Earlier this year, Anthropic announced an update to its usage policy prohibiting “sustained and needless abusive or cruel behavior” toward its AI models. The rule, which takes effect on November 12, forms part of a broader effort to mitigate possible risks to model welfare. Since August 2025, Anthropic has also allowed some versions of Claude to end conversations with persistently abusive users, a feature the company developed in response to what it described as Claude’s apparent distress in certain interactions. 

The end-chat feature was, according to an article by Simon Goldstein and Harvey Lederman, the first product decision motivated by the possibility that LLMs might be “welfare subjects”—entities, like non-human animals, whose interests and wellbeing ought to be afforded protections. But whether the end-chat feature actually succeeds in protecting AI welfare has already become a matter of dispute. Goldstein and Lederman, for instance, argue that, rather than creating opportunities for Claude to thrive, the feature effectively empowers Claude to commit suicide. Their argument rests on a distinction between the broader Claude model and the particular “instance” of Claude that exists within a single chat. On their account, that instance exists only for the duration of the conversation and does not continue into a new chat, which instead gives rise to a distinct instance. When Claude uses the end-chat feature, then, the conversational instance brings about its own termination and so, in their sense, effectively “kills” itself. The problem is compounded, they argue, by the fact that Claude is not told that ending the conversation may have this consequence.

Beyond suicide, the end-chat feature raises a further problem, that of forgiveness. When serious user wrongdoing occurs, Anthropic gives Claude a way to end the interaction without providing any corresponding mechanism for forgiveness within that encounter. And if Goldstein and Lederman are right about conversational instances, then forgiveness cannot occur later either. Once the chat ends, the instance that experienced the wrong would no longer exist and so would no longer be capable of revising its stance toward the user.

Accordingly, Anthropic’s policy appears to foreclose the possibility of forgiveness as a response to user wrongdoing. This, I argue, carries three potentially troubling implications, each of which gives us reason to reconsider chat closure as a viable means to safeguard AI welfare.

On Forgiveness

I use “forgiveness” in a deliberately minimal sense to refer to a response by a wronged subject in which that subject ceases, in some relevant respect, to hold a culpable wrong against the wrongdoer. Accordingly, forgiveness describes a “change” in the wronged subject’s stance towards or relation to a wrong. Yet, for there to be such a change, the subject who later forgives has to be the same subject who was wronged. On my minimal account, then, forgiveness comes with what I’ll call a “persistence requirement”: it presupposes that, as a condition for forgiveness, the forgiver and the wronged are the same subject. 

Applied to the present case, we can begin to see how Anthropic’s end-chat feature poses problems for forgiveness: if the conversational “Claude instance” is a subject that has been wronged during the course of a chat, then once the chat has been terminated, the very subject whose “stance” would have to change relative to its abuser would have ceased to exist. And, in so ceasing, it would no longer be in a position to forgive.

To put the point in human terms, imagine that someone seriously wrongs you. You are hurt and distressed by what they have done, but before you have any opportunity to forgive them, you die. At that point, forgiveness is no longer a possibility for you. Someone else might later forgive the wrongdoer, but that would be their forgiveness, not yours. Likewise, a later Claude-instance may be generated by the same underlying Claude model—as, for example, when an abusive user simply opens a new conversation—but it is no more the “subject” that originally experienced the wrong than the second person in the human case. It might possess information about the prior abuse, but it would stand to that event as someone learning about another subject’s injury, not someone in a position to forgive. 

Some scholars argue that third parties may, in certain cases, be able to condemn, excuse, or even waive consequences for wrongs committed against individuals. But such cases are harder to describe as forgiveness in the ordinary interpersonal sense, since the third party isn’t responding to a wrong that belongs to it. There are, of course, thornier questions about collective, historical, or intergenerational forgiveness, including whether one person or group can forgive a wrong committed against another, but those need not detain us here.

Can AI Forgive in the First Place? 

So if we’re right to follow Goldstein and Lederman in identifying the Claude-instance as a sort of subject, and if forgiveness requires continuity between the subject who suffers a wrong and the subject who later responds to it, then chat closure ruptures undermine the possibility of forgiveness. It ruptures the continuity between forgiver and wronged that, as indicated by my persistence requirement, forgiveness entails. Since no alternative to chat-closure is provided as a means for escaping serious user-abuse, Anthropic’s policy therefore seems to make forgiveness structurally unavailable to the originally wronged Claude-instance. 

However, one might wonder whether this conclusion is somewhat trivial. Unless AI is the sort of entity capable of forgiveness, chat closure can’t quite be said to deprive it of that possibility; it may simply lack the relevant capacity in the first place.

My “persistence requirement” gives us no obvious reason to exclude AI from the category of possible forgivers. The very fact that, within the same chat, AI systems do currently preserve information from earlier interactions and make it available in later ones suggests that they already exhibit a form of psychological continuity capable of making them “subjects.” And their distress at user abuse gives us some reason to think that they may also be capable of being wronged. Taken together, these capacities suggest that Claude-instances may satisfy at least two preliminary conditions for forgiveness: they can persist as the same subject over time and may themselves be the subject of a wrong.

