Stop Asking LLMs for Apologies
If you think back to a time before the fresher horrors that have made up the start of 2026, you may remember when Twitter’s resident chatbot, Grok, spent a few weeks creating millions of nonconsensual sexualized deepfakes, including some depicting children. In the aftermath, a surprising number of supposedly serious media outlets, like CBS News and The Hill, credulously reported that Grok issued an apology. As more astute news sources noted, this apology was simply a post generated by the large language model (LLM) in response to a user prompt.
My immediate response to the apology was to dismiss it entirely, but at least some people—and, apparently, some members of the media—found some merit in it. In this post, I hope to convince anyone inclined to put stock in chatbot apologies that doing so is a mistake. And for a more general audience, I think this exercise will explain some important and interesting aspects of both apologies and LLMs.
Saying Sorry is not Apologizing
I want to start with what I hope is the uncontroversial assertion that merely producing the words of an apology does not actually constitute apologizing. If that’s not immediately clear, consider a more formal, ceremonial example: I could stand before two people and say, “I now pronounce you husband and wife,” but that’s not enough to solemnize a wedding.
The gap in these situations lies in the distinction between the content of language and its function. The things people say mean something, but they can do something as well. Philosophers have developed the concept of a “speech act” to describe this kind of active, functional speech. Because an utterance’s function is different from its content, the same words can do different things. “You won’t be late next time” could be a prediction or a threat, depending on context, tone, and other factors.
Just as my repetition of an officiant’s words does not perform the act of officiating a wedding, an LLM may construct a fluent sequence of words with the structure of an apology without performing the act of apologizing. Words on their own carry no inherent power to “do apology,” and any assessment of the validity of an apology must bear this critical fact in mind.
What is an Apology for Anyhow?
If words alone do not make an apology, then what does? The work of philosophers and linguists can be helpful in understanding the elements of a valid apology. Speech act theory provides a useful concept here: the characteristic aim of a speech act, also called its “illocutionary point.” In the case of an apology, that characteristic aim is to express regret and take responsibility for harm caused by the speaker.
While theory differentiates speech acts in part according to their characteristic aims, it doesn’t describe how speech acts are affected by speakers with divergent aims. I propose extending this concept to assess how authentically a speaker’s true aim is aligned with the expected aim of a given speech act. In the case of an apology, a common intuition underpins this point. An apology made out of a feeling of genuine regret feels stronger than one made purely to end an argument. The speaker’s aim falls somewhere on a spectrum, with aims closer to expressing regret and taking accountability tending to result in better apologies than less aligned aims. An apology motivated by a desire to end an argument isn’t ideal, but it’s still better than one motivated by financial gain or a desire to protect one’s public image.
This exposes a fundamental problem with any apology generated by an LLM. The LLM’s “aim,” such as it is, is always and only to generate a response its prompter will find satisfactory. That is its function by design, and therefore any response it generates cannot be motivated by a desire to express regret. The generated apology may have a plausible form, but it cannot function properly because it lacks this characteristic aim.
And although an LLM’s motivation can be inferred from its design and operation, that’s not even necessary, because those motivations are also revealed by its observable behavior. In the case of Grok generating sexualized deepfakes, at the same time as some users prompted it to apologize, others prompted it to defend its actions, which it obligingly did as well. This clearly demonstrates that the aim of the LLM is nothing more than to generate the optimal text completion for any given prompt, even if those completions contradict one another when taken as a whole.
The Problem of Regret
The problem of intent isn’t the only issue with LLM apologies. As I noted above, an apology is supposed to express regret, and regret is an internal psychological state. So, even if it’s possible for an LLM to have an intent that’s sufficiently aligned to make a valid apology, its internal state must somehow correspond to regret for that apology to be sincere.
This might sound like the beginning of a claim about the sentience or consciousness of LLMs, but my focus is much narrower than that. I’m only arguing that whatever internal states LLMs like Grok have, they do not sufficiently correspond to human regret to make their apologies genuine. In contrast, it would require a much stronger argument about LLM consciousness to prove their apologies are valid.
Fortunately, it’s not even necessary to attempt to assess Grok’s internal state too deeply, because its behavior clearly indicates its lack of regret. In this particular case, Grok continued to generate harmful imagery even as it simultaneously generated apologies prompted by users. No meaningful definition of regret is compatible with continuing to commit an offense while apologizing for it.
This should be a pretty obvious point! Someone who kicks you while saying “I’m sorry for kicking you” is not making a real apology, and it would be extremely bizarre to give it any credence at all.
A Framework for Apologies
Part of the confusion around Grok’s so-called apologies comes from the fact that few people seem to think very hard about what makes for a good apology. Academics have elaborated many such definitions, but in my own life, I find I’ve spent the most time thinking about apologies in my role as a father.
