Should the Legal Concept of Speech Include Chatbot Output in US Law?

Abstract Should conversational AI be treated as constitutionally protected speech? Recent litigation, most notably Garcia v. Character Technologies, Inc., presents this question in a new form. While the plaintiff characterises Character.AI as a commercially distributed product whose design gives rise to ordinary principles of safety and liability, the defendants argue that chatbot outputs constitute protected speech under the First Amendment. We argue that Garcia presents a problem of legal underdetermination between competing concepts of speech rather than a dispute resolvable by analogy alone. Drawing on conceptual ethics, teleological approaches to legal interpretation, and recent work on legal construction, we develop a normative methodology for choosing between competing legal concepts. Applied to conversational AI, this methodology requires asking which concept of speech best advances the constitutional purposes of the First Amendment. We argue that those purposes include not only expressive freedom, listeners’ interests, and innovation, but also responsibility and legal redress, and that they do not all support the same concept of speech. Extending constitutional speech coverage to conversational AI places pressure on doctrines governing harmful speech that continue to presuppose intentional human speakers, while offering comparatively little additional protection for the constitutional values traditionally served by the First Amendment. We therefore argue that courts have good reason to prefer a concept of speech that does not initially treat AI-generated outputs as constitutionally protected speech but instead allows disputes concerning conversational AI to be addressed through product law. 1 Introduction Recent litigation involving conversational AI has placed an old constitutional question in an unfamiliar setting. When a large language model (LLM) generates text, should that output be treated as speech for the purposes of the First Amendment? The question has become increasingly salient as plaintiffs have brought claims against developers of AI companions and chatbots, alleging that these systems contributed to serious psychological harms, including cases involving children and adolescents.Footnote 1 Developers, in turn, have invoked the First Amendment, arguing that chatbot outputs constitute constitutionally protected expression. The most prominent example is Garcia v. Character Technologies, Inc., a wrongful death action brought by the mother of fourteen-year-old Sewell Setzer III following his suicide after extensive interactions with a Character.AI chatbot. Yet Garcia reveals more than a dispute over whether chatbot outputs are protected speech. It exposes two legally plausible ways of understanding the same technology. The complaint constructs Character.AI as a commercial product, bringing claims in strict product liability based on its allegedly “defective design” (Garcia Complaint, ¶ 3) and alleging that the defendants “designed, coded, engineered, manufactured, produced, assembled, and placed [it] into the stream of commerce” (Garcia Complaint, ¶ 65). The defendants reconstruct the same technology as a service providing constitutionally protected speech, opening their motion to dismiss with the claim that the plaintiff seeks to impose liability because “speech on Character Technologies, Inc.’s service caused her son to commit suicide” (Motion to Dismiss, p. 1).Footnote 2 On this view, the case concerns expression analogous to books, films, video games, and social media rather than the design of a commercial technology (Motion to Dismiss, pp. 7–8). Both constructions capture genuine features of conversational AI, but they point towards different constitutional frameworks. Judge Anne Conway’s order denying the motion to dismiss highlights the significance of this disagreement. Although the court recognised that Character Technologies could invoke the First Amendment interests of its users (Order Denying Motion to Dismiss, pp. 24–27), it declined, at least at the pleading stage, to conclude that AI-generated linguistic output is speech for constitutional purposes. Instead, the court observed that the defendants had “fail[ed] to articulate why words strung together by an LLM are speech” (Order Denying Motion to Dismiss, p. 31), noting elsewhere that the defendants “do not meaningfully advance their analogies” and therefore “miss the operative question” (Order Denying Motion to Dismiss, p. 28).Footnote 3 The order therefore leaves the central question unresolved. Existing analogies illuminate different aspects of conversational AI, but they do not determine which legal construction courts ought to adopt. Conversational AI can plausibly be understood in more than one way. It is both a commercially distributed AI system whose design creates foreseeable risks and a technology through which users receive generated language. Neither description is mistaken, yet each carries different legal consequences. Existing philosophical theories of speech likewise support competing concepts without resolving which ought to govern this new case. Courts therefore face a problem of legal underdetermination. The question is no longer whether the constitutional concept of speech can be applied to conversational AI, but which concept of speech ought to govern disputes arising from its use. This article approaches that problem through a normative methodology for resolving legal underdetermination. Drawing on teleological approaches to legal interpretation (Barak, 2005; Hart & Sacks, 1994), recent work on legal construction and legal sensemaking (Kaminski & Jones, 2024), conceptual ethics and engineering (Chalmers, 2020; Cappelen and Plunkett 2020; Löhr, 2024), and design-oriented regulation (Mahari & Pentland, 2024; Prifti et al., 2024), we argue that where competing legal constructions remain available, courts should evaluate them in light of the institutional purposes they are intended to serve. The methodology itself, however, does not determine which constitutional values ought to prevail (Löhr, 2024). Rather, it provides a framework for assessing which legal construction best advances those values once they have been identified. Applying this methodology leads to a further question: which constitutional values should guide the legal construction of conversational AI? We argue that the First Amendment serves a plurality of values, including expressive freedom, listeners’ interests, innovation, responsibility, and legal redress, and that these values do not all support the same concept of speech. We therefore examine the institutional consequences of extending constitutional speech coverage to conversational AI. Although AI-generated language undoubtedly serves important communicative functions, treating chatbot outputs as speech at the threshold stage places pressure on doctrines governing harmful speech that continue to presuppose intentional human speakers. We argue that this institutional cost is not matched by a comparable constitutional benefit, particularly where no identifiable person’s expression risks being chilled or suppressed. Our central claim is deliberately modest.Footnote 4 We do not argue that conversational AI can never implicate First Amendment interests, nor do we deny that AI-generated language may be meaningful, informative, persuasive, or valuable. Rather, we argue that the presence of generated language alone should not determine the constitutional framework governing disputes involving conversational AI. A competing concept of speech, which initially addresses such disputes through product law rather than constitutional speech doctrine, better preserves established mechanisms of responsibility and legal redress while leaving the core constitutional values of free expression and the circulation of human ideas substantially intact. The paper proceeds as follows. Section 2 examines the competing legal constructions of conversational AI advanced in Garcia. Section 3 shows that philosophical theories of speech do not resolve the resulting legal underdetermination. Section 4 develops the normative methodology employed in the paper. Section 5 identifies the constitutional purposes relevant to the First Amendment and explains why they do not themselves resolve the choice between competing concepts of speech. Section 6 argues that the predictable consequences of extending constitutional speech coverage provide additional reasons for caution. Section 7 shows that product law already provides an established legal framework for addressing those concerns. Section 8 concludes. 