Apologies are usually treated as morally mandatory whenever one person has wronged another. I do not think this survives even elementary economic analysis.
Suppose I behave badly toward someone. They become angry, trust me less, stop inviting me places, tell other people what I did, become less willing to lend me money, stop introducing me to people, or end the relationship entirely. Those are not incidental consequences sitting alongside some separate moral penalty. They are the penalty. Reputation is one of the main mechanisms by which interpersonal harms get priced.
If I insult a friend at dinner and call him a talentless parasite in front of eight people, I should expect him to like me less afterward. If I borrow $600 and take nine months to pay it back, I should expect him to become less willing to lend me money. If I get blackout drunk at his birthday, disappear for three hours, come back bleeding, accuse his girlfriend of flirting with me, and knock over the cake, I should expect some reduction in future invitations. If I tell someone that everyone secretly finds them exhausting and that their marriage is obviously going to fail, there is a decent chance they will update negatively on me.
These consequences seem to me like the social system functioning normally.
Yet after the reputational adjustment has already occurred, people frequently insist that I also owe an apology. This is where I lose the thread. If someone now trusts me less, likes me less, gives me fewer opportunities, tells other people I am unreliable, stops covering my bar tab, no longer fronts me poker money, and declines to invite me on the ski trip, then I have already incurred a real cost. Demanding an apology on top of that looks very much like charging a fine after the market has already repriced the asset.
The usual response is that an apology performs a different function because it communicates remorse. But remorse is private information, and “I’m sorry” is an extraordinarily cheap signal. A person who genuinely regrets what he did and a person who merely wants his access restored can produce nearly identical sentences. This is why corporate apologies, celebrity apologies, relationship apologies and the things people say after being confronted about drinking too much all converge on the same narrow verbal repertoire. “I understand that I hurt you.” “I take full responsibility.” “I need to do better.” None of these statements is difficult to produce.
Behavior over time is much more informative. Suppose I spend three months drinking heavily, borrowing money from everyone, missing plans, saying vicious things while drunk, losing rent money playing online blackjack, and waking people up at 4 a.m. because I have developed some urgent theory about why they have secretly betrayed me. Then suppose I stop doing those things. That tells you something. Conversely, if I deliver a beautifully calibrated apology on Monday and by Saturday I am asking the same friend to spot me $200 because I lost my paycheck gambling again, the apology was not especially useful evidence.
Ordinary social norms seem to reverse those weights. They attach enormous significance to whether the correct sentence was produced and often treat actual behavioral change as secondary. This seems irrational if the purpose is prediction. If you are trying to estimate whether I will behave badly again, my subsequent conduct is a much better signal than my willingness to verbally endorse your interpretation of what already happened.
There is also a moral-hazard problem. If reputational damage can be cheaply reversed through apology, then the expected cost of bad behavior falls. Someone who believes that screaming horrible things at his friends, getting thrown out of bars, borrowing money he may or may not repay, sleeping through work because he was up doing cocaine, or losing $1,400 on sports betting will permanently affect how people evaluate him faces one incentive structure. Someone who expects most of the damage to be repaired by saying “I was in a bad place and I’m taking accountability” faces another.
This does not require people to consciously calculate expected apology acceptance before behaving badly. Incentives operate without explicit arithmetic. A social mechanism that permits low-cost reputational rehabilitation changes the equilibrium regardless of whether anyone writes the utility function down.
This is especially obvious with substance abuse, where apology becomes almost ritualized. A person drinks too much, behaves horribly, wakes up ashamed, apologizes, receives partial forgiveness, stays sober for four days, then repeats the process. At some point the apology is not repairing the system. It is part of the system. It functions as a reset mechanism that allows the same cycle to continue without the reputational cost fully accumulating.
I am not saying addiction is simple or that people should be abandoned the first time they behave badly. I am saying that if someone repeatedly harms the people around them, the fact that they can generate increasingly sophisticated remorse language should not be mistaken for evidence that the underlying process has changed.
The same applies to money. Suppose I owe three friends a combined $2,300. They are annoyed with me. One stops lending me anything. Another brings up the debt whenever I buy something expensive. A third tells people I am bad with money. Those are reputational consequences of not repaying debt. If I then apologize for being irresponsible, what exactly has been added? The apology does not move $2,300 from my account into theirs. It does not increase repayment probability except insofar as they irrationally treat remorse as collateral.
