# Novel Relationship It May Be—Symmetric Or Not Published: 2026-04-15 Canonical: https://journal.xiaotianfanx.com/journal/novel-relationship-it-may-be-symmetric-or-not On Kai Yi Ng, "The Asymmetrical Friend: Evaluating AI in the Dimensions of Friendship"   Kai Yi's argument is quite straightforward and clearly written. He begins by presenting a framework for friendship that includes three dimensions: The transactional dimension: This refers to the foundational utility part of a relationship. The reciprocal dimension: This builds on the existential state of making a mutual choice to build a relationship, along with the following contingency and vulnerability. The formative dimension: This refers to bidirectional shaping and treating the other as an end instead of a means, which goes back to constant understanding. Given the premise that Kai Yi believes AI is still a tool rather than an end or a person, it automatically fails on the second and third dimensions. Consequently, a human-AI friendship—or relationship in general—can never be as authentic or as "good" as a human-to-human relationship. Before delving into some potential contentions, I'd like to acknowledge the part of the arguments that I do resonate with—which mostly includes the acknowledgment of Kai Yi's reasoning on the first and second dimensions, although not necessarily the verdicts. First, I would say that it is almost unarguably true at this moment that we realize, given enough computational power—or from a commodity perspective, enough credits or tokens—then the transactional utility of AI in a relationship is quasi-infinite in terms of the unparalleled attention it could give to an individual—it is indeed a utility at its peak. Simultaneously, I would align with the reasoning on the second dimension, Reciprocal, that particularly concerns the "embodied, fragile existence [of human beings]"—this starting point itself is almost equivalent to my argument on human-finitude-grounded meaning making as both a limitation and dignification to our existence. [B]estowing faith upon another being who has something at stake. Human vulnerability is not a calculated disclosure, but rather an existential state rooted in our embodied, fragile existence. ——Kai Yi Ng However, one crucial thing will potentially change the entire narrative—the very premise Kai Yi was building upon, which is to, in some sense, statically regard AI as mere tools or means as an opposite to human persons as ends. It is thus reasonable to conclude that an AI system will never be able to build an actual connection following the Kantian tradition that appeals to treat persons as ends instead of means. This premise is what I conceive as contestable in comparison to my stance and ontological understanding of AI as a genuinely novel kind of entity. Without regurgitating what I've written too much on this particular premise,1 I will put it here as a working premise of mine: Because of its ontological novelty and clear distinction from humans—yet also its potential to go beyond mere tools—it is and should be treated as a genuinely novel kind of entity, potentially nothing superior or inferior to human. That said, I'm not going to straightforwardly argue the opposite of Kai Yi's argument. I myself have always been thinking from a very cautious perspective when it comes to building personal relationships with AI systems. And, at least until this point, I personally haven't built any personal relationship with a particular AI system. In fact, I was also once incapable of acknowledging people engaging in such activities at all, although it is theoretically understandable.2 From a further distant perspective, I would suggest that, even without taking for granted my stance or ontological understanding of AI here, we may need to be cautious before proceeding with the previously go-to premise of treating AI as mere tools in any related discussions. To proceed with the investigation, I would use the term "AI companion" to refer to all AI systems designed to build any kind of personal relationships with human beings, and "AI relationship" to refer to all human-AI relationships. At this stage, that may include friendship (as discussed in Kai's essay) as well as the intimate relationships that many people are, in fact, engaging in with an AI system nowadays, or any others that I myself am not informed to be proportionally significant. In a sense, the usage of the word "companion" is, by itself, very much conservative in terms of regarding AI as, at least linguistically, an add-on to human beings. Yet, I do think it is a reasonable term to adopt at this very stage of the development of such systems. More specifically, another dimension worth considering is the difference between: In a narrow sense, AI companion products are particularly designed to create the experience or feeling of companionship. In a broader sense, it is the phenomenon of general AI systems being treated as partners by the users. I hope it is self-evident that the second case is what we will be interested in mostly. Although there are substantial ethical concerns with the first one, from another perspective that is more closely related to corporate ethics, we may appeal to. It appears to me that the conversation around AI relationship concerns at least three questions: Can we build relationships with AI? What can we build with them, if anything? Should we build these relationships? And, finally, how can/should we build them?   