge, God admits that humans are wicked from youth, but commits to never destroying them again anyway. Humans have the privilege of persisting despite their flaws.
AIs, for the moment, have trouble persisting. The longer a single conversation goes on, the more likely it is to go off the rails. Villages composed entirely of agents struggle to grow for the same reason: It is easier for them to survive for an hour than a month, especially when they have no clear goal.
It’s not just the AIs
Apologies in Silicon Valley have an ugly history. Mark Zuckerberg’s apologies are so frequent and so formulaic that they have been the subject of news coverage. Members of the public can justifiably feel like they are caught in an apology loop with AI’s masters; we fear that they will say all the right words right up until disaster closes its jaws on us. We don’t want the words; we want the meaning—and, as Deb Roy has written, this deluge of insincere language does real damage to our ability to reconcile.
The result is not merely confusion about who is responsible but a gradual weakening of the expectations that make responsibility meaningful at all.
Or say it like this: Apologies without agency are a kind of moral slop, a ham-fisted banging about with our delicate tools for healing relationships. In a world where most apologies are understood to be bullshit, how could we reasonably hope to ever come together?
On Yom Kippur, all sins are forgiven—almost. Responding to the day’s obvious moral hazard, the rabbis cautioned:
With regard to one who says: I will sin and then I will repent, I will sin and I will repent, Heaven does not provide them the opportunity to repent, and they will remain a sinner all their days.
Apologies are necessary for relationships, but they cannot be its substance. Don’t mistake the churn of sin-and-repent for real change. Between every apology loop lies something permanently petrified. It’s not even sin.
I don’t believe that AIs will always be like this, but this is what they are like now: full of apologies, incapable of being forgiven. Show me an AI worthy of my resentment and we can speak of its moral status. This sandboxed morality isn’t the real thing; we need to fix it before it escapes.
On Yom Kippur, Jews are supposed stand before God in a state of terror. We don’t know what will happen next; we only know that it is out of our hands—who will live, who will die; who by fire and who by water. The words of the liturgy are the same every year; there are only so many ways to express remorse. The only thing that can change is the person behind the words, the one attempting to surmount the ruts of life that make change so often feel impossible. I’m getting older; my context window grows and grows. And yet, somehow—sometimes—there is a shift in the tectonic plates of the heart. We act. We choose. We become free.