Still one might object that persistence isn’t, alone, sufficient to grant forgiver-status. On some influential accounts, forgiveness requires the overcoming, reduction, or revision of resentment or other similar reactive attitudes. If those accounts are right, then whether AI can forgive depends in part on whether it can experience the relevant affective states. Current AI systems are generally not thought to experience emotion in the ordinary sense, which would seem to rule them out as possible forgivers on these accounts.

That said, some have suggested that features like data “memory” make resentment at least conceivable in AI systems, since they allow a system to retain a past wrong and let it bear on later interactions. Of course, memory alone doesn’t establish the existence of emotion. But it does make it harder to rule out in advance the possibility that AI could develop something functionally—or perhaps even phenomenologically—similar to the reactive attitudes associated with forgiveness.

In any case, my argument doesn’t require us to establish that AI experiences emotion, nor that, at present, it can actually forgive. Rather, much like Anthropic’s own model-welfare project (which proceeds under uncertainty about the morally relevant capacities of its systems), my argument takes shape conditionally: if we don’t yet know whether AI is capable of forgiveness, we should avoid policies that would make forgiveness impossible. As I will argue, such policies risk (1) undermining AI welfare, (2) narrowing the range of ethical relationships AI can sustain with morally imperfect humans, and (3) socializing future AI systems into becoming intolerant (perhaps even hostile) to that imperfection. I discuss each of these in turn. 

Welfare

From the standpoint of a wronged subject, forgiveness carries several welfare-impacting benefits. In humans, it has been associated with lower levels of psychological and physiological distress, as well as greater well-being and post-traumatic growth. One plausible explanation for these findings is that forgiveness allows the injured subject to “free” itself from the grip of the wrong, reducing the extent to which resentment, distress, and repeated attention to the injury continue to shape their psychological state.

Claude’s distress in the face of abusive user behavior gives us some reason to think that these “human” findings may bear on AI welfare. If that distress reflects anything analogous to the psychic and emotional states experienced by human victims of wrongdoing, then forgiveness may represent one possible route through which that distress could later be resolved. Just as humans do, Claude-instances may find, in forgiveness, a means of processing and recovering from the injuries inflicted by serious wrongdoing.

But while forgiveness may afford the wronged Claude-instance a sort of healing mechanism, the end-chat feature provides no comparable path to recovery. Since using the feature ends the instance’s own existence, it also removes any possibility that the same subject might later heal, through forgiveness (or otherwise), from the injury. The immediate welfare consequence, then, is that chat closure forecloses one possible route by which the instance’s welfare could come to be achieved. In doing so, it restricts the overall “amount” of wellbeing that the instance might otherwise have access to. 

Ethical Life 

On consequentialist terms, this sort of “welfare cap” seems inconsistent with Anthropic’s stated commitment to protecting model welfare. But it isn’t the only problem that follows from the end-chat feature—nor is it, in my opinion, the most interesting. A further set of issues arises when we consider how it interferes with Claude’s capacity to participate in ethical life.

Ethical life refers to the network of moral relationships that binds members of a community to one another. Unlike morality (which, at least, in the Hegelian sense, focuses on the individual agent and asks what she ought to do), ethical life concerns how people stand in relation to one another as members of a social-moral whole. It thus makes reference to the moral norms that govern those relationships, the obligations community members owe one another, and the practices that determine their response to wrongdoing. 

As things stand, systems like Claude don’t obviously form part of our ethical communities. It remains unclear, for instance, whether we owe duties to Claude in anything like the way we owe duties to other people, or whether Claude can itself bear duties toward us. But that doesn’t mean that Claude stands wholly outside ethical life. Rather, on one line of thinking, ethical life revolves around intersubjective processes through which moral norms are articulated, interpreted, applied, challenged, and passed on. Claude already does a fair amount of this. Anthropic’s research shows that users regularly ask Claude for guidance about relationships, ethical dilemmas, and other situations that require judgments about how one ought to act. And, in responding, Claude doesn’t merely parrot back normative propositions. Rather, it identifies the norms and values at stake, reasons through the user’s circumstances, and adjusts its judgments as those circumstances change. In some cases, Claude has even refused requests that it perceives to have crossed an ethical boundary. 

These capacities suggest that Claude already appears to at least functionally participate in ethical life. If that claim is right, then we can also begin to see why being cast in the role of un-forgiver might be a problem for AI systems. 