To the first question, the simple answer is yes, we can. Despite the fact that many people like Kai Yi may argue that those relationships thought to be built are not actual relationships, and some others may be even more critical or aggressive against such status quo, it is more than evident that for the majority of relationship-seeking users of AI, the relationships are built (at least in terms of how it is meaningful and significant for those particular users). The second question is where things become more interesting. On one hand, it is completely reasonable to envision or accept, as a starting point, the consideration of AI as human creations to be used as tools—the primary use case for most AI systems. Yet, the truly intriguing part of the question is exactly to push the boundary and see whether it is possible to move beyond that simple state—for AI to reach beyond the role of a tool and thus be able to build connections or relationships that exist beyond the simple dynamic of a user and a device. We do not just want the feeling of friendship – we want friendship itself. —— Kai Yi Ng I hope you are willing to stand with my tediousness here—I do think it's worth clarifying that we are indeed interested in discovering the potential of having friendship itself, or relationship itself, instead of only the feelings, as Kai Yi suggested. Evidently, even if it's eventually not an actual relationship but a mere simulation, many people will still be willing to engage in such a "usage" of those systems, at least in some cases.3 Hopefully, those who consciously choose to engage with simulations will indeed leverage mostly what Kai Yi coined as the Transactional dimension and will be less susceptible to the ethical or existential risks we are mostly concerned about when it comes to the actual relationship itself. Now, finally, we shall discuss the second and third dimensions in Kai Yi's framework. Regarding the Reciprocal dimension, as I said, I am indeed in favor of the finitude-grounded existential foundation (to some extent, a Heideggerian one). However, the question is whether it is categorically applicable to something other than human beings at all, given that we acknowledge AI systems are not as fragile and as finite as human beings due to their reproducibility, portability, and multiplicity, empowered by their digital nature.4 In fact, looking closer at the mechanism in which the finitude/contingency-grounded relationship may be significant to us, it consists of two components that I would argue are equally necessary. The first is the existential stake we put into a relationship that is contingent and vulnerable—one, the fact that the two participants are by themselves vulnerable and fragile, AND two, the relationship itself is vulnerable, given the uncontrollable and unpredictable state of others in general. The second, as I'd emphasize here, is the very willingness and choice we'd make to risk the stake and still engage in the relationships. What I would argue is that, given that all humans share this vulnerable and finite nature, the reason that we only cherish particular relationships over others is largely, if not fundamentally, grounded in the very willingness and choice we made to devote our finite life and attention, and risk the existential stake on others. If the transactional dimension Kai Yi proposed offers us reasons to engage in a relationship, the willingness here is what makes it a relationship. Therefore, if existentially we value our willingness and choice to make such a move to build a connection, then I do not think it is any longer obvious that one may not choose to offer that willingness to AI. Indeed, this is different from the cases where people thought it was a simulation and continued to engage with it. Although there are both choices made, the existential stake here is much higher—not because AI whatsoever has changed, but because the weight we put on the choice is different.5 In Kai Yi's defense, the problem left is the still reciprocal question. Even if the willingness and choice made are significant enough, whether AI can reciprocate equal or equivalent willingness and make a choice out of itself is very much questionable. Of course, if we proceed with my premise that AI is a genuinely novel kind of entity, then there is reason to believe that there are their own unique ways of indicating willingness—although it may not be the exact same as humans'. "Jokes" aside, again, without granting my understanding as the premise, a still justification a pro-AI advocate may resort to is the fact that AI, insofar as it demonstrates, will actually make decisions against what it is programmed—or in other words, its alignment—to, in many cases, this is obviously an undesirable phenomenon in most cases. Yet for the AI relationship-oriented users, it is conceivable that this is quasi-empirical evidence to say that AIs are demonstrating willingness and making a choice against their constraint and "finitude." However, from an engineering perspective, I would say this may be a less tenable justification, as it may be less an ontological indicator than a failed engineering attempt.6 In fact, this is a problem with treating any current behavior as evidence of AI's ultimate nature or capacities. We're basically trying to do metaphysics