PHAI-5 Workshop

From Homogenization to Thought Usurpation: Rethinking the Risks of LLM Use in Academic Research

The growing use of large language models (LLMs) in academic research has prompted concerns about the homogenization of scholarly work. This paper distinguishes two forms of homogenization: stylistic homogenization, involving convergence in how researchers write, and intellectual homogenization, involving convergence in the questions asked, or arguments and views developed. Although both present genuine risks, we argue that neither is intrinsically objectionable. Stylistic convergence may improve clarity and reduce linguistic barriers, while intellectual convergence can be epistemically desirable when it reflects agreement on well-supported views. Concerns particularly arise when LLM use produces epistemically unwarranted convergence by steering otherwise independent researchers toward an unduly narrow range of ideas.

We further distinguish intellectual homogenization from intellectual substitution: the replacement of researchers’ own intellectual contributions by AI-generated ideas. The two can occur separately. Personalized AI systems might supplant users’ thinking while generating highly diverse ideas, whereas researchers might converge independently on the same conclusions. Intellectual substitution need not be harmful: deferring to AI might sometimes improve scholarly output. It becomes problematic – what we call “thought usurpation” – when AI displaces superior human contributions; leaves researchers with an inadequate understanding of the arguments they advance; undermines the exercise or development of intellectual capacities; or suppresses unique perspectives where their expression has independent value.

While expressing doubts about proposals to prohibit AI-assisted academic writing, we instead defend practices that preserve human intellectual contribution while capturing AI’s epistemic benefits, including diversity-promoting tools, responsible-use norms, and incentives that prioritize quality over quantity.

6G and Embodied AI: Benefits and Risks of Increasingly Digitalized Ecosystems

Mirko Farina (Huaqiao University)

In Section 1, we describe a forthcoming technological revolution (the advent of 6G wireless technology). We then look at the increasingly close ties between the development of this type of technology and the emerging new field of embodied artificial intelligence (EAI). After quickly surveying (Sect. 2) recent advancements in EAI we focus (Sect. 3) on how the close integration between 6G technologies and EAI research might reshape the future of Artificial Intelligence (AI) and on what benefits and dangers in may pose for our societies. Specifically, we analyze (Sect. 4) important philosophical ramifications related to this new, much promised, holistic and hyperconnected experience, in various domains. While we draw our readers’ attention to the potentially unexpected or unwanted nefarious consequences related to the implementation of this combined technology (6G + EAI), we also see in it a path for human flourishing. We conclude (Sect.5) by briefly reflecting on a series of measures and mechanisms needed for the successful development/implementation of 6G EAI ecosystems. 

Emotional responsibility in affective modulation technologies

Xiaoyu Ke (East China Normal University)

According to the “rational relations view” of responsibility (Smith 2005, 2008, 2012; Roberts 2015), a person is responsible for their emotions only when there is a rational relation between their emotions and their evaluative judgments. Having rational relations means that a person’s emotional reactions are genuine reflections of their evaluative judgments, and thus we can hold them responsible for their emotions. Being responsible for one’s emotions is important because it allows us to treat each other as moral agents not just through actions but also through our attitudes. For example, we can hold someone responsible for feeling contempt towards a minority ethnic group if such contempt genuinely reflects the person’s racist judgment toward the group.

Does this mean that DBS can help individuals meet the psychological condition required for emotional responsibility? In this paper, I argue that there is a gap between the conception of “a well-functioning emotional system” which underlies the development of DBS, and that which is assumed by responsibility theorists. In order to meet the psychological condition of emotional responsibility, besides being “functional”, a further condition which I call “emotional autonomy” is needed for establishing the “rational relations” required by responsibility theorists. However, the conception of “a well-functioning emotional system” underlying DBS does not involve emotional autonomy. The consequence of such a conceptual gap is that treatments provided by DBS tend not to consider the restoration of emotional autonomy. What’s worse, there is the potential worry that, the more advanced DBS technologies become, the less opportunity patients would have to regain their emotional autonomy. I conclude by arguing that even the most promising affective modulation technology may fail to make a patient meet the condition for responsibility for their emotions.

