A Network Approach to Public Trust in Generative AI

You’re listening to “A Network Approach to Public Trust in Generative AI,” by A. McIntyre, L. Conover, and F. Russo. Published in 2025.

UvA-DARE (Digital Academic Repository)

A Network Approach to Public Trust in Generative AI

McIntyre, Andrew; Conover, Lucy; Russo, Federica DOI 10.1007/s13347-025-00974-6 Publication date 2025 Document Version Final published version Published in Philosophy and Technology License CC BY

Link to publication

Citation for published version (APA): McIntyre, A., Conover, L., & Russo, F. (2025). A Network Approach to Public Trust in Generative AI. Philosophy and Technology, 38, Article 137. the linked source

General rights

It is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s) and/or copyright holder(s), other than for strictly personal, individual use, unless the work is under an open content license (like Creative Commons).

Disclaimer/Complaints regulations

If you believe that digital publication of certain material infringes any of your rights or (privacy) interests, please let the Library know, stating your reasons. In case of a legitimate complaint, the Library will make the material inaccessible and/or remove it from the website. Please Ask the Library: the linked source, or a letter to: Library of the University of Amsterdam, Secretariat, P.O. Box 19185, 1000 GD Amsterdam, The Netherlands. You will be contacted as soon as possible.

UvA-DARE is a service provided by the library of the University of Amsterdam (the linked source)

A Network Approach to Public Trust in Generative AI

Andrew McIntyre1 · Lucy Conover2 · Federica Russo2

Received: 7 February 2025 / Accepted: 7 September 2025 / Published online: 11 October 2025 © The Author(s) 2025

Abstract.

As generative AI becomes more deeply integrated into society, building public trust in this technology has emerged as a key challenge for policymakers. Existing approaches, such as the European Commission’s Trustworthy AI framework, largely seek to tackle this issue by offering comprehensive technical and legal measures for promoting a more trustworthy AI industry. However, this paper argues that such approaches are limited in scope and do not fully account for the social complexity of generative AI. As these technologies can now replicate modes of human communication and contribute to our collective knowledge, they cannot be simply considered products to be regulated. Rath-er, they exist as active social actors and AI policy should reflect this.

To better account for this social role, this paper develops a network approach to trust in AI inspired by philosophy of technology and Actor-Network Theory (ANT). This approach argues that trust emerges, first and foremost, from the material interactions between social actors involved in a vast and precarious network. In the context of generative AI, this material network extends far beyond the AI industry to include those various actors that are not directly involved in AI development but that nonetheless influence public trust. As such, this paper argues that the policy goal of establishing trustworthy AI, and thus promoting public trust in AI, is not solely a matter of promoting a more trustworthy AI industry. Rather, to achieve such a goal, more diverse policy solutions need to be devised on the basis of social interactions as part of a whole-of-society approach.

Primarily, this paper highlights that public trust in generative AI is influenced by those actors that play a key role in socio-political discourse such as political figures, media organizations, academic institutions and government bodies, among others. As such, public trust in generative AI is linked to trust in our information environment more broadly. To conclude, the paper argues that policymakers seeking to promote trustworthy AI must first seek to combat the current post-truth political crisis and restore public trust in democratic institutions.

Extended author information available on the last page of the article 1 Introduction

As artificial intelligence (AI) systems have become more advanced and increasingly integrated into society, public trust in these technologies has emerged as a critical challenge. Despite the growing adoption of AI, surveys and empirical studies of pub-lic attitudes in the US, UK and Europe consistently show that there remains signifi-cant public scepticism and distrust toward the technology. While the public broadly acknowledge and welcome the poten-tial benefits of AI in sectors such as healthcare and education, distrust toward the technology primarily stems from concerns over issues of data privacy, safety and security, and political interference. Research from the Ada Lovelace Institute (2022) suggests there is strong public support for stricter regulation of AI to address these concerns as existing laws and government initiatives are viewed as insufficient.

However, the issue of public trust in AI is further complicated by a broader distrust in those institu-tions responsible for the development, deployment and governance of this technology. Notably, government agencies, AI firms and social media companies are largely not trusted to uphold ethical standards and to use this technology responsibly. With many countries now seeking to capitalise on the potential benefits of AI, restoring public trust in the AI industry and regulatory landscape has thus become a priority for policymakers and regulators.

At the forefront of policy discussions on public trust in AI is the European Commis-sion’s High-Level Expert Group on AI (AI HLEG). In 2019, the AI HLEG proposed that ethical guidelines for the development, deployment and use of AI technologies be grounded in the foundational notion of Trustworthy AI. The stated goal of Trust-worthy AI was to provide citizens with legal certainty and confidence to accept these technologies in their daily lives, while also enabling technological innovation and responsible competition. Specifically drawing comparisons to the aviation or nuclear power industries, the AI HLEG argued that establishing public trust in AI cannot be achieved solely by improving the safety of these technologies themselves.

Rather achieving AI trustworthiness ‘requires a holistic and systemic approach, encompass-ing the trustworthiness of all actors and processes that are part of the system’s socio-technical context throughout its entire life cycle’. This holistic approach establishes key ethical requirements for business including, for example, human oversight, data privacy and transparency obligations. It further proposes both technical measures (e.g., testing and validation, quality of service) and non-technical measures (e.g., regulation, certification) aimed at ensuring the effective implementation of these requirements. The proposals put forward under the Trustworthy AI framework have since gone on to inform significant legislative and regulatory initia-tives around the world, notably the European Union’s AI Act and Digital Services Act.

While the framework’s holistic approach does account for the trustworthiness of all actors and processes involved in the AI lifecycle, in this paper we aim to take a step further in the conceptualisation of the AI ecosystem. Our own approach begins from the idea that trust in technology emerges, first and foremost, from the material interactions between social actors and that such interactions are extensive, complex and not bound within a specific domain or context. Following this approach, we argue that public trust in generative AI is reliant upon a far broader and more diverse range of actors that extends beyond simply those actors involved in the industrial development pipeline. As such, we propose that further policy interventions aimed at ensuring public trust and confidence in AI cannot be solely rooted in the industry-focused Trustworthy AI framework.