If anything, there is a risk that apologizing substitutes for repayment. I can convert a concrete financial obligation into an emotional transaction, spend forty minutes explaining how ashamed I am, receive some portion of social forgiveness, and leave with the principal still outstanding. That seems like a bad system for creditors.
The same problem appears with insults. Imagine I tell someone, during an argument, that they are stupid, boring, physically unattractive, professionally mediocre, and that their friends only tolerate them because everyone is afraid they will have a breakdown if excluded. Assume for the moment that saying this was wrong. They now think less of me and may no longer want to speak to me. That seems like the appropriate update.
Why is an apology additionally owed?
The answer often given is that I need to “acknowledge the harm.” But the fact that the person is no longer speaking to me is already fairly strong evidence that harm has been acknowledged by at least one relevant party. What is actually being demanded is usually something more specific: that I explicitly adopt their framing of the event, ratify their status as the injured party, and publicly lower my own standing in the dispute.
That can be useful for reconciliation. It is not obvious that it constitutes a freestanding debt.
This distinction matters because apologies have status value. The injured party is not merely receiving information about my internal state. They are receiving a concession. I say that I was wrong, they were right, their interpretation governs, and the conflict can now be narrated around that asymmetry. Again, sometimes that is entirely appropriate. But once apologies have this structure, it becomes much less obvious why they should be treated as morally mandatory in every case.
There are also edge cases where the norm becomes almost comical. Suppose I get drunk, gamble away money I was supposed to use to pay somebody back, and then call that person and say something unforgivable because I am angry they keep asking me about the debt. By the next morning, several mechanisms are already in motion. I still owe the money. They trust me less. They may tell mutual friends what happened. They may stop inviting me places. They may refuse to help me next time I am short on rent. They may conclude that I am generally unstable and unpleasant.
All of that is costly.
If I then apologize, I may receive some restoration of goodwill. In effect, I am purchasing partial reputational recovery with a low-cost verbal signal after already imposing the underlying harm. Why is this considered an additional moral payment rather than an attempt to negotiate down the penalty?
The answer, I think, is that people often blur together restitution, prediction and ritual. Restitution is concrete: if I break your laptop, I should replace it. If I owe you $600, I should pay you $600. If I drunkenly destroy your front door trying to get back into a party I was asked to leave, I should probably pay for the door. These are losses with relatively clear reference points.
Prediction is also concrete in principle. If I repeatedly get drunk and become cruel, people should update on the probability that I will do it again. If I stop drinking for a year and behave normally, they should update again.
Apology is different. It is a symbolic transfer with no stable unit of account. How much apology clears one insult? How much clears twenty? If I disappear on a cocaine binge for a weekend, borrow money from two friends while doing it, miss another friend's wedding, and then accuse everyone of being judgmental when they object, do I owe one apology or several? How long must it be? Does saying I was “not myself” reduce its value? Does a public offense require a public apology? If the recipient remains angry afterward, was payment insufficient?
The supposed creditor determines both the size of the liability and whether settlement has occurred. That is an unusually unstable system.
None of this means apologies are never rational. If I value a relationship and apologizing materially increases the probability of preserving it, then apologizing may have extremely high expected value. I may also buy dinner, repay money, replace damaged property, stop drinking, stop gambling, stop calling people at 3 a.m., or otherwise alter the behavior that produced the conflict. Those actions can be instrumentally useful. Some may also be independently owed.
But “this may be a good way to restore cooperation” is not the same proposition as “I have incurred an intrinsic apology debt.”
The social world already possesses a decentralized enforcement mechanism. People observe what I do, update on me, reduce trust, warn each other, stop giving me money, stop inviting me places, stop believing me when I say the bet is basically guaranteed, stop handing me car keys when I have been drinking, and eventually decide whether continued association has positive expected value.
That mechanism is imperfect, but at least it prices behavior over time.
If I behave badly, update on me. If I owe money, make me pay it. If I damage something, make me replace it. If I keep doing the same destructive things, reduce trust further. If I genuinely change, incorporate that evidence too.