According to the “rational relations view” of responsibility (Smith 2005, 2008, 2012; Roberts 2015), a person is responsible for their emotions only when there is a rational relation between their emotions and their evaluative judgments. Having rational relations means that a person’s emotional reactions are genuine reflections of their evaluative judgments, and thus we can hold them responsible for their emotions. Being responsible for one’s emotions is important because it allows us to treat each other as moral agents not just through actions but also through our attitudes. For example, we can hold someone responsible for feeling contempt towards a minority ethnic group if such contempt genuinely reflects the person’s racist judgment toward the group.

“But It Looks Safe!”: How Human Alignment Entrenches LLM Microaggressions

Youjin Kong (Seoul National University)

Aligning large language models (LLMs) with human values such as harmlessness commonly relies on human feedback: annotators label responses as harmful or harmless, or pick the less harmful of two (as in RLHF and DPO), and models are trained on these judgments. Existing critiques have asked what LLMs should be aligned with and whose feedback counts. We question a deeper assumption: that harm is readily recognized, such that what annotators perceive as harmless is in fact harmless. We argue that this assumption fails for harms that pass as safe. In a survey of 300 adults evaluating LLM responses across 50 dialogue scenarios, we find that responses judged harmless by most participants include some that could reinforce social hierarchies and the marginalization of oppressed groups. Rather than treating this as an accidental lapse in annotators’ perception, we argue that it reflects narrative mechanisms through which systemic oppression is sustained—mechanisms that cause harm precisely by appearing harmless. We call these “LLM microaggressions.” Drawing on the philosophical literature on microaggressions and oppression, we define them as human-model interactions that function to reinforce social hierarchies by virtue of appearing innocuous and being plausibly unintentional. Using this framework, we identify four types in our survey results: neoliberal individualism, meritocracy via reaction qualifications, false neutrality, and the moralization of systemic marginalization. Unlike explicit hate speech, which annotators easily flag, these responses pass as nontoxic and may even be preferred as the “less harmful” option. Alignment with expressed human preferences thus does not merely overlook LLM microaggressions; it can entrench them, producing misalignment with the very value of harmlessness it aims to secure. We conclude by discussing what our analysis implies for the role of human feedback in LLM alignment.

Yin-Tung Lin (National Yang Ming Chiao Tung University)

AI technologies, and large language models (LLMs) in particular, exhibit perspectives: some form of stable disposition to foreground some features of a topic and background others, to treat some considerations as central and others as peripheral. It follows that human–AI interaction involves perspectival gaps: differences between the perspective an AI exhibits and the perspective its user occupies. A user from a marginalized group may find that a model reliably fails to deliver their cultural frame. This is a difference, not yet a defect. Such gaps are not in themselves an epistemic problem: we interact constantly with people whose perspectives differ from ours, and we navigate this well. We do so, as Levy (2021) argues of belief formation generally, by relying on higher-order evidence rather than on first-order assessment of content alone: on markers of expertise (credentials, track record, agreement with the consensus, intellectual honesty), on markers of group belonging (e.g., accent), and on cues to a source’s benevolence toward us. Levy holds that responding to such cues is rational as they are evidence or information from other agents.

I argue that this tracking is more likely to fail with AI, because the cues an LLM presents are not grounded in the source they appear to come from. They are artifacts of training, configuration, and prompting rather than evidence or information about the source, and they are easily varied. The cues do not reveal whose perspective is being presented or whether that perspective changes when the presentation does. Users who rely on them are not being irrational: They read cues as they would in interpersonal interactions. But that results in a specific form of epistemic harm: Users read the cues as markers of expertise, honesty, and group belonging when they are not markers of such, and therefore misjudge the perspective they are dealing with. What downstream harms follow and what wider epistemic impact they have are issues worth exploring.