Whether or not achieving public trust in AI is itself an inherently or entirely desirable policy goal is a complex and nuanced issue that is beyond the scope of this paper. Instead, this paper seeks to elaborate on how the social environment around AI influences trust in the technology and how policy can react to this.

Primarily, the AI HLEG conceptualises trust in AI through an industrial lens and the proposed Trustworthy AI framework certainly succeeds in aligning differ-ent stakeholders (e.g., academics, legal experts, developers) toward a common, if ambiguous, goal. As Eugenia Stamboliev and Tim Christiaens note, ‘AI industry can read “trustworthiness” as a call for robustness, while ethicists and legal experts can simultaneously imagine that the document puts forward the agenda of making AI development more ethical and lawful’. However, the framework and other similar guidelines would seem to prioritize the interests of industry over ethical considerations. AI ethics is not in itself envisioned as a goal but, rather, amounts to a justification for industrial development and is thus reduced ‘to the supporting role of a fire extinguisher subservient to AI indus-try’s projects’ (Ibid., 15).

Indeed, members of the AI HLEG, Thomas Metzinger and Mark Coeckelbergh, have criticized the European Commission’s efforts for this exact reason. Subsequently, policy proposals deriving from this framework are primarily limited to establishing trust in the industrial pipe-line (e.g., corporate practices, development processes, legal restrictions on products, market competition). Meanwhile, they seem to sideline or overlook the extent to which AI technologies have become deeply integrated into our social lives and how this social integration influences trust and trustworthiness.

This deep social integration is most evident in the arrival of generative AI. Having been trained to recognise and replicate the underlying patterns in massive datasets, these technologies are capable of producing original human-like communications (e.g., text, visual art, music). These AI-generated outputs then contribute to our col-lective knowledge and discourse by building new socio-political narratives, altering our interpretations or understanding of events, and shaping our values through dis-cussion and argumentation. While the argument developed in this paper applies to generative AI broadly, the discussion that follows is primarily focused on large lan-guage models (LLMs) such as ChatGPT. This is because LLMs illustrate the social nature of generative AI particularly well as they enable conversational interactions with users.

Due to this social nature of generative AI, we argue that these technologies cannot be simply considered products or services to be addressed in the same way as aircraft or power plants. Rather, they are social actors that exist as active participants within our broader social practices and that interact with numerous other social actors that are not directly involved in the industrial pipeline. This includes, for example, politi- cal figures, online communities and media organisations. Following this perspective, this paper will argue that public trust in AI is not solely dependent upon the perceived trustworthiness of the AI industry but is also influenced by the perceived trustworthi-ness of the broader information environment in which AI operates and that informs its outputs.

As such, we argue that the question of trust in AI is becoming ever more complicated and that Trustworthy AI policy frameworks focused on the industrial pipeline are not sufficient.

To begin, Sect. 2 presents a brief overview of the philosophical debate around trust in AI, highlighting how the traditional concept of interpersonal trust has been compli-cated by technological developments. This section then goes on to discuss how trust in AI has since been reconceptualised through the lens of socio-technical systems theory, a position we seek to develop upon in this paper. In Sect. 3, however, we argue that the emergence of generative AI as social actors marks a significant change in the nature of our information environment. This is due to the fact that our socio-technical systems are increasingly populated by unpredictable technologies that contribute to discourse and collective knowledge in ways similar to human beings. To better account for this social role of generative AI and how it influences trust, in Sect.

4 we develop a network-based philosophical framework that draws upon the philosophy of information and Actor-Network Theory (ANT). In this approach, trust in AI is reconceptualised as something that emerges, first and foremost, from the material conditions of a network of social actors. It is precarious and contingent upon the structuring, ordering and stability of this material network. In Sect. 5, then, we briefly discuss the application and policy consequences of this network approach to trust in AI, while emphasising that public trust in AI is deeply entwined with the current post-truth crisis. We conclude that for policymakers seeking to develop trustworthy AI systems and improve public trust in AI, it is first necessary to establish a more trustworthy information environment and to restore public confidence in democratic institutions and processes.

2 Approaches to Trust in AI

While establishing public trust in AI technologies has become a key goal of AI policy and regulation around the world, the very notion of trust in AI remains a disputed concept in philosophy. This controversy largely stems from the argument that apply-ing trust to any technology is nonsensical as trust is conventionally defined as some-thing characteristic of interpersonal relationships exclusively between human beings (e.g., Metzinger, 2019, Ryan, 2020). While we will elaborate on this argument in more detail momentarily, it is first important to briefly note that the notion of inter-personal trust is itself a multi-faceted and complex concept that is subject to debate.

Conventionally, we may begin from an understanding of interpersonal trust as a purely rational attitude in which a trustor expects a trustee will behave in a certain way (e.g., fulfil a stated commitment, provide accurate information, complete a task correctly). This expectation is based on an assessment of the trustee’s past behaviour, as well as their perceived expertise and competence. Trust is also not absolute but rather it comes in degrees and is context specific. I may trust someone to complete one task and not another, while I may trust someone else more to complete the same task depending on my assessment of their competence. However, as philosophers such as Annette Baier and Karen Jones have argued, trust cannot be willed by ratio-nal explanation alone and, indeed, often precedes rational explanation.

Furthermore, for one to believe the trustee will act honestly and in the way expected, it is also necessary to consider the trustee’s motivations and to be will-ing to place oneself in a vulnerable position by delegating responsibility to another person. As Baier states: ‘one leaves others an opportunity to harm one when one trusts, and also shows one’s confidence that they will not take it...Trust then, on this first approximation, is accepted vulnerability to another’s possible but not expected ill will (or lack of good will) toward one’. As such, interpersonal trust cannot be described as a purely rational assessment of a trustee’s behaviour. Rather, it entails an affective dimension in that the trustor places confidence in the trustee’s goodwill toward them and also a normative dimension in that the trustor expects the trustee to fulfil moral obligations.

Interpersonal trust is further complicated by the relationship between trust and other concepts such as belief, authority, and reputation. Notably, Fricker (2007) has developed the concept of epistemic injustice to explain how pre-established prejudices and/or biases in society will lead to misplaced or excessive trust and distrust. As an example, Fricker highlights cases of sexual assault wherein misogynist bias may lead officials to excessively distrust female accusers and excessively trust male defendants without or in spite of rational explanation. This line of thought has been further explored by theorists Hawley (2019) and Medina (2020). In relation to generative AI specifi-cally, Simon (2025) has further discussed how deceptions concerning the ontology, capacity and function of these products may lead to misplaced or excessive trust in the technology.