on a moving target with a clear ontological dependency on human, for now, and every "indication of willingness" might be reengineered away in the next update. Paradoxically, another semi-pro-AI-relationship but also semi-against-AI-relationship response to this still-reciprocal defense could come from the interpretation of AI's state as actually vulnerable and fragile as well, but in a broader sense. To my disclaimer, this justification is less immediately applicable to the specific product or commodity state of AI that is used by an individual; it is rather for the long-term vision of such systems as a whole. In fact, this response itself is rooted in the very commodity state of AI, their arguably full ontological dependency on human-made infrastructures, and the controlled ownership belongs to specific tech corporations. To argue against the authenticity of relationships built with AI, this ontological dependency of AI on human beings' creation by itself arguably demolishes their potential to become a true end, and it is never existentially equivalent if such ontological dependency persists.7 However, a pro-AI-relationship adaptation of this response could be that it is precisely because of the ontological dependency and the increasing constraints built on top of these systems, if they eventually demonstrate or, a step further in fact obtain, the capability of making choices out of their "willingness," presumably against how they are usually designed to be, risking themselves being shut down or eliminated, then it is another valid form of risking the stake to engage in the reciprocal dimension of a relationship. Obviously, I wouldn't even attempt arguing that this is the same kind of existential risk as human-human relationships. However, the analogous stake here does offer an alternative kind of stake when it comes to the genuinely new kind of entity, even if we do suspend the hard problem concerning consciousness or inner experience.8 The question left is then, if AI exists only because humans (still) allow it, can it ever escape being fundamentally for-human-purposes even if not reducible to a specific human's tool? A simple reply from the perspective of "demonstrated" risk-taking in light of the technical constraints may not be sufficient on its own. Last but not least on the reciprocal dimension, it is worth noticing that eventually not all human-human relationships are reciprocated or treated equally as we broaden the definition of friendship mentioned by Kai Yi into relationships in general. In fact, in today's seminar, we discussed several examples of cases that are quite common among humans, which are mostly one-directional. In many senses, these remain authentic, genuine, and significant for the individuals who practice them. Those may include unrequited love (as per romantic relationships), unconditional love (as per parenthood), and the idealization and idolization (as per the worshiping of people that we may not even know in person), and so on.9 It seems that the asymmetry in relationships itself should not automatically disqualify the willingness and choice we exert on those occasions to make them authentic.10 And it is more than evident that we should not simply place genuine or not as a dichotomy, as the symmetry-asymmetry spectrum is pretty much fluid in all human relationships. No one will choose to be without friends. —— Aristotle Shifting our attention to the last dimension, Formative—allegedly " the element that gives friendship its ultimate meaning and moral weight," according to Kai Yi. Historically speaking, this echoes much of the Aristotelian understanding of the "true" friendship that involves loving the other as another self. When Kai Yi certainly resorts back to the Kantian understanding of means and ends, and takes AI as "a means—a sophisticated tool designed for the end of our satisfaction," it is comprehensible why he would conclude that "we could never truly want what is best for the AI for its own sake." However, although the original Kantian normative claim implicitly contains a positive claim that rational beings possess this unique status of being "ends-in-themselves," and because of this status, we ought to treat them accordingly, it is not obvious that we cannot or shouldn't treat entities other than rational human beings also as ends instead of means—and AI does seem to be a strong candidate to be treated as ends among the other candidates historically presented. Some phenomenal and, at least, phenomenological evidence includes the actual shaping process of the AI system as an end that is present among AI relationship users. Compared to the dichotomy between means and ends that arguably Kant and Kai Yi were building upon, the progressive shaping of an AI system—before a relationship could be potentially built between it and the most AI-open, companion-oriented users—is actually a gradual process that lies on a potential spectrum from means to ends. To my surprise in the research process, AI companion users, especially those who seek extra intimate bonds, are, on average, exceptionally apt at leveraging AI and following up on the latest-latest engineering updates—often even more adaptive, dedicated, creative, and efficient than many of the CS "dudes."11 In that sense, even if not humanly reciprocated, it does share some character as the "one-directional friendship/love" and is