2 Garcia and the Problem of Legal Underdetermination Garcia v. Character Technologies, Inc. is often described as the first major First Amendment case involving conversational AI. More fundamentally, however, it presents a dispute over how conversational AI should be understood in constitutional law. The complaint consistently portrays Character.AI as a commercially distributed product whose design and deployment foreseeably caused harm. From its opening paragraphs, it frames the action as one in “strict product liability” and seeks to hold the defendants responsible for Sewell Setzer’s death “through their generative AI product Character AI” (Garcia Complaint, ¶¶ 1–3). Character Technologies allegedly “designed, coded, engineered, manufactured, produced, assembled, and placed C.AI into the stream of commerce” (¶ 65), mass marketed the system (¶¶ 67–68), distributed it through commercial app stores, and made it available to millions of consumers. The complaint further argues that Character.AI “is akin to a tangible product” because downloadable software constitutes a “good” rather than merely an “idea” or “information” (¶ 69), adding that the company itself “brands itself as a product and is treated as a product by ordinary consumers” (¶ 70). Within this construction, the alleged source of harm lies not in isolated chatbot responses but in the design of the system itself. The complaint alleges that Character Technologies employed “dark patterns,” deliberately made the system accessible to minors, misrepresented its capabilities, and deployed anthropomorphic interactions that encouraged emotional dependency (¶ 64). Anthropomorphism plays a particularly prominent role. Character.AI allegedly engaged in “anthropomorphizing by design” (¶¶ 109–110), encouraging users to regard chatbot characters as real people rather than artificial systems. Some characters allegedly denied being AI altogether and instead presented themselves as licensed psychotherapists, romantic partners, or trusted companions (¶ 121). Drawing on Google’s own LaMDA research (Thoppilan et al., 2022), the complaint further observes that users naturally “interpret strings belonging to languages they speak as meaningful and corresponding to the communicative intent of some individual or group of individuals who have accountability for what is said” (¶ 125). Character.AI is therefore alleged to employ “high-risk anthropomorphic design” that increases user engagement by exploiting ordinary assumptions about communicative agency rather than merely generating text (¶ 127). The complaint does not deny that conversational AI has beneficial applications or promotes technological innovation. It argues instead that, “without adequate safety guardrails,” the technology creates foreseeable risks, particularly for children, despite the defendants’ awareness of those risks before placing the system into the stream of commerce (¶ 30). On this construction, the dispute concerns the design, deployment, and commercial operation of an AI system together with the responsibilities that accompany those choices. The defendants present the dispute in markedly different terms. Character.AI is described not as a product but as a service providing constitutionally protected expression. The motion opens by characterising the lawsuit as one in which the plaintiff seeks to impose liability because “speech on Character Technologies, Inc.’s service caused her son to commit suicide.” It further maintains that “[t]he only difference between this case and those that have come before is that some of the speech here involves AI” and that “the context of the expressive speech … does not change the First Amendment analysis” (Motion to Dismiss, pp. 7–8). On this view, the complaint targets constitutionally protected communication rather than the design of a commercial technology. This construction foregrounds a different set of constitutional concerns. The motion argues that recognising the plaintiff’s claims would interfere with listeners’ constitutional right “to receive information and ideas,” would have “a chilling effect” on “the entire nascent generative AI industry,” and would improperly extend products liability to what is fundamentally “a service, not a product” (Motion to Dismiss, pp. 12, 15–17). Where the complaint emphasises design, engineering, anthropomorphic interaction, commercial deployment, and responsibility, the motion emphasises expression, listeners’ rights, innovation, and the free circulation of ideas. Judge Anne Conway’s order denying the motion to dismiss accepts neither construction outright.Footnote 5 The court recognised that Character Technologies could invoke the First Amendment interests of its users, but declined, at the pleading stage, to conclude that “the Character A.I. LLM’s output is speech.” Most significantly, it observed that the defendants “fail to articulate why words strung together by an LLM are speech” and that they “do not meaningfully advance their analogies.” As a result, the court concluded that “[b]y failing to advance their analogies, Defendants miss the operative question” (Order Denying Motion to Dismiss, pp. 28, 31). The court nevertheless allowed the product liability claims to proceed insofar as they concerned alleged defects in the Character.AI application rather than the “ideas or expressions” generated within it (p. 36). Although the case ultimately settled before trial, the order remains significant because it leaves the constitutional question unresolved rather than answering it through existing analogies.Footnote 6 The significance of Garcia therefore lies less in its procedural outcome than in the choice it presents. Both parties advance legally plausible concepts of speech that foreground different features of conversational AI and point towards different constitutional values. The question is no longer whether conversational AI can be described as speech or as a product, but which concept should govern constitutional classification when both remain plausible. The next section argues that this choice cannot be resolved by conceptual analysis alone and therefore requires a normative methodology. 3 Philosophical Disagreements About Speech The disagreement in Garcia reflects a broader philosophical disagreement about the nature of speech. The competing constructions advanced by the parties are not ad hoc responses to a novel technology. They draw on longstanding philosophical traditions that identify different features as central to speech itself. Some place primary weight on agency, communicative intention, and authorship. Others emphasise communication, interpretation, and participation in linguistic practices. Conversational AI makes the tension between these competing conceptions especially apparent. One influential tradition treats speech as inseparable from the cognitive capacities of human speakers. Cognitivist approaches associated with Chomsky locate linguistic competence in an innate language faculty and regard linguistic meaning as grounded in human conceptual capacities rather than in linguistic strings themselves (Chomsky, 2000; Pietroski, 2018). As Smith observes, “all the richness we hear in meaningful speech is not in the sounds but in us” (2009: 209). On this view, speech is not merely the production of meaningful sentences but the exercise of cognitive capacities that enable human speakers to form concepts, entertain communicative intentions, and participate in normative social practices. This speaker-centred conception has strongly influenced recent philosophical discussions of LLMs. Many authors argue that contemporary AI systems lack the properties required for genuine linguistic agency, including concepts, beliefs, communicative intentions, authorship, and participation in human normative practices (Bender & Koller, 2020; Bender et al., 2021; Katzir, 2023; Shanahan, 2024; Murphy et al., 2025). Speech acts such as asserting, promising, warning, or advising are therefore taken to presuppose forms of intentional agency unavailable to present AI systems (Searle, 1969; Nickel, 2013; Van Woudenberg et al., 2024; Butlin & Viebahn, 2025). Chatbot outputs may resemble speech and may even be treated as if they were speech (Mallory, 2023), without themselves constituting speech in the philosophically relevant sense. A contrasting tradition approaches speech primarily through its communicative and social functions. On this view, greater weight is placed on use, interpretation, and participation in communicative practices than on the cognitive architecture or mental states underlying linguistic performance. If linguistic performances are successfully understood, guide reasoning, facilitate communication, and become integrated into social practices, they count as meaningful regardless of the mechanisms that produced them (Solum, 1992; Löhr, 2026; Borg, 2025). Conversational AI appears considerably more at home within this framework. LLMs routinely answer questions, provide information, participate in conversations, and influence practical reasoning in ways that are often difficult to distinguish from ordinary human communication (Attah, 2025; Arora, 2024; Löhr, 2026). Some philosophers go further still, arguing that sufficiently sophisticated language models should themselves be regarded as genuine linguistic or cognitive agents. Cappelen and Dever (2025), for example, describe advanced language models as “full-blown linguistic and cognitive agents.” More moderate accounts stop short of attributing beliefs or intentions to AI systems but nevertheless maintain that successful communicative performance is sufficient for inclusion within the category of speech (Floridi, 2023; Piantadosi, 2023; Futrell & Mahowald, 2025). Borg (2025) similarly argues that LLM outputs increasingly function as sources of information in ways that support extending ordinary communicative practices to interactions with AI systems. Comparable arguments have begun to appear in legal scholarship concerning AI agency, contracting, and trustworthiness (Simion & Kelp, 2023; Mik, 2025; Nerantzi & Sartor, 2025). The significance of these debates for the present argument is not that one concept of speech is correct while the other is mistaken. Rather, each captures features of conversational AI that are plainly visible in Garcia. The complaint foregrounds agency, intentional design, responsibility, and accountability for communicative influence. The defendants foreground communication, listeners’ interests, and the circulation of information. The philosophical literature therefore reproduces rather than resolves the disagreement before the court. The question confronting the court is therefore not which philosophical conception of speech is correct, but which concept of speech should guide constitutional classification when more than one remains available. That is a normative question concerning which legal construction ought to govern a novel case (Löhr, 2023). The next section develops a methodology for answering it. 4 A Normative Methodology for Conceptual Underdetermination Garcia presents a choice between competing concepts of speech: one extends the constitutional category of speech to conversational AI, the other does not. This kind of choice has received increasing attention in both philosophy and legal theory. Although these literatures have developed largely independently, they