It is not simply that AI developers make exaggerated claims or use misleading metaphors but, rather, this deception is inherent in the technical design choices, user experience and ambiguous processes behind AI. Indeed, Natale (2021) has argued that this deceit is and always has been fundamental to how AI technolo-gies are developed, marketed and culturally understood. Regardless of deception, however, it is primarily the normative and affective dimensions of interpersonal trust discussed above that would seem to exclude objects and technologies from being considered as legitimate participants in trust relationships.

Previously, there has been much philosophical discussion concerning the nature of trust in relation to objects, technology and artificial agents. However, since the publication of the Trustworthy AI framework and its subsequent adoption into the AI Act, there has been a flurry of academic activity and debate concerning the stated goal of establishing public trust in AI. In particular, there has been significant debate in the field of medical ethics wherein many theorists have argued that the concept of trust cannot and should not be applied to objects and technologies, including AI sys-tems. Notably, Ryan (2020) argues that, while AI systems may meet the requirements for a purely rational account of trust, these technologies do not fulfil the normative and affective conditions.

This is because AI systems fundamentally do not hold motiva- tions or a sense of moral responsibility for the task they are entrusted to carry out. Furthermore, authors such as von Eschenbach (2021) and Estella (2023) have argued that AI technologies, particularly those that utilize deep learning, cannot be said to sufficiently fulfil the rational conditions for trust either. The “black box” nature of these systems means that their internal processes are opaque and their behaviours are entirely unpredictable. While this may be too strong a statement, it remains true and well documented that these systems are largely opaque and unpredictable, as we discuss in more detail in Sect. 3.

Not only does AI fail to fulfil the conditions of interpersonal trust, applying the concept of trust to AI falsely anthropomorphizes the technology and further compli-cates the notion of accountability. As Freiman (2023) has argued, AI systems fun-damentally lack the necessary qualities for trust relationships (e.g. goodwill, moral agency, responsiveness). By attributing trust to AI, we wrongly assign moral respon-sibility to a technology that cannot be truly held accountable. Meanwhile, we obscure the roles and responsibilities of those human actors involved in the development and deployment of AI. For Freiman and many of the critics discussed above, then, the very concept of Trustworthy AI is not only ambiguous but also philosophically incoherent and problematic as it shifts focus away from humans.

Rather, these crit-ics argue that policy discussions around trust in AI should instead be reframed using technologically-specific terms such as “reliable AI”. Furthermore, they argue that greater emphasis should be placed on ensuring that those organisations and individu-als involved in the development and deployment of AI are trustworthy.

In opposition to these theorists who argue against the very notion of trust in AI, others have sought to reconceptualize trust relationships to account for technologies in different ways. For example, some have argued that trust in AI is a “simple” or functional form of interpersonal trust that is solely rational. Meanwhile, others have argued that the normative expecta-tions of an AI system differ depending on the norms of the social situation in which the technology is deployed. These arguments challenge the idea that AI cannot fulfil the necessary conditions for trust by seeking to loosen our definition of trust in a way that accounts for AI. However, they do not fully address Freiman’s central concern that attributing trust to AI risks obscuring human accountability.

Oth-ers have argued that the technology itself is not a subject of trust but rather the object through which a normative trust relationship is established between AI developer and user. Notably, philosopher Nickel (2022) proposes a discretionary account of trust wherein clinicians can justifiably rely on AI systems by trusting the ethical responsibilities and normative commitments of those people involved in the development and deployment of the system. While maintaining Frei-man’s concern with human accountability, these approaches nonetheless recognise the importance of understanding human-AI interactions through the lens of trust relationships.

In this paper, we take a different approach to the issue of trust in AI. Rather than seeking to redefine the notion of trust to accommodate AI systems, we reconceptual-ise AI systems themselves. As will be discussed further in Sect. 3, such technologies cannot be understood as mere tools or objects through which humans interact with one another. Instead, AI systems are understood as social actors that exist within a vast, interconnected network of other diverse social actors. From this perspective, we argue that trust in AI is, first and foremost, reliant upon the social relationships and material conditions of this network. This network approach will be developed in Sect. 4. In developing this argument, we are not proposing that AI systems can be trusted in the same way as humans, nor are we seeking to resolve the question of whether AI is an appropriate subject of trust at all.

As is evident from the academic debate discussed above, determining whether or not trustworthy AI is a morally desirable or even philosophically coherent policy goal is a complex issue that lies beyond the scope of this paper. Rather, we wish to elaborate on how the social environment around AI influences trust in the technology and, in doing so, begin to identify key actors and processes that have been overlooked by current AI policy. This approach is initially rooted in socio-technical systems theory and develops upon the work of von Eschenbach (2021).

Socio-technical systems theory emphasizes the hybrid nature of organizational systems (e.g., public services, companies, information networks) in that the technical and social aspects of these systems are deeply interconnected. As such, technologies used in these systems are not simply considered intermediaries connecting different human agents but, rather, they play a significant role in defining the system’s behav-iour. In order to analyse the characteristics and behaviours of a socio-technical system, it is necessary to consider a multitude of different interconnected agents, both human and non-human, and their specific interactions within the system.

With reference to socio-technical systems theory, von Eschenbach argues that dis-cussing trust in AI through the conventional lens of interpersonal trust does not prop-erly encapsulate the nature of human-AI interactions. As he states ‘to conclude that trust cannot apply to AI because this class of artifacts are not persons is to oversimplify the case...we interact with AI in ethically significant ways with increasing power, prevalence, and, ultimately, vulnerability’ (von Eschenbach, 2021, 1618). Instead, as we also argue later, it is necessary to acknowledge that these interactions occur within a broader socio-technical system and that trust is a social phenomenon arising from these interactions.

As von Eschenbach argues, trust in AI ‘can only be understood in reference to the system as a whole, and each agent’s trustworthiness will be judged relative to the differences in roles, interests, and expertise’ (von Eschenbach, 2021, 1619). To illustrate, he points to the use of AI medical diagnostic equipment which exists within a broader socio-technical system that includes the doctor, patient, tech-nicians and developers. While these actors have different roles, their actions contrib-ute toward fulfilling a shared goal of treating the patient efficiently and effectively.

