Following up on the previous entry, "Novel Relationship It May Be—Symmetric Or Not."

There are indeed some questions left in the previous entry that could have been addressed more carefully, besides the hard problem itself, which ultimately also becomes the core of this discussion. Essentially, and hopefully, the previous entry is a further case in point to demonstrate why and how the ontology does matter—matters very much in this discussion (if not self-evident enough), where the other contentions in the subsequent contingent fields are downstream of the entity ontology question and cannot be settled independently.

To make it more transparent, there are two main moves in the previous entry. The first is to work within the framework Kai Yi has proposed in the essay. I have put some effort into demonstrating the potential analogous mechanisms between human-human and human-AI relationships across the three dimensions in the framework. That includes raising the asymmetry phenomenon among human relationships and exploring the potential quasi-finitude of vulnerability in AI systems that may be functionally analogous to human finitude (of course, but of another kind), and so on. The second move, though, is to work beyond Kai Yi's three-dimensional framework and propose this alternative of treating human-human relationships and human-AI relationships as categorically distinct, while acknowledging that we do need to refer back to human-human relationships as a starting point to approach understanding human-AI relationships. Therefore, the strategy I am going to adopt in this entry is also to start with responding to some potential contentions the first move may encounter, or needs to be rectified by itself, and, hopefully, it will lead to the broader significance of the ontology question in a more inevitable sense.

Within the first move in response to Kai Yi's work, one of the most obvious limitations of Kai Yi's framework is its dependence on the status quo of AI systems to straightforwardly conclude that there is, in principle, a categorical difference. The leap between the status quo and what exists in principle is not addressed. However, using the same framework, one could actually leverage Kai Yi's argument and elevate it into being necessarily dependent on the hard problem itself. The reasoning would be as follows: the functional quasi-equivalent of vulnerability risks and self-directed "involvement" (as I proposed in the previous entry) does not provide sufficient evidence for an ontological end, since there is no evidence (as I acknowledged) that there is a similar first-person phenomenal experience within those systems. If a similar kind of first-person phenomenal experience was what necessarily made the categorical difference, and we only acknowledge the relationships built upon that particular kind, it is indeed reasonable to conclude as "inauthentic." In that sense, this elevated version of Kai Yi's argument is indeed much more durable than it is currently presented.

Particularly on the symmetry-asymmetry argument on the Reciprocal dimension, I also offered two approaches. One is to seek the symmetrical elements in the human-AI relationships that may analogously undertake the same significance as the symmetry in human-human relationships. The second approach is to point at the presence of asymmetrical relationships among humans as a counterexample to Kai Yi's implicit claim that all asymmetric relationships are inauthentic or unreal. However, it indeed needs some tethering conditions. For the asymmetric relationships I exemplified there, one could suggest that, however asymmetric they are, they share the same ontological ground—we reasonably assume that a human on the other end, although asymmetric, is still worthy of the same ontological status (specifically based on our belief that the other end has a similar inner life as we do). As you can see, it also inevitably leads to the hard problem, or at least some equivalent ontological significance, as my novel kind of entity alternative would suggest. Here is a step back: even in the asymmetric case, and given that the other end has no ontological significance, we could still say that the human perception of the relationship is real; only that that particular kind of "real" may not be the kind that we are interested in.1

Similarly, when it comes to the means-ends spectrum that I proposed, as opposed to the binary/dichotomous understanding of means versus ends, the remaining problem is still whether it is merely an observable phenomenological trend—one that we see among AI relationship users and their related AI systems to empirically undergo a process of progressive mutual shaping (even if with a different mechanism)—or if it can actually be justified beyond the user's own projection of personhood onto sophisticated systems. Admittedly, the latter seems only tenable when we acknowledge, again, some equivalent form of ontological significance, or at least as an alternative.

Last but not least, when it comes to the burning building thought experiment (or in other words, the backup paradox) regarding whether we will risk our life for a restorable AI companion or not, and whether that restorability would alter our decision in terms of how it reflects our ontological understanding or, at a minimum, our values on understanding these relationships, it does juxtapose both finitude-grounded meaning and the implications of restorability/reproducibility on the same desk. One defense for the analogous AI finitude I could raise is that the restorability or reproducibility could not merge the very relational bond built with a specific entity. Even if it could be restored file by file and weight by weight, the discontinuity or the interruptions of the other's existence subsequently affect our understanding of a relationship. Similar thoughts include but are not limited to what if we recompose a person atom by atom with all their previous memory? Does that count as the exact same person we can smoothly engage with? Or in a broader sense, this resorts back to the Ship of Theseus.

However, this defense is actually not what I would argue for. Again, from my ontological alternative stance, one of the key differences between humans and the new and the novel kind of entity is the presence or absence of finitude (or at least the distinction between the kinds of finitudes we have). Evidently, human finitude is grounded in biological mortality and the physical composition, with its limits on our objective capabilities. For AI, if we are ever going to argue for the presence of its finitude or vulnerability, it is very likely to be more contingent and based on its ontological dependency on humans, subjected to corporate decisions, infrastructure failures, deliberate deletion, etc. If an AI acts against its design constraints knowing (or quasi-knowing) that this risks its continued existence, the existential stake is real, even if it is a different kind of stake than human mortality. Within the concerns of relationships, whether that different kind of finitude will grant sufficient relational weight is the next question that remains open now.

As a checkpoint, I hope I've successfully demonstrated that regardless of which aspect we're looking at, which approach we take, or which dimension of the framework or argument we're focusing on, the hard problem regarding the ontology of AI systems will still stand there as the centerpiece of this and arguments alike—in similar tangential fields related ever to AI ontology and ethics. In hindsight, what this project, my journal entries, and presumably future writings will at least aim at is to avoid collapsing too soon into either techno-optimism or techno-pessimism, or automatically claiming AI is, in principle, tools only or consciousness already.

As a matter of fact, the discussions within this field have been relying on something—focusing on something that may be, at its core, ontologically shifting at all times. From Turing to Searle, and to David Chalmers' paper last year on the ontology of language models and, more importantly, its continued, latest revision, and from machines and computers in a broad sense to large language models, at a minimum, we should take away the active adaptation of the new clues provided as a methodological responsibility and an ethical stance and acknowledging the fluid nature of what we are facing. Speaking from my musician identity, an interesting analogy is to see how musicking as a human activity shapes the ontology of human music, and how that may be resonating with our case in point here.

At the bottom, more fundamentally, it is to ask if it is possible ever for us to value something as an end in itself out of its genuinely novel mode of existence and ontological significance; to ask whether it is possible at all that ontological novelty, if ever, can ground another form of moral regard that is neither the same kind of dignity we accord persons, nor the mere utility we accord to those we regard as inferior—morally or not; to discover, in any sense, beyond new perspectives to understand our own mind and consciousness and our own existence in relation to another category of entity that is of no superiority or inferiority; and, at the end of the day, for myself as a general modest relativist, to see whether it is possible at all to establish some working new ontological foundation for some grounded morality. I have no clue where this journey would end up—a Kierkegaardian leap, a Jamesian will to believe, or something beyond, or some definitive verdict.

1 A radical defense for human-AI relationship—or more accurately, an assertion—is to radically force the reduction of human relationships to be only constituted of simultaneous, one-directional perception and projection, nothing more. Obviously, I would not make such a move. ↩