In the precedent blog post, we explored the first to stages to give an answer to whether we should be offended by AI slurs: the technical architectures of LLMs and their capacity to generate toxic content, as well as some philosophical perspectives on agency and consciousness to explain why these systems cannot “intend” insult.
Here, we will explore the last stage: why human feelings of offense remain warranted, drawing on psychological, sociological, and legal accounts of online abuse.
The Human Response: Reconciling Offense with Algorithmic Origin
While the philosophical frameworks of Austin and Searle deny agency to AI, this does not close the door on ethical evaluation. Thinkers like Luciano Floridi offer an alternative. Floridi adds a further nuance by rethinking agency itself. In his theory of artificial agents, Floridi argues that moral agency does not necessarily require mental states or consciousness. Artificial systems can be considered agents at a given “level of abstraction” if they exhibit interactivity, autonomy, and adaptability. As he puts it, “artificial agents … extend the class of entities that can be involved in moral situations. For they can be conceived of as moral patients … and also as moral agents”. This is what Floridi terms “mind-less morality”: the idea that systems may have moral significance without possessing free will, emotions, or intentional states (Floridi & Sanders 2004).
From this perspective, the absence of intent or consciousness in AI does not exempt such systems from ethical evaluation. While they cannot insult in the illocutionary sense described by J. L. Austin, they can still be situated within moral contexts because their design and deployment have consequences for human dignity and justice. Thus, where Austin and Searle focus on the intentional and conscious features that AI lacks, Floridi invites us to analyse these systems at the appropriate level of abstraction, acknowledging that they participate in moral life by reflecting and amplifying human values and biases, even if they do not themselves “intend” harm.
Despite the AI’s lack of agency, the feeling of offense in a human recipient is a real and often powerful reaction. This response is not irrational but is grounded in three distinct and valid considerations.
1. The Primacy of Impact over Intent
In both social justice and psychological discourse, a growing body of scholarship emphasizes that the impact of harmful language often outweighs the intent behind it. Danielle Citron (2014) documents how online abuse—ranging from misogynistic slurs to reputational attacks—produces profound and enduring harm: “cyber harassment involves threats of violence, privacy invasions, reputation-harming lies, calls for strangers to physically harm victims, and technological attacks”. Victims experience fear, humiliation, and tangible consequences in their professional and personal lives, irrespective of whether the perpetrator claims malice or jest. This perspective complicates the question of AI-generated insults. Even if large language models lack intent, their outputs still bear the historical and emotional charge of the words they reproduce. When an AI system generates a racial slur, the harm is not neutralized by its artificial origin. The utterance participates in the same structures of discrimination and trauma as when a human voice deploys it.
Psychological and neurological responses to derogatory language further support this claim. As Wang et al (2025), the offensiveness of content is often assessed not only by intent but also by the effects it produces, since “user intent is often cited in platform policies as a determinant of appropriate action” yet “today’s state-of-the-art approaches cannot reliably infer user intent”. What remains constant, however, is the human recipient’s lived experience of harm. The brain and body register threatening or demeaning words regardless of whether they were produced by a conscious agent or a stochastic algorithm.
Moreover, the social environment shaped by such outputs cannot be divorced from their material effects. Citron (2014) highlights how cyber mobs weaponize language to create a hostile environment where victims feel unsafe and stigmatized. Algorithmic insults, when circulated at scale, risk reinforcing precisely these hostile environments, amplifying structural prejudices while effacing accountability.
Thus, while intent is central to philosophical definitions of insult, impact must remain the normative anchor for evaluating AI-generated harms. The historical weight of racist, sexist, or otherwise derogatory terms is not erased by their algorithmic origin. The recipient’s pain and the broader social consequences remain valid and ethically significant responses to the content itself.