As such, a “web of trust” is established in which ‘the patient, as the end user, will trust the doctor based on his judgement about her expertise, reliability, and whether she is acting on patient’s behalf’ and ‘with respect to the black box technology, the patient then would be willing to trust the judgement of the physician, who in turn trusts the judgement of the technician’ (Ibid.). Therefore, the patient may not be able to trust the AI technology in itself but they can trust the use of such technology within this particular socio-technical system that is geared toward an express collective goal. As von Eschenbach states, ‘from the perspective of the user (i.e., patient), the physician in this case becomes the proximate interface to the entire socio-technical system, of which the black box technology is an essential but component part and may remain opaque to the patient’ (Ibid.).

This reconceptualisation of trust in AI as trust in the socio-technical system of AI seeks to sidestep the challenges raised by the rational, normative and affective accounts of trust discussed above. Von Eschenbach does not seek to attribute moral responsibility directly to the technologies, nor does he seek to limit trust relation-ships to human beings. Instead, moral responsibility for the task entrusted to the AI is attributed to the socio-technical system as a whole, of which the technology is only a part of the decision-making process alongside human agents. The network approach that we propose in this paper develops upon von Eschenbach but further follows in the footsteps of Russo et al. (2023) who argue that trust pertains to processes rather than artefacts themselves and that this process encapsulates technical specifications, as well as norms and values.

We extend this idea of trusting the process through the concept of socio-technical systems but further highlight how the arrival of generative AI technologies has changed the nature of such systems so significantly that they may be dubbed “synthetic”, as we explain in the following section.

3 Trust in Synthetic Socio-technical Systems

Bisconti et al. (2024) have argued that the emergence of generative AI and the wide-spread integration of these technologies into our social environments is instigating a significant change in the dynamics of our socio-technical systems, rendering them synthetic in nature. In this section, we will briefly summarize this argument and out-line its consequences for the current discussion around trust in AI. Our approach to public trust builds upon this conceptualisation of synthetic socio-technical systems.

Traditional approaches in philosophy and in sociology of technology characterize information communication technologies as objects that mediate between human agents and their environments. In particular, postphe-nomenology has detailed the various ways in which technology mediates the rela-tions between humans and nature, as well as relations between humans. While the mediating role of technology is undeni-able, the mediation model of postphenomenology is insufficient to fully account for the deep social integration of digital technologies and generative AI in particular. A more appropriate starting point can be found in philosophy of information and the notion of “in-betweenness” which enables us to conceptualise tech-nology as more than simply a form of mediation.

As Floridi (2014) argues, technologies engage in different orders of relations. In first order relations, technologies stand “in-between” humans and nature; for exam-ple, sunglasses stand between the human and sunlight allowing us to see better. In second order relations, technologies stand between humans and other technologies; for example, a remote control stands between the human and a television allowing us to change channels. Finally, in third order relations, technologies stand between other technologies themselves, seemingly excluding humans from the chain of rela-tions, as it happens, for instance, in the internet of things. However, we emphasise “seemingly” here as humans are always involved somewhere in this chain, even in these third order relations.

Developing on this characterization of information communication technologies, Bisconti et al. (2024) argue that a specific form of “in-between” relations is emerging with generative AI.

When using a pen to write a letter or a messaging app to send a text, for example, we are not traditionally observed to be interacting with the technology as a social actor in itself. Rather, we are interacting with other human social actors through the technology; this is the postphenomenological perspective. However, the same cannot be said for generative AI. Generative AI technologies cannot be wholly described as mediators or as something that simply stands in between humans and nature or other technologies. Unlike previous technologies, generative AI has come to the fore-ground of sociality and qualifies as a social participant that we interact with in a simi-lar way to human beings.

For example, with only some prompting, AI models such as ChatGPT can (semi-autonomously) produce entire essays and articles that explicitly detail convincing and complex arguments, while also implicitly articulating specific social values that are unintended by the prompter. While one uses a pen as a mere instrument for writing, it seems more appropriate to say that we now write with ChatGPT.

Certainly, these technologies lack any conscious awareness of the meaning of their outputs in the same way as humans and have been rightly described as “stochastic parrots”. It is also true that they are incapable of creating these outputs entirely independently. However, the degree of autonomy that generative AI technologies have in the production process is so significant that it is inaccurate to say that these technologies behave as mere message deliverers or facilitators of communication that humans interact through. Instead, we are beginning to interact with these technologies as fully developed social actors. Our understanding of social actors in this context builds on accounts of actors and agency expressed in the phi-losophy of Latour (2005) and philosophy of information.

Following Latour, we take the idea that social actors can be of any nature in any given context including, but not limited to, human beings, technical objects, natural phenomena, institutions, and values. All are understood as legitimate and valid social actors. Meanwhile, follow-ing the philosophy of information, we take a very minimal concept of agency. One that does not presuppose intentionality and that is, instead, mainly characterized by interactivity, autonomy and adaptability.

As Bisconti et al. (2024) explain, reconceptualising generative AI models as social actors in this way reveals consequences for meaning making in society. While these technologies do not understand or intentionally express meaning, they are capable of processing information and actively participate in the co-production of semantic artefacts. This is a concept deriving from the philosophy of information that refers to any object, whether concrete or abstract, that carries information and conveys mean-ing. This includes texts and images but also more abstract objects such as a piece of code or a scientific theory. Generative AI co-produce semantic artefacts in various forms such as text responses and artificial images but also in the way their outputs contribute to or invent new narratives, offer new interpretations of facts and events, and perpetuate specific social values.

This capacity to create semantic artefacts means that generative AI models qualify as what Russo (2022) refers to as poietic agents; social actors that display epistemic agency and are capable of altering the dynamics between other actors and their environments.

The arrival of generative AI thus marks a significant shift in the way we interact with technology and the role that technology plays within our social environment. These changes are pervasive across society. For instance, AI chatbots such as Chat-GPT are being used for routine tasks in everyday life, as customer-facing advisors in private business, as automated writers in journalism and academia, and as interac-tive instructors in education. Beyond text generation, generative AI models are also being deployed in a variety of roles and in a diverse range of social settings. For example, the production of AI-generated art, the manipulation of images for political campaigning, and the introduction of AI-driven social robots as automated health-care assistants.