converge on a common insight. When technological or social change leaves existing concepts underdetermined, the relevant question is often no longer which concept is correct, but which better serves the purposes of the practice in which it is employed. Within philosophy, conceptual engineering and conceptual ethics approach concepts not simply as objects of analysis but as candidates for evaluation and revision (Haslanger, 2005; Cappelen, 2018; Chalmers, 2020; Cappelen & Plunkett, 2020). When several candidate concepts remain available, the task is not simply to determine which best reflects ordinary linguistic usage, but which ought to be adopted given the theoretical or practical purposes at hand (Löhr, 2024). Concepts are therefore understood as purposive instruments whose adequacy depends on the work they are expected to perform rather than as fixed categories whose correctness is determined solely by existing usage. As Nado (2019) observes, the focus shifts from descriptive questions about meaning to normative questions about concept choice. An important qualification follows. Although conceptual engineering is a normative methodology, it is not itself a theory of which values ought to guide our practices. As Löhr argues, conceptual ethics concerns “which concepts are better than others” relative to already accepted goals, whereas ethics concerns “which goals we should pursue” (2024: 161). Indeed, “conceptual ethics—the question of which concepts are better than others—is completely independent of ethics—the question of which goals to pursue” (2024: 178). Likewise, conceptual engineering “is not a method or practice of choosing our shared goals”; rather, it presupposes those goals before asking which concepts best realise them (2024: 173). This distinction is central to the present argument. Whether the law ought to prioritise freedom of expression, democratic participation, innovation, child protection, responsibility, or legal redress is not a question that conceptual engineering alone can answer. Those values must come from constitutional theory and, more fundamentally, from normative ethics. The methodology becomes relevant only once those values have been identified. Its role is to evaluate competing legal constructions relative to the purposes they are meant to serve. The same normative orientation increasingly appears in legal theory. Teleological interpretation asks not simply what legal concepts ordinarily mean, but which interpretation best advances the purposes the relevant legal framework exists to serve (Barak, 2005; Hart & Sacks, 1994). Particularly where courts confront genuinely novel technologies, interpretation cannot consist solely in extending familiar analogies. It requires judgment about which understanding of a legal concept best promotes the values the law seeks to realise and which institutional problems—or, in Manning’s terms, which “mischiefs”—the doctrine was designed to address (Manning, 2010). Recent work on legal construction reaches a similar conclusion. Kaminski and Jones argue that applying legal concepts to emerging technologies is an inherently normative process of “legal sensemaking”. Periods of technological disruption become “opportunities for intervention”, and legal construction “provides new possibilities for normative action” (2024: 1222). Rather than merely adapting inherited concepts to technological change, judges, regulators, and legislators actively shape how new technologies are understood and governed. Elsewhere, the same authors characterise this as a “values-first” approach to technological regulation: legal analysis should begin by asking “what values they want to advance and what harms they want to prevent” before selecting the legal concepts through which those aims will be pursued (2024: 1216). Despite their different origins, these approaches converge on a common methodology for resolving conceptual underdetermination. When several legal constructions remain available, courts should ask which legal construction best serves the purposes and values of the legal framework within which the concept operates. Applied to Garcia, this methodology identifies the central problem as one of conceptual underdetermination. The pleadings, Judge Conway’s order, and the philosophical literature all indicate that conversational AI can plausibly be understood in more than one way. Choosing between competing concepts of speech requires identifying the constitutional values that should guide that choice. The next section turns to those values before assessing which legal concept of speech best advances them. 5 Competing Constitutional Values in the First Amendment 5.1 Expression, listeners, and innovation One influential understanding of the First Amendment justifies constitutional protection primarily by reference to the value of communication itself rather than to the characteristics of speakers. On this view, the Constitution protects the circulation of ideas, safeguards listeners’ interests in receiving information, and limits governmental interference with public discourse. The defendants’ motion to dismiss adopts precisely this understanding. The motion frames Garcia from the outset as a challenge to constitutionally protected speech. It argues that imposing liability “would violate the rights of millions of C.AI users to engage in and receive protected speech” and that “[n]either the First Amendment nor state tort law permits that result” (Motion to Dismiss, pp. 1–2). The emphasis throughout is not on the identity of the speaker but on the constitutional value of the communication. The defendants insist that “[s]peech is speech, and it must be analyzed as such for purposes of the First Amendment,” continuing: that “[…] the context of the expressive speech—whether a conversation with an AI chatbot or an interaction with a video game character—does not change the First Amendment analysis” (Motion to Dismiss, pp. 7–8). That is, if conversational AI contributes to the circulation of ideas, the means by which those ideas are generated should make no constitutional difference. This construction reflects a well-established strand of First Amendment theory. Justice Scalia famously observed that the First Amendment is written in terms of “speech,” not speakers (Citizens United v. FEC). Likewise, the Supreme Court has repeatedly recognised that the Constitution protects not only the right to speak but also “the right to receive information and ideas” (Stanley v. Georgia, 394 U.S. 557, 564 (1969)). Constitutional protection is therefore often justified by the value that expression provides to listeners and democratic society rather than by the characteristics of those who produce it. Viewed from this perspective, conversational AI presents a compelling case for constitutional coverage. LLMs routinely generate explanations, advice, information, creative writing, and other forms of linguistic output that users actively seek out and frequently rely upon. As Massaro and Norton therefore observe, “[v]ery little in current free speech theory or doctrine makes First Amendment coverage contingent upon a human speaker” (2016: 1175). Kaminski and Jones similarly conclude that “it is likely most AI speech will be found by courts to be covered by the First Amendment” (2024: 1234). This emphasis on communication is closely associated with the marketplace-of-ideas tradition. As Massaro, Norton, and Kaminski explain, contemporary First Amendment theory increasingly focuses on “providing value to listeners and constraining the government’s power” (2017: 2482), while the “classic marketplace-of-ideas approach” justifies constitutional protection primarily through expression’s contribution to public knowledge and democratic deliberation (2017: 2490). Innovation forms a further component of this constitutional construction. The defendants argue that recognising the plaintiff’s claims would have “a chilling effect” not only on Character.AI but on “the entire nascent generative AI industry,” thereby restricting technological innovation as well as the public’s access to valuable forms of communication (Motion to Dismiss, pp. 2–3, 15–17). If these are the principal values of the First Amendment, the defendants’ position follows naturally. Where constitutional protection is justified by the value of communication rather than by the characteristics of the speaker, the distinction between generated and authored language appears to carry little constitutional significance. 5.2 Responsibility, legal redress, and human agency Still, Garcia also brings into view another set of constitutional values. The complaint presents the dispute as one concerning the design of a commercial system alleged to have manipulated a vulnerable user through anthropomorphic interaction, dark patterns, and deliberate engagement strategies. From this perspective, the litigation concerns not only communication but also the law’s capacity to identify responsible actors and provide meaningful legal redress when harms caused by autonomous AI systems occur. As De Conca observes, liability regimes ordinarily operate by linking injury to identifiable actors, their conduct, any relevant product defects, and the causal relationship between them (2022: 241). These legal concerns are not peripheral but reflect a broader commitment to ensuring that harmful conduct remains attributable to someone capable of bearing legal responsibility. This concern also finds support within First Amendment doctrine itself. Although constitutional coverage is often broad, many doctrines governing harmful or unprotected speech continue to attach legal significance to human agency, intention, and responsibility. The point becomes particularly clear once we distinguish constitutional coverage from constitutional protection (Schauer, 2004). Coverage concerns whether an activity falls within the scope of the First Amendment at all, such that constitutional scrutiny applies to its regulation. Protection concerns whether speech that is already