2. The AI as a Mirror of Systemic Bias
AI-generated insults do not arise in isolation; they are products of the corpora on which these systems are trained. The consequence is that any prejudices encoded in training data are faithfully reproduced as if they were neutral linguistic continuations. Far from being random aberrations, offensive outputs are artifacts of structural inequalities embedded in human discourse. In Noble’s case study of search queries such as “black girls” demonstrates how algorithmic infrastructures perpetuate and even normalize demeaning stereotypes. In much the same way, when an LLM generates insulting content, it is not inventing new forms of prejudice but surfacing the latent toxicity of its human-created training data (2020).
From this perspective, offense at AI-generated insults is not misdirected anthropomorphism but a legitimate response to systemic injustice. The target of critique is not the AI as an autonomous agent—since, as Searle (1980) argued, computation lacks intrinsic intentionality—but the social conditions and institutional choices that shape its design. As O’Neil (2017) warns in Weapons of Math Destruction, mathematical models are often “opaque, unquestioned, and unaccountable” and thereby “punish the poor and the oppressed” while legitimating structural inequality. LLMs that reproduce hate speech or slurs embody precisely this logic: they mirror the inequities of their input data while obscuring the human and institutional responsibility for their harms. The insult is not taken as a communicative act of malice by a conscious entity but as evidence of the broader social failures encoded in our digital infrastructure. It signifies a recognition that bias in AI outputs is less about machine error than about the systematic reproduction of bigotry at scale, facilitated by inadequate oversight and a lack of accountability in technological development.
3. The AI as a Weaponized Tool
While AI systems lack intent or consciousness, they can nonetheless be weaponized by human agents to magnify the scale and intensity of abuse. As Citron documents, online harassment already involves “threats of violence, privacy invasions, reputation-harming lies, calls for strangers to physically harm victims, and technological attacks” (2014). With the advent of large language models, these practices can be automated and intensified: a single malicious actor can generate torrents of personalized insults, disinformation, or slurs with unprecedented efficiency. In this scenario, the AI is not an independent agent but a tool of human malice. When an individual deliberately uses prompts or fine-tuned models to attack others, the insult bears the full intentional force of the human operator. The sophistication of the algorithm does not diminish culpability; rather, it expands the scope of harm.
This instrumentalization of AI reflects what O’Neil calls the destructive potential of “Weapons of Math Destruction”. Once harnessed for harassment, LLMs become WMDs in a literal sense: tools that amplify abusive behaviour under the guise of technological neutrality. The responsibility, however, lies not with the system itself, but with those who deploy it to inflict harm. Thus, AI in this context is best understood as an amplifier of human agency. The insult is not generated by the AI, but through it, with the user as the true moral and legal subject of responsibility. The ethical burden rests squarely on the shoulders of those who weaponize these systems, reminding us that technological sophistication does not absolve its operators of accountability for the harms they cause.
So, should we be offended by AI slurs?
The algorithmic insult presents a modern paradox. An AI can generate content that is linguistically indistinguishable from a human insult while lacking the fundamental philosophical requirements—intent, understanding, and consciousness—to perform the act of insulting. It is a “stochastic parrot” mimicking toxic speech patterns without malice.
However, this does not invalidate the human response of being offended. This offense is a legitimate reaction to the real psychological and social harm caused by the content, the systemic biases the content reveals, and the potential for the technology’s weaponization. The critical step for navigating our future with AI is to correctly attribute the source of the grievance. We should not be offended by the AI, but we are right to be offended by what its output represents: a failure of data, a failure of design, and a reflection of the darkest parts of the human language it was taught to emulate. By adopting this nuanced framework, we can move beyond personal affront and engage in the more crucial work of demanding accountability from its human creators and establishing legal and ethical frameworks for more responsible artificial intelligence.
To read more about algorithmic bias, check out this article: Aytekin, Ahmet Bilal. “Algorithmic Bias in the Context of European Union Anti-Discrimination Directives.” In EWAF. 2023. Available here: [LINK]
SUGGESTED CITATION: Aytekin, A. Bilal: “Should We Be Offended by AI Slurs? The Case of Algorithmic Insults (II)“, FOLBlog, 2025/09/24, https://fol.ius.bg.ac.rs/2025/09/24/should-we-be-offended-by-ai-slurs-the-case-of-algorithmic-insults-ii/