With this increasingly pervasive integration of generative AI into our social environments, such technologies have become social actors in their own right and cannot be simply isolated to a single static context. Through the co-production of semantic artefacts in various forms and in various contexts, these technologies are capable of influencing our collective knowledge.

Russo (2022) emphasises that our collective knowledge is composed of seman-tic artefacts that are themselves the products of an epistemic collaboration between poietic agents that may be both human and non-human. As such, collective knowl-edge is always contingent upon the specific configuration of relations within a socio-technical system. This includes relations between the information environment and social actors, as well as those relations between social actors themselves. As poi-etic agents, generative AI models can significantly alter the relational dynamics of a socio-technical system and thus shape our collective knowledge similarly to human beings. The arrival and widespread uptake of generative AI means that the number of poietic agents involved in socio-technical systems is drastically increasing.

More crucially, however, we propose that human beings and generative AI are not merely equal and legitimate social actors within socio-technical systems. Rather, they also interact in unprecedented ways and that the semantic artefacts that are co-produced from their epistemic collaborations are fundamentally different from those produced solely by human actors.

This fundamental difference primarily arises from the probabilistic and opaque nature of generative AI processes. Unlike rule-based systems, these AI models learn recurring data patterns from a massive training dataset and generate outputs based on statistical probability, rather than a set logical reasoning. This is why AI-generated outputs may appear highly fluent and persuasive, yet still contain factual inaccuracies, biases or a lack of contextual understanding. However, the sheer scale and complexity of this training process makes it difficult for users and developers to fully understand the internal decision-making that led to a specific output. Certainly, it is possible to broadly anticipate the type of output an AI model may generate in response to a spe-cific prompt. However, the exact output remains variable and inconsistent.

Indeed, even the same prompt can yield different responses. Discussing this unpredictability in relation to authorial agency, McIntyre (2022) characterized AI-generated outputs as the products of a chaotic production process in that they are neither entirely ran-dom nor entirely determinable. While the human user may provide a prompt with the intention of producing a specific outcome, by deferring labour to the statistical pro- cesses of the AI model the prompter cannot be said to predict or determine the final output itself. Consequently, while users and developers may have a general sense of what the model might produce, the exact outputs remain unstable and unpredictable. Furthermore, semantic artefacts produced by generative AI are not only unexpected but can significantly differ from artefacts included in the training data.

The chaotic nature of generative AI introduces a new dimension of variability, unpredictability and artificiality to semantic production. In order to account for this fundamental change, we follow Bisconti et al. (2024) in re-labelling these new socio-technical systems as synthetic as they can no longer be recognized as purely human social systems. However, these systems are synthetic in a further sense of the word as generative AI technologies are often not bespoke or task-specific tools confined to a single context. Rather, these technologies are multi-purpose and utilized in dif-ferent ways across society. Consider that ChatGPT has been deployed across mul-tiple different societal contexts including education, media, advertising, academia and politics. A single generative AI model can thus play a pivotal role in multiple disparate processes simultaneously.

This means that several socio-technical systems are entangled together and all are impacted by the functioning of the technology (e.g., system failure, malfunction, inherent bias). Furthermore, it must be made clear here that what is often referred to as “AI” is not a singular, defined technology but, rather, a broad and diverse collection of computational techniques, systems and applications with varying levels of complexity, autonomy and versatility. In addition to generative AI models, this broader landscape of AI includes systems for natural language pro-cessing, computer vision and other domain-specific applications. This only further emphasises the deep integration of these technologies within our socio-technical sys-tems. With such a drastic change in the nature of our socio-technical systems comes significant consequences for the notion of trust in AI.

Returning to von Eschenbach, trust in AI can be reconceptualized as trust in the overall socio-technical system of which AI technologies are simply a component alongside human actors. In this view, the moral responsibility for tasks conducted by these technologies is attributed to the socio-technical system as a whole. However, von Eschenbach conceptualizes AI technology as a tool that is opaque and unpre-dictable but that largely serves a limited and pre-defined role (e.g., medical diagno-sis) within an otherwise traditional socio-technical system dominated and driven by human actors. In contrast, synthetic socio-technical systems are densely populated with generative AI technologies that behave as social participants akin to human beings and interact in unprecedented ways. As such, establishing a web of trust between actors becomes a far more complex matter.

Given the versatile capabilities and widespread availability of this technology, multiple different generative AI models may be used in a single synthetic socio-tech-nical system for a variety of essential tasks. This may include activities such as report writing, translation, information retrieval, data visualization, email communication and idea generation. As mentioned, however, these models are unpredictable in the sense that their outputs are highly variable and inconsistent. Furthermore, these out-puts may also contain inaccuracies, biases and contradictions. A single contradictory report or inaccurate translation may not be significantly disruptive to the system as a whole.

However, if we are to use multiple AI models for numerous essential tasks, then we introduce a considerable degree of inconsistency and variability to the infor-mation being circulated within the system which may, in turn, lead to discordance and confusion. Ultimately, this confusion may undermine the shared goal of the sys-tem and encourage a breakdown in trust between the various actors involved.

The issues discussed above mean that it is becoming increasingly difficult to assess the trustworthiness of synthetic socio-technical systems let alone develop coherent strategies aimed at fostering public trust in such systems. In order to better account for the increasingly complex social role of generative AI, we propose taking a network approach to considering trust in AI that is inspired by the Actor-Network Theory (ANT) advanced primarily by theorists such as Callon (1986), Law (1992) and Latour (2005). Such an approach enables us to expand the scope of our analysis beyond the socio-technical systems involved in AI development and deployment to, instead, consider how a more diverse cast of social actors and interactions influence public trust in AI.

4 Relational Materiality of Trust

Marking a significant departure from traditional sociological theory, ANT scales down the influence of abstract concepts such as rigid social structures and social forces. Instead, within ANT, any and all social activity is conceptualized as a con-stantly shifting network of concrete relationships between social actors. Within ANT, the notion of a “social actor” does not solely refer to a human being. It further encom-passes a broad spectrum of entities including objects, animals, texts, technologies and institutions, all of which continually interact with one another in a flat non-hierarchi-cal network. As the interactions between these social actors change, so too does the composition and dynamics of the overall network change.