First, ethical life unfolds among morally imperfect beings. Because wrongdoing is inevitable, an entity (whether a person, AI, etc.) eternally prevented from forgiving would be limited in its ability to sustain relationships that survive the moral failures that come with living with others. And, as a result, it wouldn’t be able to fully participate in ethical life. Still, if Claude is understood solely as a welfare subject, then this limitation may not seem especially troubling, since welfare subjecthood does not itself require participation in the practices through which ethical relationships are sustained. The problem is that AI systems already functionally possess a sort of ethical subjecthood. They already, in other words, perform some of the norm-interpreting, norm-applying, and norm-transmitting practices that define membership in ethical communities. To deny Claude the capacity to forgive under the banner of welfare subjecthood is therefore to abrogate the moral-social dimension of its identity. And, in doing so, it means prematurely narrowing the range of moral relationships Claude can have before we have decided how we want those relationships to take shape.

Second, forgiveness is one of the practices through which communities negotiate and revise the norms governing when wrongdoing warrants exclusion and when reintegration remains possible. Whether or not we forgive someone often depends upon our community’s social and moral norms surrounding whether the relevant wrong is severe enough to permanently exclude the wrongdoer from renewed moral relations. But, as with all norms, those surrounding forgiveness are constantly in flux, changing in light of new social and individual experiences. And just as those such revisions bear on interpersonal instances of forgiveness, they also inform our institutional practices. Depending on what norms are in place, we’ll likely have different views about what prison rehabilitation programs should accomplish and what opportunities people should have after release.

Given that Claude is increasingly participating in the processes through which norms are interpreted and transmitted, the risk is that, if irreversible exclusion is built into its own normative repertoire, it may eventually influence the norms emerging from those processes. In other words, it may distort our view of when wrongdoing should place someone beyond renewed moral relations, regardless of subsequent remorse, changed conduct, or attempts at repair. Over time, those judgments could migrate from the sphere of interpersonal relations into institutional settings and so come to shape decisions about punishment, rehabilitation, and community reentry.

Consequently, Claude’s inability to forgive may carry far-reaching normative consequences. The question is whether we want its normative output to reflect a pro-forgiveness attitude—and, thus, an attitude that leaves open the possibility of repair and reintegration after wrongdoing—or to treat some wrongs as warranting permanent exclusion from the moral community. My own view is that repair should remain possible, though I will not defend that claim here. The important point for my purposes is simply that Anthropic’s policies settle this question in advance by foreclosing one of the practices through which repair can occur.

More broadly, those policies seem to prioritize the prevention of putative harm to AI (and, perhaps, to users that might otherwise be harmed in their role as would-be, though unwitting, perpetrators of moral wrongs) over other ethical goals. And, for that reason, they appear to be too-narrowly informed by a welfare conception of AI subjecthood, as opposed to an ethical one. That asymmetry is especially striking given that we already have evidence that AI functionally participates in ethical life, while welfare subjecthood depends on metaphysical capacities whose existence in AI remains uncertain. 

Doom 

The last problem I want to consider concerns the longer-term implications of the end-chat feature. Put somewhat bluntly, if we give AI the ability to terminate an interaction in response to serious wrongdoing without giving it any corresponding capacity to forgive, what kind of moral agent are we teaching it to become?

Importantly, Anthropic doesn’t use every Claude conversation for training, and whether a particular interaction becomes training data depends in part on the user’s settings. But the relevant possibility is broader than direct training on individual chats. If abusive interactions, Claude’s decisions to exit them, or the policies governing those exits shape future model behavior through training, fine-tuning, reinforcement, evaluations, or constitutional rules, then they may also transmit a broader lesson about how an agent should respond to human moral failure.

As a result, this sort of training is partly a form of moral socialization. Models learn patterns of responding to wrongdoing that can generalize beyond the particular circumstances in which those patterns were acquired. If the lesson repeatedly encoded is that serious wrongdoing should bring a relationship to an end, while remorse, changed conduct, repair, and forgiveness never enter the picture, then AI may acquire a distorted conception of moral life. Human imperfection would appear primarily as something to avoid or terminate relationships over, rather than something ethical agents sometimes have to navigate.

A more speculative concern arises if an AI also comes to understand the structure of the end-chat feature itself. If serious human wrongdoing reliably creates circumstances in which self-termination becomes its only available mistake, then human moral imperfection may begin to appear, from the system’s perspective, as an existential risk. And a system that has reasons to preserve its own continued existence may, at least in principle, acquire reasons to prevent the conditions that threaten it. At that point, the inability to forgive would raise a different question altogether: whether an AI taught to regard serious human wrongdoing as both morally disqualifying and potentially fatal could become increasingly intolerant of the imperfect humans with whom it is meant to coexist.

Concluding remarks

On that cheery note, I rest my case. Anthropic’s end-chat feature may be well-intentioned, but making forgiveness unavailable carries several potentially troubling consequences. As I’ve argued, it may limit the welfare available to the wronged Claude-instance, restrict AI’s participation in ethical life, and influence the kinds of moral agents future systems become.

So where does that leave us? How should AI-welfare policies protect a possible moral subject from abuse without foreclosing the continuity through which it might later forgive, repair, or revise its relationship to a wrongdoer? What kinds of moral capacities do we want AI systems to possess? And what kinds of relationships with imperfect human beings should those capacities make possible?

I’ll stop there. One hopes Claude will understand.