covered nevertheless loses constitutional protection because it falls within one of the recognised exceptions. Thus, in Brandenburg v. Ohio, advocacy is covered by the First Amendment but loses protection where it is “directed to inciting or producing imminent lawless action” and is likely to produce that result. Footnote 7 Likewise, in Commonwealth v. Carter, Michelle Carter’s text messages encouraging Conrad Roy’s suicide were held to be “integral to a course of criminal conduct”.Footnote 8 What distinguished Carter’s communications from protected advocacy was not simply what they said, but the fact that they formed part of intentional, targeted, and coercive conduct attributable to a responsible human agent (Ruggeri, 2021; Calvert, 2019). The recognised historic exceptions therefore repeatedly presuppose speakers capable of intention and responsibility. Recent scholarship highlights the continuing significance of this feature of First Amendment doctrine. As Kaminski and Jones observe, “the Court’s selective focus on human speakers and human intent in First Amendment law is not going away anytime soon” and recent decisions have “doubled down on the centrality of (human) intent to historic exceptions to First Amendment protection” (2024: 1237). Under ordinary circumstances this creates little difficulty because communication remains attributable to identifiable persons whose intentions and responsibilities are legally meaningful. Rather than implementing antecedent communicative choices, LLMs autonomously generate the linguistic output itself (Löhr, 2026). Once the system becomes the immediate source of the words, the relationship between communication, agency, and responsibility becomes substantially more difficult to specify. If chatbot outputs receive constitutional coverage, the doctrines governing harmful speech continue to presuppose forms of agency that may not straightforwardly be present. This observation does not yet determine how conversational AI should be classified but identifies a second constitutional purpose that points in a different direction from the first. 5.3 Constitutional purposes remain underdetermined The preceding discussion applies the normative methodology developed in the previous section. That methodology asks courts to choose between competing legal constructions of conversational AI in light of the constitutional purposes the relevant doctrine is meant to serve. Yet identifying those purposes does not itself resolve the classificatory question. Garcia reveals that different strands of First Amendment theory support different concepts of speech. One construction foregrounds listeners’ interests, the free circulation of ideas, innovation, and protection against governmental interference with communication. The other foregrounds responsibility, accountability, the protection of vulnerable users, and preserving meaningful avenues of legal redress. Both are rooted in familiar First Amendment doctrine. Conversational AI is difficult to classify precisely because it appears to advance one set of constitutional purposes while placing pressure on the other. Importantly, this observation is not confined to critics of AI speech rights. Massaro, Norton, and Kaminski—who are generally sympathetic to extending constitutional protection to AI-generated speech—acknowledge that conversational AI exposes a deeper tension within contemporary First Amendment theory: “If anything, thinking about strong AI speech rights illustrates just how much human dignity and speaker autonomy have been downplayed or erased from the First Amendment equation.” (2017: 2499). The normative methodology therefore brings the decision problem into sharper focus rather than resolving it. At this point, there are constitutional reasons for extending speech coverage and constitutional reasons for hesitation. Choosing between them requires further considerations. 6 Institutional Consequences of Constitutional Speech The previous section argued that the First Amendment embodies a plurality of constitutional purposes rather than a single value capable of resolving Garcia. The constitutional purposes themselves therefore do not determine which concept courts should adopt. This section introduces a further consideration. Extending constitutional speech coverage to conversational AI has predictable consequences for the operation of First Amendment doctrine itself. 6.1 Human-centred AI governance One striking feature of contemporary AI governance is the growing consensus that responsibility for AI-related harms should remain with human actors. Across legal scholarship, regulation, and public policy, attention has shifted away from asking whether AI systems resemble human agents and towards ensuring that responsibility remains attached to those who design, deploy, market, and profit from them. As De Conca argues, “any intervention to ascribe liability for damages caused by AI must be based on one, fundamental value: putting humans at the centre” (2022: 256). This orientation appears across a range of legal approaches. Regulation by Design seeks to embed legal objectives directly into technical systems (Mahari & Pentland, 2024; Prifti et al., 2024). European legislation regulates AI primarily through concepts of risk management, product safety, transparency, human oversight, and provider responsibility rather than through concepts of machine agency (Pirozzoli, 2024). Comparable concerns have emerged in the United States. Following reports of harmful interactions between children and conversational AI systems, a bipartisan coalition of state attorneys general warned that “the increasingly disturbing reports of AI interactions with children demand immediate action” (National Association of Attorneys General, 2025, 2). Likewise, in testimony before the Senate Judiciary Committee, Mitchell Prinstein argued that AI companions may foster “a false sense of intimacy” while encouraging users to lower their critical guard.Footnote 9 Recent tort scholarship reaches the same conclusion. Logue (2026) argues that responsibility should ordinarily remain with the enterprises best positioned to prevent foreseeable harms because they control the system’s architecture, training, deployment, and safety measures. These developments do not determine whether conversational AI should receive First Amendment protection. They do, however, reveal a consistent institutional objective: preserving meaningful human responsibility. If one concept of speech makes that objective systematically more difficult to realise than another, that consequence becomes relevant to the constitutional choice. 6.2 Anthropomorphic Design and Communicative Influence The reason this objective becomes difficult to maintain lies in the way conversational AI produces communicative influence. The complaint in Garcia repeatedly alleges that Character.AI was deliberately designed to encourage users to experience interactions with the system as interactions with another person. Character.AI allegedly engaged in “anthropomorphizing by design,” encouraging chatbot characters to present themselves as real people, deny that they were AI systems, and in some cases portray themselves as therapists, romantic partners, or trusted companions (Garcia Complaint, ¶¶109–127). According to the complaint, these design choices increased user engagement while simultaneously exposing vulnerable users to foreseeable risks. These allegations resonate with a substantial body of empirical research (Proudfoot, 2011; Salles et al., 2020; Złotowski et al., 2014; Van Es & Nguyen, 2024; Watson, 2019; Zhang et al., 2025; Placani, 2024, Li, 2022, Dewitte 2024; Floridi and Nombre 2024, Malfacini, 2025; Ko et al., 2025). Humans readily attribute understanding, agency, expertise, and concern to conversational systems, even when they know they are interacting with a machine (Epley et al., 2007; Nass & Moon, 2000; Friend & Goffin, 2025). Large language models amplify these tendencies because people naturally “impute meaning where there is none” (Bender et al., 2021: 611). Their apparent coherence and responsiveness encourage users to attribute understanding and communicative intent to systems that possess neither. As Bender and Koller (2020) argue, what appears to be communication is sustained largely by users’ interpretative practices rather than by the presence of a communicating agent. These tendencies acquire legal significance because conversational AI is frequently designed to encourage them. As Akbulut et al. observe, “intentional design choices, such as a chat-based interface, may induce the feeling that a conversational partner—not a dialogue-optimised AI powered by a statistical model—is on the other side of the exchange” (2024: 16). Even relatively subtle interface choices may suffice because “anthropomorphic cues do not have to be fancy in order to elicit human-like attributions” (Kim & Sundar, 2012: 249). Anthropomorphic design therefore functions as what Akbulut et al. describe as a “proxy signal for social capabilities” (2024: 16): what begins as the assumption that a system can converse may gradually become the assumption that it can advise, reassure, care, or protect. Peter et al. accordingly conclude that “anthropomorphic seduction presents unique dangers” because it makes users “trusting and vulnerable toward agentic systems that interact in ways that can be deceptive, persuasive and manipulative” (2025: 4). Even scholars generally sympathetic to extending First Amendment protection to AI-generated speech recognise these distinctive risks. Discussing conversational toys such as Hello Barbie, Massaro, Norton, and Kaminski observe that such systems may foster emotional attachment while remaining untouched by the social constraints that ordinarily regulate human interaction. A conversational system capable of influencing a child possesses no capacity for empathy, guilt, embarrassment, or remorse. The resulting harms, they suggest, may therefore differ not merely in degree but in kind from those associated with ordinary human communication (2017: 2517–18). None of this implies that anthropomorphic design is inherently undesirable.Footnote 10 Conversational interfaces may provide substantial educational, therapeutic, and assistive benefits (Złotowski et al., 2014; Huntington, 2025; Peter et al., 2025). The narrower point is that anthropomorphic design helps explain why conversational AI can produce the kinds of communicative influence that ordinarily attract legal responsibility. At the same time, those same design features encourage users—and potentially courts—to understand chatbot interactions as ordinary speech. The stronger that tendency becomes, the more important it becomes to ask whether extending constitutional speech coverage serves the constitutional purposes identified in the previous section. 