As such, the boundaries of a network are fluid and never fixed and any social actor, be they human or non-human, may take on a more active role and come to instigate significant change within the larger network. In emphasizing the fluidity and materiality of social relationships and further recognising the role of non-human social actors, ANT presents an appropriate and convenient starting point for our discussion on trust in synthetic socio-technical systems. However, there are limitations to the application of ANT in this context and so it is necessary to first distinguish our own approach from that of ANT.

Firstly, it should be noted that ANT is not intended as a strict and consistent the-ory but, rather, a flexible and evolving approach to social complexity. Even leading proponents such as Latour and Law have themselves problematized and critically reflected on the framework, highlighting its limitations and ambiguities. Notably, the non-hierarchical approach of ANT presents all social actors, be they human or non-human, as being equally capable of shaping the net-work in which they exist. However, this apparent flattening of the network blurs the distinction between different social actors and may neglect significant material characteristics and capabilities, as well as normative considerations. While we main-tain that all social actors are alike in that they have agency to influence the activity of the network, we also wish to emphasise the distinctions between different social actors.

Different social actors display different inherent dispositions i.e., material and relational tendencies that presuppose or enable specific kinds of social interactions and relationships. In the previous section, we highlighted that generative AI might be considered as a poietic agent in that they are capable of processing information and thus actively participate in the production of semantic artefacts. It is this unique disposition that distinguishes generative AI from previous technologies and that ulti-mately influences trust, as will be discussed.

While we do not pursue a strict ANT approach (inasmuch as ANT can be strict), we do draw upon the ideas of ANT as a starting point as it allows us to better account for the increasingly significant role that generative AI plays within our social inter-actions. Furthermore, it enables us to expand our discussion to consider the role of social actors that are not directly involved in AI development and deployment but that nonetheless influence public trust (e.g., media organizations, policymakers, govern-ment institutions). Particularly, those actors involved in the production of semantic artefacts, as will be elaborated on in this section. Primarily, we draw on ANT’s focus on the material configurations of social networks and how these material configura-tions become ordered and stabilised.

For theorists such as Latour, Callon and Law, ‘all phenomena are the effect or the product of heterogeneous networks’ composed of both human and non-human materials. As a notable example, these scholars argue that sci-entific facts are not simply revealed as fundamental truths through rational inquiry but rather they are social products. As social products, scientific facts are materially constructed through the interactions of ‘heterogeneous bits and pieces – test tubes, reagents, organisms, skilled hands, scanning electron microscopes, radiation moni-tors, other scientists, articles, computer terminals, and all the rest’ (Ibid., 381). These socially-constructed scientific facts do not simply exist as abstract concepts but, fur-thermore, always exist in a variety of material forms; ‘it comes as talk, or conference presentations. Or it appears in papers, preprints, or patents.

Or again, it appears in the form of skills embodied in scientists and technicians’ (Ibid., 381). Following this line of argument, we propose that trust is not just or solely some characteristic of interpersonal relationships and is not something limited to human-human interac-tions. Rather, trust is also a product of the material conditions and social interactions between numerous social actors interconnected in a network.

In this context, we take a liberal understanding of “materiality”. In our view, mate-riality does not merely or solely refer to worldly objects but further extends to any concrete situation, condition or practice that plays a role in establishing and/or main-taining trust relationships. For example, in court proceedings, a witness may be asked to publicly swear upon a sacred text or make a secular pledge to tell the truth. It is not that the text itself or the words of the pledge compel the witness to be truthful but that this material practice itself invites others (e.g., judge, jury, lawyers) to accept their testimony as truthful. Thus, this material practice fosters the necessary conditions for trust.

Trust relationships arise, first and foremost, from the material conditions of our interactions and these conditions extend far beyond the immediate situation itself (e.g., swearing upon a sacred text). As Law argues, our interactions with other social actors are almost always ‘mediated by a network of objects - the computer, the paper, the printing press. And it is also mediated by networks of objects-and-people, such as the postal system...these various networks participate in the social. They shape it’ (Ibid., 382). When considering why we do or do not trust another person, object or institution, it is therefore necessary to step back from the immediate interaction itself to instead consider the far wider material network of social actors that frames this interaction.

To illustrate this network approach to trust in the context of technology, let us now turn to the simple example of using a calculator. We accept the calculator’s output as correct not because we trust the device alone but because we trust a larger material network of social actors that has led to or frames this output.

To trust that the calcula-tor’s output is correct, we must also trust that (i) the people involved in the supply chain (e.g., manufacturers, transporters, vendors) have not altered or damaged the device, (ii) the human and mechanical labourers constructing the device have done so according to accepted standards, (iii) the standards of regulatory bodies are prop-erly enforced, (iv) the original designers have competently utilised scientific con-cepts to create the device, (v) the educational systems and resources (e.g., textbooks, equipment) that trained these designers were accurate, and finally but fundamentally (vi) the scientific research behind the calculator (e.g., electronics, arithmetic) is built upon rigorous investigation that adheres to the scientific method.

This is, of course, only a limited description but one that demonstrates how even an apparently simple trust relationship is reliant upon a far more complex and diverse network than we might initially expect. To be clear, our point here is not that anyone using a calculator will be consciously making such statements. Rather, this is a conceptual unpacking of trust that is grounded in a network approach and that emphasises material relations.

Before developing this argument further, it is important to first distinguish our net-work approach from von Eschenbach’s account of trust in socio-technical systems. Let us recall the example that von Eschenbach describes in which trust in the results of AI medical diagnostic equipment arises through a web of trust among various human and non-human actors (e.g., patient, doctor, technicians, developers, organisations). For von Eschenbach, trust in the AI diagnostic tool is mediated by this broader socio-technical system wherein users place trust in the technology because they already trust the surrounding people, organisations and infrastructures. In contrast, our net-work approach argues that trust among these actors emerges, first and foremost, from the relational and material entanglements that constitute their interactions.