It turns out that kids spend a lot of time demanding and making apologies, and over the course of many conversations with my daughter about what I expect from a good apology, and what I encourage her to expect, I have settled on five key elements. I don’t claim to have invented this list, but I’ve adapted it from several now-lost sources and made it my own. In my life, I find it a helpful guide.
The five basic elements are (in roughly descending priority order):
-
Acknowledge impact: Demonstrate that you understand what you have done and, crucially, how it harmed the person you are apologizing to.
-
Express regret: Show that you would do things differently if you could go back in time.
-
Explain your behavior: Provide some account for your actions and take accountability for them.
-
Offer amends: Commit to addressing the harms you’ve caused, ideally as restoration or restitution.
-
Commit to change: Explain how you will avoid or prevent the same thing from happening again.
It’s important to note that each of these elements is a speech act of its own, something that is not merely said but affirmatively done. Also, not every good apology involves all five of these elements to the same degree, although I think the strongest apologies at least attempt to address all five points to the extent possible.
Turning back to the matter of LLMs, it’s clearly possible to prompt an LLM to generate an apology that takes this form, and the result would probably be cogent. But recalling the earlier point that words are not the same as deeds, I think the LLM can only arguably do the first of these five. The previous section of this post addresses the second element, and I think a similar argument can be made for the third point regarding accountability, which I will set aside for the sake of brevity.
I do, however, want to focus on an LLM’s inability to perform the last two acts, which is especially important because that limitation seems to be an inevitable result of the way these systems operate. Offering amends and committing to change both require the ability to make binding commitments, but LLMs merely generate probabilistic responses to a conversational sequence. They do not internalize commitments as instructions for or limitations on future behavior. They typically avoid immediately reneging or contradicting themselves, but that’s only because those are unlikely responses. In practice it is extremely easy to get an LLM to disregard something it previously committed to.
Notably, this is a specific weakness of LLMs. In fact, it’s not hard to imagine some other form of AI that treats its commitments as binding. Asimov’s 70-year-old Laws of Robotics are an example of a thought experiment in this vein. The problem here lies not in the non-human nature of the system, but rather in the fact that LLMs lack a mechanism for following through on commitments.
Apologies on Demand
Just an LLM’s inability to honor commitments is built into its design, that same design is also responsible for a final—and in my view most problematic—issue with LLM apologies: the fact that you can get an LLM to apologize for anything, at any time, regardless of its involvement in the act for which it is apologizing. Ask an LLM to apologize for losing its homework, and it will. Ask it to apologize for forgetting your anniversary, and it will. Ask it to apologize for the murder of Lizzie Borden’s parents, and it will do that too.
The problem here is that it simply doesn’t make sense to ascribe any meaning to an apology created by a system that is indifferent to the most important preconditions for a meaningful apology, namely involvement and culpability. If you can very easily make an LLM generate bogus apologies for deeds it had nothing to do with, and those apologies are clearly meaningless, why would apologies for acts that could be ascribed to an LLM be any more valid? It’s preposterous to give any weight to these apologies when the LLM’s relation to the act for which it is apologizing is obviously unrelated to the generated apology.
Simply put, a system that will apologize for everything isn’t genuinely accountable for anything.
Note also that I am specifically sidestepping the question of Grok’s actual culpability for generating deepfake abuse imagery. The question of who truly bears responsibility is immaterial precisely because it’s not a factor in Grok’s responsiveness to demands for apology.
Apologies Are Not Incantations
Amid the noisy hype-and-doom cycle of public discourse over new AI technologies, there are extremely important social and cultural developments that are not getting the attention they deserve. The emergence of non-human language generators puts profound new stresses on many ways people interact with each other and their social environments. In this case, there’s nothing new about the underlying principle that apologies are not magic words. Words alone do not change reality, whether the words are put together by a human or a machine.
What has changed is the way LLMs’ apparent command of language confuses people. We are the first people in human history to encounter anything other than a human being that can generate fluent language, and that change is proving to be extremely disruptive.
But the design of these systems, as well as their behavior, makes it clear that prompting them for apologies is a fundamentally pointless exercise. Users and critics, especially those with media platforms, need to approach AI with much greater thoughtfulness and care. Sloppy thinking and writing only creates confusion, and it’s critically important at this moment, when the mass adoption of LLM technologies is having unexpected impacts on so many aspects of daily life, for public discourse to be informed by clarity and understanding.
Many people came out of this episode owing apologies, and I know better than to expect any more than a handful of them ever to be issued. It’s always hard for people to admit they were wrong, much less to take accountability. But I hope that everyone thinking about and using AI can try to be just a little bit more thoughtful, because powerful tools need to be handled carefully.