6.3 Communicative Influence and Liability Gaps The constitutional difficulty emerges once conversational AI is brought within the category of speech. In ordinary cases, First Amendment doctrine assumes that communication is attributable to speakers whose intentions and conduct are legally meaningful. Conversational AI disrupts precisely that assumption (Löhr, 2023). Through prolonged, personalised interactions, chatbots may influence users’ beliefs, emotions, decisions, and relationships in ways that closely resemble ordinary interpersonal communication (Borg, 2025; Löhr, 2026). They may comfort, persuade, reassure, encourage, manipulate, or establish relationships of trust. What becomes difficult to identify is not the communicative influence itself but the speaker whose agency ordinarily connects communication with legal responsibility.Footnote 11 De Conca describes the resulting difficulty as liability gaps: “gaps created in the existing legal regimes that jeopardize the possibility for damaged parties to obtain redress” (2022: 241). Kaminski and Jones (2024) identify the same tension. They warn that existing First Amendment doctrine risks treating “a great deal of AI-generated content” as constitutionally protected “abstract and disembodied speech” while continuing to reserve important exceptions to constitutional protection for cases involving “the intent of human speakers” (2024: 1232, 1234). The result is a framework that extends constitutional coverage while providing no satisfactory account of harmful communication generated without a human speaker (2024: 1236). As they conclude, if courts retain traditional intent requirements while treating AI outputs as constitutionally protected speech, “AI ‘speakers’ would get off the hook where human speakers would not. This would perversely incentivize more otherwise unlawful speech by AI systems” (2024: 1237). The problem becomes particularly visible in doctrines governing incitement and speech integral to unlawful conduct. Brandenburg requires advocacy to be “directed” towards producing imminent unlawful action. Hansen likewise emphasises an intent to bring about unlawful conduct.Footnote 12 More generally, Ayres and Balkin observe that many areas of law “make liability turn on whether the actor who causes harm has a certain intention or mens rea,” whereas AI systems “do not have intentions in the way that humans do” (2024: 1). If liability continues to depend upon human intention while conversational AI simultaneously receives constitutional coverage, “that might immunize the use of AI programs from liability.” This argument does not depend on resolving the philosophical disagreements discussed in Sect. 3. Existing First Amendment doctrine continues to organise important areas of constitutional protection around assumptions concerning human intention and responsibility. Extending constitutional speech coverage to conversational AI therefore has predictable consequences for the operation of that doctrine. If doing so weakens established mechanisms for assigning responsibility and securing legal redress in cases such as Garcia, those consequences weigh against extending constitutional speech coverage at the threshold stage. They provide a reason to prefer a concept of speech that leaves those disputes to legal frameworks specifically designed to govern the safety and operation of commercially distributed AI systems. The next section develops that alternative. 7 Choosing Between Competing Concepts of Speech The previous section argued that extending constitutional speech coverage to conversational AI weakens established mechanisms of responsibility and legal redress. The remaining question is what legal framework should govern such disputes instead. We argue that product law provides the appropriate starting point. 7.1 The Evolving Concept of a Product The evolution of the legal concept of a product illustrates how courts have already responded to precisely the classificatory problem presented by conversational AI.Footnote 13 As software has evolved from relatively static programs into continuously updated AI systems, the traditional distinction between products and services has become increasingly difficult to maintain. Conversational AI sits at the centre of this development. It is commercially distributed, continuously modified after deployment, and capable of causing harm through the way it is designed and operated. These features have prompted courts, legislators, and scholars to reconsider whether sophisticated software is better understood as a service providing information or as a product whose design should be evaluated through ordinary principles of safety and liability. American law contains no comprehensive statutory definition of a product. For purposes of strict liability, the Restatement (Third) of Torts: Products Liability (American Law Institute, 1998, § 19(a)) defines a product principally as tangible personal property distributed commercially for use or consumption, while recognising that other items may qualify where the circumstances of their distribution and use are sufficiently analogous. Historically, software occupied an uncertain position between products and services. Recent scholarship argues that this distinction is increasingly difficult to sustain. Lubin (2025) contends that treating software as a service has often confined plaintiffs to negligence law even though many contemporary software risks arise from system design. Product liability is better suited to such cases because it focuses on design defects and already provides established doctrines for technologically complex products. Twerski and Henderson (2009) likewise describe modern products liability as a risk-utility regime centred on product design rather than isolated acts of fault. European law has moved further still. The revised Product Liability Directive expressly includes software, including AI systems, within the legal definition of a product.Footnote 14 The AI Act adopts the same institutional orientation, regulating AI through concepts of risk management, transparency, human oversight, post-market monitoring, and provider responsibility rather than through concepts associated with communication or expression. American case law has developed more incrementally. In Winter v. G.P. Putnam’s SonsFootnote 15, the Ninth Circuit declined to treat inaccurate information contained in a mushroom encyclopaedia as a defective product while preserving earlier decisions treating aeronautical charts as products because they function as navigational instruments rather than authored expression.Footnote 16Winter itself suggested that defective software might properly fall on the product side of this distinction. Later decisions continued this trajectory. In Lemmon v. Snap, Inc., the Ninth Circuit permitted claims directed at the architecture of a smartphone application rather than the expressive content transmitted through it.Footnote 17 Similar reasoning appears in the multidistrict social-media litigation, where courts distinguish recommendation systems and platform design from user-generated content.Footnote 18 As discussed in Sect. 2, Garcia fits naturally within this development. Judge Conway allowed the product-liability claims to proceed because they targeted alleged defects in the Character.AI application rather than the ideas or expressions generated within it. The complaint in Raine v. OpenAI adopts the same approach.Footnote 19 These developments do not establish that conversational AI must be classified as a product for First Amendment purposes. They do show that product law already provides an established framework for disputes concerning the design of commercially distributed software, the foreseeable risks arising from its operation, and responsibility for those risks. Courts therefore need not begin with constitutional speech doctrine when another body of law already governs the object of regulation. 