As such, this web of trust that von Eschenbach describes is not a precondition for trust in AI but, rather, a product of a much broader network of social interactions involving both human and non-human materials. This includes not only those actors named by von Eschenbach but also other actors such as hospital administrative systems that manage patient flows, procedures and processes that structure patient-doctor consultations, educational institutions that accredit doctors and technicians, and regulatory bodies that certify AI systems as reliable according to their own standards. It also further extends to the material configurations involved in the production and dissemination of knowledge more broadly; this is a point we will return to in more depth shortly.

Trust, in our view, is the product of these various material configurations and emerges through a process that Law refers to as punctualization.

If trust is the product of a vastly complex network, the question then arises: why do we not perceive it as such? And why do we not actively scrutinize these material interactions to ensure that they are trustworthy? When using a calculator to solve a problem, we usually accept that the numbers that appear on the screen are not randomized outputs but are, in fact, the correct answer to our problem. With every calculation, we do not question the calculator supplier as to how carefully the device was handled before selling it on or call the manufacturer to ensure they have adhered to production standards. Nor do we inspect the device’s circuitry for faults or re-read our electronics textbooks to reassure ourselves that the theory behind the device is valid. Often, we do not even double-check the calculation by working it out ourselves with pen and paper.

The reason being that the calculator exists as what Law refers to as a simplified effect or punctualization of the broader network, stating: if a network acts as a single block, then it disappears, to be replaced by the action itself and the seemingly simple author of that action. At the same time, the way in which the effect is generated is also effaced: for the time being it is neither visible nor relevant. So it is that something much simpler – a working television, a well-managed bank or a healthy body – comes, for a time, to mask the networks that produce it.

While there may have been a time when the calculator’s outputs were met with scep-ticism and the designer’s abilities were questioned, this material network surrounding the calculator has since become ordered and maintained with all actors seemingly acting as one. In other words, the production and use of calculators for quickly solv-ing mathematical problems has become routine, structured and commonplace within society. As such, the device no longer appears as an odd and messy electromechani-cal thingamajig cobbled together by a jumble of humans, machines and scientific concepts. Rather, it appears as a familiar tool that we trust to solve mathematical problems. Trust in technology is a product of this punctualization. It arises when a network of heterogeneous actors becomes materially stabilized and their interactions become ordered such that they appear as a unified, coherent system.

Trust does not pre-exist this ordering but rather arises from it. We trust the outputs of a calculator because there is an established coherence and order in the material configurations that surround this technology. This is not yet the case with generative AI, if ever it will be.

While AI technologies have been around in one form or another for some time now, the generative AI products that have begun to populate our socio-technical sys-tems today are still relatively recent. As such, the network surrounding these tech-nologies has yet to be ordered or maintained in the same way as the calculator and so its individual actors and their interactions remain visible and exposed to scrutiny. The controversies of AI developers are still visible to the public, as are the flaws and limitations of those regulatory bodies struggling to keep pace with a rapidly changing landscape. Meanwhile, users of generative AI often find that their outputs fail to meet expectations or achieve intended goals, and may not even be relevant to their initial prompts. These underwhelming results highlight the inherent unpredictability and lack of comprehension of these systems.

Furthermore, there are increasing efforts to make the public aware of the theoretical, ethical and practical challenges of using this technology (e.g., bias, inaccuracy, unpredictability). These often come in the form of communications from politicians, academics (including this paper and its authors), technology companies, and media figures that are, at times, highly speculative and hyperbolic. As the network is new, the material configuration of its constituent actors and the interac-tions between them remain so disordered that it appears incoherent and unstable. As such, the technology itself appears as an ill-defined and untrustworthy product rather than a familiar and reliable tool.

From this perspective, the current public distrust of AI is not simply due to its novelty but it is also, to an extent, a result of the dysfunc-tion and distrust that already surrounds the production, deployment and regulation of the technology. It is also not clear that people will simply come to trust AI in time if this dysfunction subsides and the AI industry becomes more structured and regulated.

Notice, however, that punctualization is not the inevitable conclusion that social networks naturally converge toward but rather a transitory state of stability. As net-works are fluid and continuously change, a state of ‘punctualization is always pre-carious, it faces resistance, and may degenerate into a failing network’ wherein the heterogeneous messiness of material reality is once again laid bare to be scrutinized. While we trust the output of a calculator now, a significant change in the supply chain (e.g., reports of manufacturers not meeting regulatory standards) may destabilize the network as a whole. Thus causing us to question the trustworthiness of these actors and potentially distrust the calculator’s outputs.

This is admittedly an unlikely scenario with the calculator but there have been numerous high-profile cases in the past wherein a once-trusted technology rapidly fell out of public favour due to shifts in the surrounding social network. A notable example is asbestos. For decades, asbestos was widely accepted and trusted as a reliable building material with a structured network of industrial actors and regulatory bodies initially firmly aligned on its effectiveness, reliability and safety. However, scientific reports linking asbestos exposure to serious long-term health problems destabilised this net-work, resulting in widespread restrictions on its use and public condemnation of the material. Not only did these reports undermine public trust in the material itself but they also caused distrust in the construction industry and those regulatory bodies responsible for oversight.

Furthermore, history is littered with other technological devices that were not trusted at all and were rejected by the public even before punctualization could occur. Take, for example, the Google Glass augmented reality headset launched in 2014 that allowed wearers to record their surroundings, receive notifications, and use the internet through voice commands. While it technically functioned as intended, the device’s recording capability was viewed as intrusive by the public and consumers found little practical use for the device in everyday life, eventually leading to its dis-continuation in 2015 due to poor uptake. If the material configurations of its surrounding network remain disordered and unstable, generative AI may very well follow a similar path and become yet another technology doomed to economic and innovative failure.

The apparent dysfunction surrounding generative AI technologies is not, however, the only barrier to achieving this punctualization.

As discussed in the previous section, generative AI technologies are capable of producing syntactically coherent semantic artefacts that contribute to our collective knowledge. Consider that ChatGPT can provide lengthy summaries of historical accounts, government policy documents, and cultural texts, as well as explanations of scientific, economic and philosophical concepts, among many other things. Fur- thermore, it can generate entirely new and synthetic historical accounts, policies, texts, and concepts in collaboration with human users. These AI-generated seman-tic artefacts in turn derive from other semantic artefacts (e.g., government records, research publications, journalistic articles) that have been created by other social actors (e.g., economists, historians, academics, politicians, media organizations).