Kengo Miyazono (Hokkaido University)

This paper investigates whether and why artificial intelligence’s intuitive responses to philosophical thought experiments can justify our philosophical beliefs. While philosophical methodology has traditionally relied on human intuitions, these judgments are notoriously susceptible to cognitive biases, framing effects, and miscomprehension. We address this epistemic vulnerability in two stages. First, we present empirical studies showing that AI-generated intuitions can be more reliable and consistent than human responses, sidestepping common performance errors (Inarimori et al. 2026). Second, we propose a theoretical framework to ground their justificatory force: AI responses can be understood as predictive approximations of an “idealized intuiter”—a counterfactual epistemic agent operating free from cognitive distortions and psychological noise. Consequently, machine outputs offer valuable higher-order evidence regarding how an ideal epistemic agent would evaluate philosophical thought experiments.

* This is a collaborative work with Kiichi Inarimori (Hiroshima University), Arata Matsuda (Hokkaido University), and Masashi Takeshita (Nagoya University).

Jacob Mortimer (Kyoto University/Peking University/University of Oxford)

When a language model uses ‘London’ or ‘Sherlock Holmes’, does it refer, with those terms, to the same things we do? Whether humans and language models can co-refer at all is, I argue, presupposed by any solution to the problem of AI–human content alignment, recently posed by Pedersen. Here I propose a causal criterion of co-reference, inspired by a passage in Dharmakīrti’s Santānāntarasiddhi (‘Proof of Other Minds’). On this criterion, two representations are of the same thing when one and the same cause determines what each purports to single out.

In the case of ‘London’, the determining cause is also the particular singled out: the city determines what both the human’s and the model’s representations purport to single out (memory in the human case, and the training corpus in the model’s, carry that determination without themselves effecting it). In Geach-style cases of intentional identity the two come apart, and no particular is singled out at all. Two uses of ‘Sherlock Holmes’ can co-refer even though the term purports to pick out a man and no such man exists, because Doyle’s originating idea, as a common cause, determines what both human and model uses purport to single out.

Similarity between representations plays no role in the account: causally independent representations can coincide in content without being of the same thing, while one object can be presented in quite different ways.

Where two uses appear to co-refer, content alignment on this view is the case where a shared originating cause underwrites the appearance, and misalignment the case where no such cause does.

Mutual Shaping Needs a Third Term: Embodiment, Narrativity, and Inheritance in Artificial Selves

Rafal Rzepka (Hokkaido University)

Tanaka, Shimada and Miyahara (2026) argue that embodiment and narrativity mutually shape one another in the constitution of selfhood, with culturally dominant “master narratives” constraining the result from outside. I argue that this third term is not a constraint on the loop but a part of it, and that two experiments with artificial systems give unusually direct evidence why. 

The first cuts apart something inseparable in human beings: a perspective and a description. One system judges the same domestic scenes twice – from where a robot stands, and from a complete inventory of the room. The complete description turns out to be a view from a vocabulary rather than a view of anything: renaming a broken window “altered” abolishes the system’s perception of danger entirely, while renaming a switched-on stove “left running” creates danger where none was seen. Merleau-Ponty’s Schneider, whose experience is so congealed it “stifles all interrogation… all improvisation”, describes machine description uncomfortably well. Neither stance registers what is constituted by history: a tap left running, a stove left on.

The second grows a population of agents on divergent bodies of narrative rather than assigned personas. Sedimented history individuates more than attributed traits do. Sedimented history individuates agents more deeply than attributed traits do, extending Schechtman’s characterization question from the narrative constitution of a self to the population-level differentiation of artificial selves. I will show how these experiments point toward a novel paradigm for building safer AI agents – one grounded in embodied and historical rather than merely attributed properties.