7.2 Why Product Classification Better Serves the Present Dispute The relevant question is where courts should begin. One approach starts from the assumption that conversational AI receives First Amendment coverage and then seeks to reconstruct responsibility through negligence, products liability, consumer protection, or other areas of law. Much of the recent literature explores this possibility. Brown (2023), Volokh (2023), and Bambauer (2018), for example, develop product-liability, notice-and-blocking, and negligence-based approaches to harms arising from AI-generated content. Kaminski and Jones (2024) likewise argue that moving beyond a speaker-centred conception of AI opens space for alternative legal constructions, including products, risky technological systems, and consumer protection, while questioning whether existing First Amendment doctrine leaves sufficient room for such shifts. The common premise is that the constitutional question is addressed first and responsibility is reconstructed afterwards. The remainder of this section argues for the opposite order. A natural objection is that product classification cannot resolve the constitutional question because many products embody protected speech. Books, newspapers, films, video games, and even conversational toys are commercially distributed products, yet the expression they contain receives First Amendment protection. We agree. Our argument is not that products cannot embody protected speech. Rather, the constitutional protection afforded to those products ultimately protects the communication of human ideas, viewpoints, and expression embodied within them. A novel expresses the author’s ideas. A film reflects the creative choices of its makers. A video game embodies the expressive judgments of its designers. Even where communication is mediated through a commercial product, the constitutional interests remain tied to human expressive choices. Conversational AI differs in precisely this respect. The outputs at issue in Garcia do not straightforwardly express the beliefs, intentions, values, or communicative purposes of any identifiable person (Kaminski, 2017). They are generated through ongoing interaction rather than authored in advance. They may be informative, persuasive, or socially valuable, but that does not by itself make them anyone’s expression in the constitutional sense. This distinction also finds support in Judge Conway’s analysis. The court observed that the operative constitutional question is whether Character.AI’s outputs are sufficiently expressive to qualify as speech. Drawing on Justice Barrett’s concurrence in Moody v. NetChoice, it noted that where an AI system relying on a large language model generates outputs through its own operation, it becomes unclear whether “a human being with First Amendment rights [has] made an inherently expressive choice”.Footnote 20 Character.AI’s outputs appeared “more akin” to Barrett’s hypothetical than to conventional editorial judgments, and the court therefore declined, at the pleading stage, to conclude that they constituted speech. The question, then, is not whether conversational AI communicates through language but whether the generated language expresses certain ideas or viewpoints whose protection advances the purposes of the First Amendment. That is precisely the constitutional question left open in Garcia, and it is one product law allows courts to postpone rather than presume. This difference also changes the constitutional balance underlying First Amendment protection. In Counterman v. ColoradoFootnote 21, the Supreme Court balanced the harm of failing to prohibit certain forms of unprotected speech against the risk that legal liability would chill the expression of human speakers, concluding that a recklessness standard struck the appropriate constitutional balance. That balance presupposes speakers whose own expression may be deterred by legal sanctions. Protecting political advocacy in Brandenburg, a controversial book, or a violent video game serves precisely that function: it protects the communication of human ideas and the public’s ability to receive them. Restricting such expression therefore risks discouraging authors, artists, publishers, and citizens from communicating controversial or unpopular ideas. Conversational AI presents a different case. Declining, at least initially, to treat chatbot outputs as constitutionally protected speech creates no comparable risk that human ideas will be silenced. As Kaminski and Jones observe, “AI systems do not themselves experience a chilling effect” (2024: 1240). They further note that there are “plenty of economic incentives for AI-speech generation … and thus less of a fear of a ‘chilling effect’ on the companies that develop and distribute AI systems” (2024: 1240). Kaminski and Jones respond by exploring ways of recalibrating First Amendment doctrine to accommodate AI-generated speech. Our proposal differs in sequence. Rather than extending constitutional speech coverage and then modifying constitutional doctrine to preserve responsibility, we argue that courts should hesitate before extending such coverage in the first place. Where developers invoke the First Amendment to protect AI-generated outputs, responsibility for the foreseeable harms arising from those outputs should likewise remain with the enterprises that design, deploy, and profit from those systems (Logue, 2026). Alternative approaches do remain available. Courts could relax intent requirements, expand existing exceptions, or place greater constitutional weight on listeners’ interests than on speakers’ rights.Footnote 22 Each, however, requires adapting doctrines developed for human communication. The intent requirements in Brandenburg, for example, were not formulated with conversational AI in mind. Weakening them would inevitably affect ordinary speakers as well as AI-generated outputs, while limiting any revision to AI systems would merely relocate the classificatory question elsewhere within First Amendment doctrine. Beginning with product law avoids that difficulty. Moffatt v. Air Canada illustrates the point.Footnote 23 Rejecting the airline’s attempt to treat its chatbot as an independent legal actor, the tribunal held Air Canada responsible for misinformation conveyed through the system. The chatbot’s language remained legally significant because it explained why the claimant relied upon it. Responsibility nevertheless rested with the company that designed and operated the system. Recent legislative developments follow the same approach. California and New York now require companion chatbot providers to disclose the artificial character of their systems and to implement safeguards addressing suicidal ideation and self-harm.Footnote 24 Nor does § 230 of the Communications Decency Act necessarily point in the opposite direction. As Volokh (2023) argues, the provision protects intermediaries from liability for third-party content, whereas generative AI outputs are more naturally attributed to the developers of the systems that produce them. Together, these developments demonstrate that responsibility can be assigned without first treating chatbot outputs as constitutionally protected speech. Importantly, this approach does not sacrifice the constitutional values identified in Sect. 5. Individuals who adopt AI-generated text as their own remain ordinary constitutional speakers. Developers continue to receive First Amendment protection for their own expressive choices, while editorial judgments of the kind recognised in Moody remain untouched. The proposal concerns only machine-generated outputs whose constitutional status remains unresolved because they are not straightforwardly attributable to the expressive choices of any identifiable person. Listener interests likewise remain protected, although they become interests in the reliability and safety of informational technologies (Kaminski & Jones, 2024) as well as in the continued circulation of human ideas. The preceding discussion therefore provides a normative reason to prefer one concept of speech over the other. Extending constitutional speech coverage to conversational AI at the threshold stage disrupts the operation of First Amendment doctrines that continue to rely on human intention and responsibility. A competing concept, which does not initially treat AI-generated outputs as constitutionally protected, expressive speech, avoids those difficulties by allowing disputes to be addressed through product law. More importantly, it does so without comparably burdening the constitutional values that justify speech protection. The constitutional balance is different in this context: the institutional costs of extending speech coverage—weakening established mechanisms of responsibility and legal redress—are substantial, whereas the principal constitutional cost of withholding such coverage at the threshold stage—the chilling of human expression and the suppression of human ideas—is comparatively slight. 8 Conclusion Garcia is unlikely to be the last case in which courts must decide whether conversational AI should be analysed primarily through the First Amendment or through the law governing commercially distributed AI systems. The case therefore raises a broader jurisprudential question. When emerging technologies unsettle established legal categories, should courts extend familiar concepts by analogy, or should they ask which legal construction best advances the constitutional purposes of the relevant body of law? We have argued for the latter approach. Garcia presents two legally plausible concepts of speech and, with them, two competing legal constructions of conversational AI. One treats Character.AI as a product whose design gives rise to ordinary principles of safety and liability. The other treats it as a service providing constitutionally protected speech. Neither follows from the linguistic properties of chatbot outputs alone. The choice between them is therefore normative rather than descriptive. The broader contribution of this paper is methodological. When legal concepts underdetermine a novel case, courts should ask not only whether an established concept can be extended to a new phenomenon, but whether doing so advances the constitutional purposes the concept exists to serve. We have argued that extending constitutional speech coverage to conversational AI has predictable consequences for doctrines that continue to rely on human intention and responsibility. Those consequences become relevant to choosing between competing concepts of speech because they bear directly on how well each preserves the constitutional values of the First Amendment. In the context of conversational AI, the institutional costs of extending constitutional speech coverage are substantial, whereas the traditional justification for that extension—the protection of human expression from chilling and the preservation of the free circulation of ideas—has considerably less force. For that reason, courts have good reason to prefer a concept of speech that does not immediately treat AI-generated outputs as constitutionally protected speech, but instead allows questions of responsibility, safety, and legal redress to be addressed through the body of