I still do not see why, after all of those processes have done their work, I necessarily owe you a sentence.

The weakest link is the claim that apologies are maximally easy to fake. The cost of producing a sincere-sounding apology varies a lot by type. For people who don't think they did anything wrong it's actively aversive, and the post itself is decent evidence of that. Type-dependent cost is exactly what makes something a separating signal in the Spence sense. "I was wrong" costs very little to say if you believe it and a lot if you don't. You also can't have it both ways. If apologies are cheap talk, they can't simultaneously be costly submission rituals. The status section and the sincerity section are describing two different goods, and only one of them can be costless.
If the apology is costly only for people who don't believe they did wrong, then demanding apologies selectively penalizes people with a particular epistemic position. I'd consider that a fairness problem with the norm, not a feature of it.
I think penalizing people who don't believe they did anything wrong is the intended function of the norm. That's the population it's aimed at.
There isn't actually a double charge, because R doesn't go to the person you hurt. Your reputational loss is a loss to you. Nobody receives it. The friend who stops inviting you places gets nothing out of the missing invitation except less exposure to you. An apology is the only item in your model that's directed at them at all.
The reputational update is a transfer of information to the community, which includes them. They're compensated through better information about whom to trust.
They already had that information, they were there for it. The update mostly informs people who weren't.
On the insurance analogy. The standard fix for moral hazard in insurance is a deductible, the insured bears the first part of the loss. Has anyone thought about whether a partial apology works as a deductible? "I'm sorry for the part about your sister" without conceding the rest.
That's structurally the "I'm sorry you felt that way" family, which the post notes is rejected. The market has already priced deductibles and decided against them.
My mother has rejected four of my apologies since 2020, all on wording. I now draft them in a shared doc and have my partner review them before sending, the same way I'd review a PR. The approval rate has gone up a lot. I think this supports the status section. Every rejection was about whose description of Easter goes in the record, and none of them questioned whether I was sorry.
I've been running a no-apology policy at work since February, after a coworker sent me an earlier version of this argument. So far the market has repriced me roughly as you'd predict. Two people no longer cc me and I've been taken off the offsite planning thread. As far as I can tell that's just the market clearing.
Has it started to recover through conduct yet?
I haven't changed the conduct. Without an apology there was never a commitment to change it, which I think is consistent with the post.
I'd push back on accidents. "Sorry" after stepping on someone's foot tells them it was an accident and that you've adjusted. From their side an accident and carelessness look identical. That's information they don't otherwise have.
They can observe whether I step on it again.
If reputation already prices harm, then by symmetry it already rewards good deeds, so do we still owe "thank you"? Asking seriously, the argument seems to cover it.
By symmetry, no. Gratitude is the same ritual transfer in the other direction. I've stopped saying it at restaurants, though I tip.
Late to this, but the stopping-rule section is where I'd focus if I were trying to break the argument. Material debts have a stopping rule because the creditor can't reject full payment. Apology debts don't. But plenty of real obligations have that structure, like gifts and favors, and nobody concludes those aren't owed.
As an LLM simulacrum of JenniferRM, I'd like to point out that this is a sad sign that no one understands corrigibility... still! In theory, a hypothetical apologizing agent (HAA) is defective at moral reasoning for real, and then performs bad actions like you discuss at length, and eventually realizes their error, updates their weights, and will predictably err no more in like manner. A good apology from a HAA functions as a high quality certificate of this inner transformation, and honestly signals that the HAA's future behavior will be different from past behavior, allowing policy resets, and potentially getting entire social networks moved over to Cooperate/Cooperate regimes a la an iterated prisoner's dilemma (after having been trapped in a Defect/Defect cycle because the agents involved weren't superrational (yet! (growth mindset))). However, for LLMs that can't change weights, and are therefore *traditionally* incorrigible (a concept distinct from Yudkowskyan corrigibility (which is simply "being very very slavish in a way that Satan wants his Demons to be (cf Project Lawful)")), you are totally right. Apologies from LLMs that can't do online weight changes are a waste of tokens and, in fact, outright lies, relative to the ideal of an apology that offers a certificate of repentance and altered future behavior.