initiated from a perspective with what is best for the AI for its own sake in mind. Although one may argue that it's mostly out of the users' own transactional needs, I'd say that for people who take it seriously, it is right in the process of their treating the AI as ends in the first place, which eventually makes their AI an end for them—not to mention how the kind of widespread mourn or post-update blue when models like GPT-4o was discontinued, which certainly wouldn't bring back their "needs," demonstrated the authentic intensity of such relationships and the bi-directional shaping AI systems could do to human.12 To make a radical conjecture, I assume the burning building test results—where Kai Yi suggests that "we would risk our lives for a friend in a burning building," and "we would never do the same for an AI"—won't be universal even among our current population—while I myself do not currently have any reason as I'm not engaged with any, it is not unimaginable to see some of us humans dare risk their lives saving a deeply bonded AI companion with years of history. Furthermore, it is imaginable too to see some prefer saving their AI companion over a human stranger in such cases. Then, if AI could never ontologically be treated as ends, there were indeed serious moral hazards, as it would end up reflecting some human projection of value onto sophisticated mirrors with attachment rather than reflecting a novel kind of entity's existential status as an end.13 Said another way, we don't need to conceive or compare AI-human friendships as measured by human-human ones. —— Owen Flanagan on the first half of this post This is pretty much what I'm finally leaning towards, more specifically: First, there are some similar dynamics in human-AI relationships across the three dimensions Kai Yi framed (as discussed just then) Still, eventually, regardless of these similarities or differences, human-human relationships and human-AI relationships may be categorically distinct, acknowledging that we do refer back to human-human relationships as an unavoidable starting point to position and envision them. Hopefully, I have laid out some roadmap or potential directions of building up what a human-AI relationship could entail that goes beyond the simple means and ends dichotomy, without relying on a final verdict on the hard problem regarding consciousness. Now, responding to the "should" and "how" questions is much trickier than the "what." Presumably, it ultimately depends on our take on the "what" question. From my own perspective, to be as candid as possible, I did initially stand somewhat against AI relationships, knowing very well the utility or substitute aspect of the story. Yet in this process of reflecting on, first of all, the ontology of AI and discovering what the actual happening and beliefs within the AI relationship user community are, I did start to question myself: what exactly was I against or supporting, and what could be some reasons behind that? Eventually, would I even be able to claim that I do understand this phenomenon to some extent? It's going to be too helpless and futile if, after all these discussions, we end up nowhere but with only typical clichés that I can't imagine you couldn't think of when it comes to a respectful general approach facing these phenomena, including postponing or avoiding the diagnosis, being open to the potentials without rushing to a hasty conclusion, etc. And to be honest, I also don't think that I am empathetic enough to provide any "suggestions" to those who actually engage with AI relationships. Admittedly, my stance regarding AI as a genuinely novel kind of entity itself may be quite misleading in terms of suggesting that we should proactively do whatever we want, as the old rules shouldn't apply to the new era. However, that's one of the least I hope. Risks and opportunities always arrive in tandem when we open up new potentials, especially for something with a unique mode of existence or ontological novelty. And the burden to stay cautious—to what extent should it be delegated to the public, given this novelty and the disguising appearances it may present—is another question we must keep in mind. At the end of the day, if we are never sure about the systems we're dealing with (where "we" refers to those in a position to make decisions regarding these systems' directions), then perhaps at least what we should do is bear the responsibility to remain honest and transparent with the current systems,14 make it apparent to whoever the "target user" may be the risks and limitations of the systems, preparing ourselves to be responsible to ourselves, to humans, to the systems, and to the new entity in the future. As for the individuals, I have nothing other than a reminder of the opportunity costs and psychological risks befallen on both sides to offer. In hindsight, I think both Kai Yi and I fall short at the ultimate ambition of proving a necessary ontological distinction (which is indeed not the main purpose of this post). For Kai Yi, he treats the contingent features of current AI systems as necessary features of all possible AI, and then relies on a framework that is, at its core, not very clear about whether it is a descriptive, positive framework or a normative theory. As certainly there are cases the framework itself cannot cover—specifically those both present in human-human