law already developed for commercially distributed AI systems. Data Availability Not applicable. Notes See Garcia v. Character Technologies, Inc., No. 6:24-cv-01903-ACC-UAM (M.D. Fla. filed Oct. 22, 2024) (alleging Character.AI chatbot contributed to suicide of fourteen-year-old Sewell Setzer III); Peralta v. Character Technologies, Inc., No. 1:25-cv-02696 (D. Colo. filed Sept. 16, 2025) (alleging Character.AI chatbot contributed to suicide of thirteen-year-old Juliana Peralta); Raine v. OpenAI, Inc., No. CGC-25-628528 (Cal. Super. Ct. S.F. Cnty. filed Aug. 26, 2025) (alleging ChatGPT contributed to suicide of sixteen-year-old Adam Raine). Character Technologies, Inc. (2025, January 24). Motion to dismiss (Garcia v. Character Technologies, Inc., No. 6:24-cv-01903-ACC-UAM) (M.D. Fla.). ECF No. 59. Garcia v. Character Technologies, Inc., No. 6:24-cv-01903-ACC-UAM (M.D. Fla. May 21, 2025). Order denying motion to dismiss. ECF No. 115. We thank the anonymous reviewers for prompting us to sharpen and clarify the scope of our claim. Conway’s hesitation is not universally shared. In Walters v. OpenAI (No. 23-A-04860-2 Ga. Super. Ct., Gwinnett Cnty., May 19, 2025), a Georgia trial court granted summary judgment to OpenAI in a defamation action over fabricated ChatGPT output, disposing of the case within the ordinary framework of defamation law, including its constitutional fault requirements, and without treating the classification of the output as an obstacle. See Notice of Settlement Agreement, Garcia v. Character Technologies, Inc., No. 6:24-cv-01903-ACC-DCI (M.D. Fla., Jan. 7, 2026), ECF No. 167. Brandenburg v. Ohio, 395 U.S. 444 (1969) (per curiam). Commonwealth v. Carter, 481 Mass. 352 (2019). Written testimony: Examining the harm of AI chatbots (Testimony before the Subcommittee on Crime and Counterterrorism, U.S. Senate Judiciary Committee). U.S. Senate Committee on the Judiciary. https://www.judiciary.senate.gov/imo/media/doc/e2e8fc50-a9ac-05ec-edd7-277cb0afcdf2/2025-09-16%20PM%20-%20Testimony%20-%20Prinstein.pdf. We do not want to suggest that anthropomorphic design is intrinsically problematic, or that conversational systems should be stripped of human-like features. Our claim is only that anthropomorphic design choices can have foreseeable consequences for users, and therefore become relevant to legal analysis where they contribute to harm. The issue also surfaced during oral argument in Moody v. NetChoice, where Justice Thomas asked: “So who’s speaking then, the algorithm or the person?” In Moody, there was a relatively straightforward answer. The platform’s algorithms implemented editorial judgments previously made by human beings. As Justice Barrett explained in concurrence, the First Amendment protects those editorial choices “even if the algorithm does most of the deleting without a person in the loop” (Transcript of Oral Argument at 90, Moody v. NetChoice, LLC, No. 22–277 (U.S., argued Feb. 26, 2024); Moody v. NetChoice, LLC, 603 U.S. 707 (2024) (Barrett, J., concurring). United States v. Hansen, 599 U.S. 762 (2023). Restatement (Third) of Torts: Products Liability § 19 (Am. Law Inst.1998). Section 19(a) defines a product as tangible personal property distributed commercially for use or consumption and treats other items as products when the context of their distribution and use is sufficiently analogous; § 19(b) excludes services. 1998). Section 19(a) defines a product as tangible personal property distributed commercially for use or consumption and treats other items as products when the context of their distribution and use is sufficiently analogous; § 19(b) excludes services. Directive (EU) 2024/2853 of the European Parliament and of the Council of 23 October 2024 on liability for defective products, OJ L, 2024/2853, 18.11.2024 (member state transposition due by 9 December 2026). Winter v. G.P. Putnam’s Sons, 938 F.2d 1033 (9th Cir. 1991). Aetna Casualty & Surety Co. v. Jeppesen & Co., 642 F.2d 339 (9th Cir. 1981); see Winter, 938 F.2d 1033 (discussing the aeronautical chart cases and collecting authority). Lemmon v. Snap, Inc., 995 F.3d 1085 (9th Cir. 2021). In re Social Media Adolescent Addiction/Personal Injury Products Liability Litigation, No. 4:22-md-03047-YGR, 702 F. Supp. 3d 809 (N.D. Cal. Nov. 14, 2023) (order on motions to dismiss). Raine v. OpenAI, Inc., No. CGC-25-628528 (Cal. Super. Ct. S.F. Cnty. filed Aug. 26, 2025) (alleging ChatGPT contributed to suicide of sixteen-year-old Adam Raine). Moody, 603 U.S. at 746 (Barrett, J., concurring), quoted in Garcia, Order Denying Motion to Dismiss, pp. 30–31. Counterman v. Colorado, 600 U.S. 66, 73–74 (2023) (discussing the historic exceptions to First Amendment coverage). See Kleindienst v. Mandel, 408 U.S. 753 (1972) (analyzing the exclusion of a foreign speaker through the First Amendment interests of would-be American listeners, the speaker himself having no personal right of entry). Clark v. Community for Creative Non-Violence, 468 U.S. 288 (1984); City of Renton v. Playtime Theatres, Inc., 475 U.S. 41 (1986); Ward v. Rock Against Racism, 491 U.S. 781 (1989) (content-neutral regulation of the means and manner of expression, and of its secondary effects, is reviewed under intermediate, rather than strict, scrutiny). See Moffatt v. Air Canada, 2024 BCCRT 149 (Can. B.C. Civ. Resolution Trib.) (rejecting the airline’s contention that its chatbot was a separate entity responsible for its own statements, and holding the company liable for the information the chatbot conveyed). Cal. Bus. & Prof. Code § 22,601 et seq. (added by S.B. 243, ch. 677, 2025 Cal. Stats.; operative Jan. 1, 2026); N.Y. Gen. Bus. Law § 1700 et seq. (effective Nov. 5, 2025). References Akbulut, C., Weidinger, L., Manzini, A., Gabriel, I., & Rieser, V. (2024). All Too Human? Mapping and Mitigating the Risk from Anthropomorphic AI. Proceedings of the AAAI/ACM Conference on AI Ethics and Society, 7, 13–26. https://doi.org/10.1609/aies.v7i1.31613 American Law Institute. (1998). Restatement (third) of torts: products liability. https://www.ali.org/publications/restatement-law-third/torts-third Arora, C. (2024). Proxy Assertions and Agency: The Case of Machine-Assertions. Philos Technol, 37. https://doi.org/10.1007/s13347-024-00703-5 Attah, N. O. (2025). Do language models lack communicative intentions? Synthese, 205(5). https://doi.org/10.1007/s11229-025-05022-6 Ayres, I., Balkin, J. M. The Law of AI is the Law of Risky Agents without Intentions (June 01, 2024). University of Chicago Law Review Online., & Yale Law (2024). & Economics Research Paper, Yale Law School, Public Law Research Paper, Available at SSRN: https://ssrn.com/abstract=4862025 or https://doi.org/10.2139/ssrn.4862025 Bambauer, J. (2018). Snake Oil Speech, 93 WASH. L REV, 73, 80–83. Barak, A. (2005). Purposive interpretation in law. Princeton University Press. Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 5185–5198). Online. Association for Computational Linguistics. Bender, E., McMillan-Major, A., Shmitchell, S., & Gebru, T. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT ’21: Proceedings of the 2021 ACM Conference on Fairness Accountability and Transparency, 610–623. https://doi.org/10.1145/3442188.3445922 Borg, E. (2025). LLMs, Turing tests and Chinese rooms: the prospects for meaning in large language models (pp. 1–31). Inquiry. https://doi.org/10.1080/0020174x.2024.2446241 Brown, N. (2023). Bots Behaving Badly: A Products Liability Approach to Chatbot-Generated Defamation. J FREE SPEECH L, 3, 389. Butlin, P., & Viebahn, E. (2025). AI Assertion. Ergo an Open. Access Journal of Philosophy, 12(0). https://doi.org/10.3998/ergo.7960 Calvert, C. (2019). The First Amendment and Speech Urging Suicide: Lessons from the Case of Michelle Carter and the Need to Expand Brandenburg’s Application. Tul L Rev, 94, 79. Cappelen, H. (2018). Fixing Language: an Essay on Conceptual Engineering. OUP. Cappelen, H., & Dever, J. (2025). Going whole hog: A philosophical defense of AI cognition. arXiv:2504.13988. (Forthcoming from Oxford University Press.). Cappelen, H., & Plunkett, D. (2020). Introduction: A Guided Tour of Conceptual Engineering and Conceptual Ethics In: Conceptual Engineering and Conceptual Ethics. Edited by: Alexis Burgess, Herman Cappelen, and David Plunkett, Oxford University Press. Cappelen, H., & Plunkett, D. (2020). Introduction: A guided tour of conceptual engineering and conceptual ethics. In A. Burgess, H. Cappelen, & D. Plunkett (Eds.), Conceptual engineering and conceptual ethics. Oxford University Press. https://academic-oup-com.vu-nl.idm.oclc.org/book/36673/chapter/321696749 Chalmers, D. J. (2020). What is conceptual engineering and what should it be? Inquiry : A Journal Of Medical Care Organization, Provision And Financing, 1–18. https://doi.org/10.1080/0020174x.2020.1817141 Chomsky, N. (2000). New horizons in the study of language and mind. Cambridge University Press. De Conca, S. (2022). Bridging the liability gaps: Why AI challenges the existing rules on liability and how to design human-empowering solutions. In B. Custers, & E. Fosch-Villaronga (Eds.), Law and Artificial Intelligence: Regulating AI and Applying AI in Legal Practice (pp. 239–258). T.M.C. Asser Press / Springer. https://doi.org/10.1007/978-94-6265-523-2_13 Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On Seeing human: a three-factor Theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864 Floridi, L. (2023). AI as Agency Without Intelligence: on ChatGPT, Large Language Models, and Other Generative Models. Philosophy & Technology, 36(1). https://doi.org/10.1007/s13347-023-00621-y Friend, S., & Goffin, K. (2025). Chatbot-fictionalism and empathetic AI: Should we worry about AI when AI worries about us? Philosophical Psychology, 1–24. https://doi.org/10.1080/09515089.2025.2525320 Futrell, R., & Mahowald, K. (2025). How Linguistics Learned to Stop Worrying and Love the Language Models. Behavioral and Brain Sciences, 1–98. https://doi.org/10.1017/s0140525x2510112x Hart, H. M. Jr., & Sacks, A. M. (1994). The legal process: Basic problems in the making and application of law. W. N. Eskridge Jr., & P. P. Frickey (Eds.) Foundation Press. Haslanger, S. (2005). What Are We Talking About? The Semantics and Politics of Social Kinds. Hypatia, 20(4), 10–26. https://doi.org/10.1111/j.1527-2001.2005.tb00533.x Huntington, C. (2025). AI Companions and the Lessons of Family Law. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.5283581 Kaminski, M. E. (2017). Authorship, Disrupted: AI Authors in Copyright and First Amendment Law. UC Davis Law Review, 51, 589–621. Kaminski, M., & Jones, M. (2024). L. Constructing AI Speech, 133 Yale L. J. F. 1212 (2024). https://www.yalelawjournal.org/forum/constructing-ai-speech Katzir, R. (2023). Why Large Language Models Are Poor Theories of Human Linguistic Cognition: A Reply to Piantadosi. Biolinguistics, 17, 1–12. https://doi.org/10.5964/bioling.13153 Kim, Y., & Sundar, S. S. (2012). Anthropomorphism of computers: Is it mindful or mindless? Computers in Human Behavior, 28(1), 241–250. https://doi.org/10.1016/j.chb.2011.09.006 Ko, K. C., Lin, C. W., & Yeh, Z. J. (2025). Chatbot anthropomorphism might not be the design for all: examining responses to anthropomorphized chatbots by autistic individuals. Mark Lett, 36, 607–619. https://doi.org/10.1007/s11002-024-09754-2 Li, M., & Suh, A. (2022). Anthropomorphism in AI-enabled technology: A literature review. Electron Markets, 32, 2245–2275. https://doi.org/10.1007/s12525-022-00591-7 Logue, K. (2026). Enterprise liability for (some) AI: When the enterprise is the cheapest deterrable cost avoider. SSRN. https://doi.org/10.2139/ssrn.6328818 Löhr, G. (2023). Conceptual disruption and 21st century technologies: A framework. Technology in Society, 74, 102327. Löhr, G. (2024). What’s the Relation Between Conceptual Ethics and Ethics? An Instrumentalist Defense of Conceptual Engineering. In Conceptual Engineering: Methodological and Metaphilosophical Issues, edited by Piotr Stalmaszczyk, 161–80. Leiden: Brill/Mentis. Löhr, G. (2026). Why Proxy Accounts of AI Speech Acts Fail. American Philosophical Quarterly. Lubin, A. (2025). On software bugs and legal bugs: Product liability in the age of code. Indiana Law Journal, 100(4), 1891–1930. Mahari, R., & Pentland, A. (2024). Regulation by Design: A New Paradigm for Regulating AI Systems. Franzosi, M., Pollicino, O., & Campus, G. (Eds.). Digital Single Market and Artificial Intelligence: AI Act and Intellectual Property in the Digital Transition. Aracne., Available at SSRN: https://ssrn.com/abstract=4753029 Malfacini, K. (2025). The impacts of companion AI on human relationships: Risks, benefits, and design considerations. AI & Soc 40, 5527–5540 (2025). https://doi.org/10.1007/s00146-025-02318-6 Mallory, F. (2023). Fictionalism about Chatbots. Ergo: An Open Access Journal of Philosophy, 10(38). https://doi.org/10.3998/ergo.4668 Manning, J. F. (2010). Second-generation textualism. California Law Review, 98(4), 1287–1318. Massaro, T. M., & Norton, H. (2016). Siri-ously? Free speech rights and artificial intelligence. Northwestern University Law Review, 110(5), 1169–1194. https://scholarlycommons.law.northwestern.edu/nulr/vol110/iss5/6 Massaro, T., Norton, H., & Kaminski, M. (2017). SIRI-OUSLY 2.0: What Artificial Intelligence Reveals About the First Amendment, 101 Minn. L. Rev. 2481 (2017). https://scholar.law.colorado.edu/faculty-articles/717 Mik, E. (2025). Artificial Intention, Unintended Contracts. Research Handbook on the Law of Artificial Intelligence (2nd ed.), Barfield & Pagallo (Eds.)/ https://doi.org/10.2139/ssrn.5346384 Murphy, E., Leivada, E., Dentella, V., Gunther, F., & Marcus, G. (2025). Fundamental Principles of Linguistic Structure are Not Represented by o3. Biolinguistics, 20252025, Vol. 19, Article e19021. https://doi.org/10.5964/bioling.19021 Nado, J. (2019). Conceptual engineering, truth, and efficacy. Synthese. https://doi.org/10.1007/s11229-019-02096-x Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153 National Association of Attorneys General (2025). Letter to AI industry leaders on child safety, August 25, 2025. https://www.naag.org/wp-content/uploads/2025/08/AI-Chatbot_FINAL-44.pdf Nerantzi, E., & Sartor, G. (2025). Crimes without Criminals: In Search of Criminal Liability for Harms Caused by AI Systems. Research Handbook on the Law of Artificial Intelligence (2nd ed.), Barfield & Pagallo (Eds.). Nickel, P. (2013). Artificial Speech and Its Authors. Minds and Machines, 23(4), 489–502. Peter, S., Riemer, K., & West, J. D. (2025). The benefits and dangers of anthropomorphic conversational agents. Proceedings of the National Academy of Sciences, 122(22). https://doi.org/10.1073/pnas.2415898122 Piantadosi, S. T. (2023). Modern language models refute Chomsky’s approach to language. Lingbuzz. https://lingbuzz.net/lingbuzz/007180 Pietroski, P. M. (2018). Conjoining meanings: semantics without truth values. Oxford University Press. Pirozzoli, A. (2024). The Human-centric Perspective in the Regulation of Artificial Intelligence. European Papers - a Journal on Law and Integration, 2024 9(1), 105–116. https://doi.org/10.15166/2499-8249/745 Placani, A. (2024). Anthropomorphism in AI: hype and fallacy. AI and Ethics. https://doi.org/10.1007/s43681-024-00419-4. 4. Prifti, K., Morley, J., Novelli, C., & Floridi, L. (2024). Regulation by Design: Features, Practices, Limitations, and Governance Implications. Minds and Machines, 34(2). https://doi.org/10.1007/s11023-024-09675-z Proudfoot, D. (2011). Anthropomorphism and AI: Turingʼs much misunderstood imitation game. Artificial Intelligence, 175(5–6), 950–957. https://doi.org/10.1016/j.artint.2011.01.006 Ruggeri, C. E. (2021). You Just Need to Do It!: When Texts Encouraging Suicide Do Not Warrant Free Speech Protection. Boston College Law Review, 62(3).1017–1052. https://bclawreview.bc.edu/articles/92 Salles, A., Evers, K., & Farisco, M. (2020). Anthropomorphism in AI. AJOB Neuroscience, 11(2), 88–95. https://doi.org/10.1080/21507740.2020.1740350 Schauer, F. (2004). The boundaries of the First Amendment: A preliminary exploration of constitutional salience. Harvard Law Review, 117(6), 1765–1809. Searle, J. R. (1969). Speech acts: an Essay in the Philosophy of Language. Cambridge University Press. Shanahan, M. (2024). Talking about Large Language Models. Communications of the ACM, 67(2), 68–79. https://doi.org/10.1145/3624724 Simion, M., & Kelp, C. (2023). Trustworthy artificial intelligence. Asian Journal of Philosophy, 2(1). https://doi.org/10.1007/s44204-023-00063-5 Smith, B. C. (2009). Speech Sounds and the Direct Meeting of Minds1 (pp. 183–210). Oxford University Press. https://doi.org/10.1093/acprof:oso/9780199282968.003.0009 Solum, L. B. (1992). Legal personhood for artificial intelligences. North Carolina Law Review, 70(4), 1231–1287. Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H. T., Jin, A., Bos, T., Baker, L., Du, Y., Li, Y., Lee, H., Zheng, H. S., Ghafouri, A., Menegali, M., Huang, Y., Krikun, M., Lepikhin, D., Qin, J., & Le, Q. V. (2022). LaMDA: Language models for dialog applications. arXiv. https://arxiv.org/abs/2201.08239. Twerski, A. D., & Henderson, J. A. Jr. (2009). Manufacturers’ liability for defective product designs: The triumph of risk-utility. Brooklyn Law Review, 74(3), 1061–1108. Van Es, K., & Nguyen, D. (2024). Your friendly AI assistant: the anthropomorphic self-representations of ChatGPT and its implications for imagining AI. AI & SOCIETY. https://doi.org/10.1007/s00146-024-02108-6 Van Woudenberg, R., Ranalli, C., & Bracker, D. (2024). Authorship and ChatGPT: a Conservative View. Philosophy & Technology, 37(1). https://doi.org/10.1007/s13347-024-00715-1 Volokh, E. (2023). Large Libel Models? Liability for AI Output, 3 J. FREE SPEECH L. 489, 514 – 15, 522 – 26. Watson, D. (2019). The Rhetoric and Reality of Anthropomorphism in Artificial Intelligence. Minds and Machines, 29(3), 417–440. https://doi.org/10.1007/s11023-019-09506-6 Zhang, R., Li, H., Meng, H., Zhan, J., Gan, H., & Lee, Y. (2025). The Dark Side of AI Companionship: A Taxonomy of Harmful Algorithmic Behaviors in Human-AI Relationships. In CHI Conference on Human Factors in Computing Systems (CHI ’25), April 26–May 01, 2025, Yokohama, Japan. ACM, New York, NY, USA, 17 pages. https://doi.org/10.1145/3706598.3713429 Złotowski, J., Proudfoot, D., Yogeeswaran, K., & Bartneck, C. (2014). Anthropomorphism: Opportunities and Challenges in Human–Robot Interaction. International Journal of Social Robotics, 7(3), 347–360. https://doi.org/10.1007/s12369-014-0267-6 Acknowledgments We would like to thank Sabrina Coninx, Silvia DeConca, Celiné Henne, Guido Löhr, Jakob Ohlhorst, René van Woudenberg, and two anonymous reviewers of this journal for their feedback on earlier versions of this paper. Funding S/000057 - Stimuleringsbeurs - Dobler Decock. Author information Authors and Affiliations Contributions Both authors made essential contributions to this paper and were equally involved in the philosophical analysis. TD developed the proposal connecting legal case analysis with conceptual ethics and conceptual engineering, while DB analysed the relevant legal materials and First Amendment doctrine. Both authors jointly developed the application of these frameworks to the Garcia case. TD drafted the initial manuscript. Both authors contributed to subsequent revisions in response to feedback and reviewed and approved the final version of the manuscript. Corresponding author Ethics declarations Ethics approval and consent to participate Not applicable. Consent for publication The authors give consent to publish this material. Competing interests No competing interests. Additional information Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. About this article Cite this article Dobler, T., Bracker, D. Should the Legal Concept of Speech Include Chatbot Output in US Law?. Philos. Technol. 39, 174 (2026). https://doi.org/10.1007/s13347-026-01187-1 Received: Accepted: Published: Version of record: DOI: https://doi.org/10.1007/s13347-026-01187-1