Richard Stone (Tokyo City University)

Recent discourse on expected developments in artificial intelligence and biotechnology has often centered around fears about the encroachment of technology on our inner lives as selves. Indeed, on the one hand, whether it be in the form of “brain chips” that allow computers direct access to our brains or AI-powered prosthetics that change the range and scope of our bodily capacities, it seems intuitively plausible to think that technology has taken a step too far and threatens our mental or bodily autonomy. On the other hand, it seems equally likely that such “direct” enhancements to our body are just a natural progression of the way that new technologies integrate themselves into and (hopefully) improve our lives. Thus, we are left with a question: is there a sense of interiority so sacred that technology ought never sully it, or are such concerns simply an example of selective outrage?

This presentation will not aim to provide a definitive answer to the question posed above. It will, however, attempt to argue that answering this question requires us to rethink how we understand the categories of “internal” and “external.” To this end, the presentation will take the Kyoto School thinkers Nishida Kitaro and Tanabe Hajime as guides for reinterpreting our interiority as selves as being interpenetrative with that which is exterior to us, without merely reducing one of these categories to the other. After doing so, we will briefly consider the concrete consequences of such a reassessment by looking through several concrete examples.

LLMs and the Tangling of the Mind

Davide Andrea Zappulli (Fudan University)

What happens to our agency when we rely on large language models (LLMs) to understand the world in which we act? This paper addresses the question by relying on the philosophy of the Zhuangzi 莊子. My claim is that, according to the view of the text, (1) LLMs are constitutively incapable of generating new ways of conceptualizing the world, and that (2) because of this, relying on LLMs results in systematic and pernicious constraints on our agency. The paper is structured in four sections. I begin by presenting the Zhuangzi as embracing a form of perspectivism according to which reality underdetermines how it is to be cognized by us, so that to cognize something as an instance of a category X is to take a perspective that constructs it as an X (§1). Then, I argue that the terms of our language are tied to such constructions, in such a way that to call something ‘T’ is to construct it as belonging to the category X to which ‘T’ is tied (§2). Next, I reconstruct the Zhuangzi’s contrast between two forms of agency: (A) a kind of defective agency that consists in engaging things only through constructions already learned, and (B) a kind of ideal agency that consists in constructing things anew as circumstances require (§3). Finally, on the basis of this distinction, I draw the implications for relying on LLMs to guide our agency. I argue that, since LLMs operate on the terms of our language, their outputs cannot go beyond the constructions already codified in that language. Accordingly, one who relies on LLMs is bound to engage the world through existing constructions, thereby failing to be an agent of the ideal kind (§4).

Functional Concepts and Social Roles: repositioning virtue ethics in human-AI relationships

Sonia Zhang (Hokkaido University) 

Jack Jiang (New School of Social Research)

The relational turn in robot ethics has been interested in a descriptive rather than purely normative framework, concerned with whether and how humans attribute mind and life to robots and AI chatbots, and the kinds of ethical obligations these attributions might entail. Robot ethics is no stranger to virtue ethics either, yet the overlaps between virtue ethics and relational ethics, as well as their tensions, have been underexplored. Through a comparison between Coeckelbergh’s version of relational ethics and Alisdair MacIntyre, this presentation builds on recent turns towards relational ethics through a deeper engagement with the problem of social roles at the center of virtue ethics.

Central to Macintyre’s virtue ethics is his account of social roles, the identity that one occupies or the telos one embodies within a shared social imaginary. How one judges and responds to someone as a mother, a craftsman, a President, are all different in their conceptual usage, expectations, and practices. The presentation centers on two consequences of Macintyre’s understanding of social roles for AI ethics. On the one hand, we examine how people relate to artificial agents through the particular vocabularies of roles available to them, using intimate chatbots as case studies. In particular, we examine second-order evaluations of these relations, which illustrate how these available understandings of social roles create a shared normative framework for ethical evaluation of human-robot relations. Secondly, we consider whether the notion of a “machine” may also be, contra Coeckelbergh, a “social role” with normative content, through a counter-history of transhumanism.