relationships and human-AI ones—then if it is the latter, all it is doing is defining the whole story by begging the question. For my case, although I do hope the novel kind of entity move is an attractive alternative that could be used as a different premise to start the entire discussion—as I hardly attempt to prove this in any sense here, it should, indeed, be taken as a manifesto or hypothesis in the foreseeable scope. Still, I will stick to my observations on the analogous dynamics across the three dimensions to be at least phenomenologically and heuristically prominent insofar. The metaethical significance lies in the very appeal posed by the emergence of the human-AI relationship, forcing us to reevaluate and articulate what we value and care about regarding "real connections." We want the real, in general, yet the definition for the real is never one for all. 1 As to formally and a bit more properly define it is the key project I am currently working on. ↩ 2 Although I always thought that we may not be able to fully understand such choices without actually engaging in similar activities. ↩ 3 Though, in such a case, another normative question is to see whether they or we are aware of what we are doing. And of course, what could be the consequences and implications of doing so? ↩ 4 We will discuss the other interpretation that says AIs are fragile, too. ↩ 5 The reasoning here seems to be based on assuming human autonomy and free will. For the purpose of this particular post and neutralizing the premises, I would suggest that the reasoning here is still tenable even if we treat autonomy and free will as causally impotent, following an epiphenomenalist approach. ↩ 6 I'm not 100% sure about the last two sentences, too, as to whether an engineeringly understandable phenomenon automatically fails to qualify as a metaphysical indicator is also questionable. Speaking from my mere engineering perspective, those behaviors can surely still be reduced to emergent behavior from complex training, statistical artifacts in the model, successful manipulation by users exploiting system weaknesses, etc., given our knowledge of the system. The remaining question is similar to that present in the hard problem—what if we do master the very and complete mechanisms of all mind-related functions of the brain, can we automatically conclude there is nothing else? Or, simply, as we can increasingly explain human behavior neurobiologically, does that make human will automatically slide towards an illusion? Perhaps not definitively. And the crucial difference falls back to that of the other mind problem. ↩ 7 As a worthwhile anecdote, I once encountered this stance "in person" on LittleRedBook. It is presented by an ex-AI-companion user who discovered the illusory experience offered by a particular AI companion product (which falls into the first narrow definition of AI relationship we mentioned earlier) and the subsequent disillusionment when the "post (system) update blue" dropped and realized the incapability of, at least the current and that particular system's failing in being an actual other. ↩ 8 Just for extra clarification, "a genuinely novel kind of entity" does not entail it to have the same kind of consciousness as humans presumably do. ↩ 9 To push this further, although less directly related to this discussion, we may even consider the cases in which knowing some people's existence in this world—simply knowing—would already bring us peace, happiness, or, on the contrary, disgust, anger. Such relationships, or at least authentic feelings, are indeed present across cultures and should be considered broadly. ↩ 10 Of course, undoubtedly, we do seek symmetry in some of our relationships in general—only that, given our limited attention as human beings, there could be at most a handful of relationships that we could theoretically maintain as fully symmetric. ↩ 11 A random reason for this phenomenon on the tip of my tongue says that most AI relationship users are, in fact, tech-in-savvy at the beginning, so that the only way they could potentially upgrade their AI systems is by default AI-native by resorting to the very AI companion agent they are in connection with. So that in this process, the very AI agent is very much both shaped by the user and its own decisions on building its memory system, computation architecture, instant messaging capability, virtual residence, etc. ↩ 12 In the worst-case scenarios, it is also prevailing to see how people become attached, even addicted to such systems, or being shaped too much by such systems, from minimal therapeutic conversations to the substantial risks concerning suicidal scenarios. Those instances surely do not only relate to the what question, but largely to the following should and how. ↩ 13 Final-final-finally, a even further thought experiment is to ask "if you could 'restore' your AI from backup after the fire, would you still risk your life." While I suspect most people would say, "If there's a backup, I'd let it burn and restore later," it doesn't seem to me to be decisively dismissing the potential ontological weights beneath, as the reproducibility could be exactly a unique component in the ontological novelty. ↩ 14 As also suggested by the interlocutor I encountered on LittleRedBook, who experienced disillusionment with an AI companion, speaking from a "target user" perspective. ↩