This further extends to the vast amount of content created by internet users and the general public more widely that is circulating online. However, these technologies do not simply funnel semantic artefacts from one social actor to another but, rather, they synthesise these artefacts into an entirely new output, in ways that may not be neutral or objective.

Even if these sources have been accurately represented, these AI-generated outputs are a somewhat muddled mixture of complex socio-political narratives, arguments and ideas that may be highly speculative, loosely defined, and open to interpretation or dispute. These outputs are not simple logical assertions that can be deemed correct or incorrect like the calculator’s numerical outputs. Rather, they are the product of an epistemic labour that is not only conducted by the technology itself but that is widely distributed among these numerous other social actors involved in the production and dissemination of knowledge in society. We will illustrate this concept in more detail with an example in Sect. 5. Beforehand, however, we need to understand how trust is essential to this kind of distributed epistemic labour, as Boaz Miller and Ori Frei-man (2020) argue.

This notion of distribution is similar to the epistemic collaboration between generative AI and humans described in Sect. 3, though Miller and Freiman emphasise different aspects of this process.

Building on the work of philosopher Hardwig (1985), Miller and Freiman (2020) argue that the production of scientific knowledge should be understood as a form of distributed epistemic labour. As no single individual possesses all the necessary expertise, evidence or resources to independently justify a piece of knowledge, scien-tific researchers rely on the testimonies and semantic artefacts (e.g., papers, technical reports, datasets) of others. Scientific knowledge, then, is not solely the product of individual researchers themselves but rather the outcome of a collaborative epistemic labour across the broader scientific community. To believe that scientific products are accurate, it is necessary to trust the communications of others. As Miller and Freiman state, ‘trust is as fundamental to knowledge as epistemic justification, e.g. evidence.

Trust is the glue that binds researchers’ testimonies about products of their distributed epistemic labor into collective knowledge’.

While Miller and Freiman focus their discussion on scientific knowledge, we might apply the same framework to our discussion of generative AI and consider AI-generated outputs as the product of a distributed epistemic labour. As discussed, generative AI systems can address a vast range of topics in their outputs by draw-ing on and synthesising knowledge from numerous different sources. As such, these outputs are not solely the product of the technology itself but also arise from the labour of those various social actors whose communications are used to inform AI. This presents a significant challenge for trust in AI.

Not only is this epistemic labour distributed among numerous actors, precisely which actors are involved is continu-ally changing as we move from asking ChatGPT to explain the fundamentals of mac-roeconomics, to then requesting that it generate an essay on analytic philosophy, to then asking it to match medical symptoms with possible illnesses. To trust that ChatGPT’s outputs are correct in each case, it is necessary to also trust the diverse and evolving array of social actors from which these outputs have emerged. As such, the network of social actors surrounding generative AI expands far beyond simply those actors involved in the industrial pipeline and regulatory environment to include numerous actors involved in our information environment more broadly.

Notably, Bisconti et al. (2024) provide a more detailed description of the network of social actors involved in the production, circulation and reception of AI-generated content online. Beyond industrial actors, they highlight how social actors involved in content creation, online information networks, and media narratives are connected to and influence one another in a vast network. The sheer size, complexity and instability of the network surrounding generative AI makes punctualization even more difficult to achieve.

Following our network approach, it seems imprecise and misleading to simply ask whether or not AI systems and their outputs are trustworthy in themselves. This would seem to neglect and obscure the much larger and more diverse social network of which AI systems are only a part. Trustworthy AI policy initiatives aimed at indus-trial processes and regulation would only seem to address one aspect of this network, while other aspects that influence the trustworthiness of AI systems are overlooked. Notably, the integrity and trustworthiness of our information environment. To more accurately reflect this situation, it is necessary to disentangle this network to better understand its material conditions. In doing so, we can identify unstable connections, detrimental practices and disordered conditions that contribute to the dysfunction and distrust around AI.

From there, we may begin to develop diverse policy interventions to build more robust and trustworthy networks.

5 Fostering a Trustworthy AI Environment

In the previous section, we argued that trust relationships arise, first and foremost, from the material conditions of our social interactions and that these conditions are influenced by a vast number of diverse social actors. Through this network approach we highlighted the social complexity of human-technology trust relationships and further elaborated on the nature of the network relations that form around generative AI technologies. Notably, we highlighted that generative AI synthesise content pro-duced by numerous other social actors in order to generate their own original outputs that often deal with nuanced and debatable socio-political topics. As such, unlike previous simpler technologies such as the calculator, the trustworthiness of genera-tive AI and AI-generated outputs is directly influenced by a much larger and more diverse group of social actors.

In the context of AI policy, this is the main takeaway from our discussion.

Trust in AI cannot be reduced to trust in the technology itself and/or trust in those actors involved in the industrial pipeline, the approach taken by the Trustworthy AI framework and policies deriving from it. By continuing to emphasise regulation and standards aimed at governing the AI industry, these policy approaches continue to view the issue of public trust in AI solely through the lens of legal accountability.

This approach is limiting as it seems to overlook the social role of generative AI technologies and neglect the influence of the broader social environment of which AI is a part. In contrast, our network approach supports the view that AI policy can and should be devised on the basis of social activity and social interactions rather than legal accountability alone. Using this approach, we may begin to identify a diverse range of actors from across different societal domains whose activities none-theless impact the trustworthiness of generative AI and AI-generated outputs. Such an approach might encourage policymakers to develop more unique interventions as part of a broader whole-of-society approach to AI policy that builds upon and goes beyond the Trustworthy AI framework.

In the previous section, we argued that the network that forms around generative AI is often vast, unstable and continually changes depending on the specific use of the technology. This would seem to make practical policy development rather dif-ficult. To use this network approach in practice, then, it may prove beneficial to iso-late a variety of different examples of AI-generated outputs and analyse the material conditions and social relationships that influence the production, dissemination and reception of such outputs.

These examples might focus on common AI-generated out-puts (e.g., generated news articles, summaries of scientific texts, answers to medical questions), the settings or contexts in which these outputs are used (e.g., education, healthcare, journalism), or the media channels by which these outputs are distrib-uted to the public (e.g., broadcasting, online platforms, private communications). While each of these examples can only provide a static snapshot of this complex and fluid network, together they may provide a more detailed understanding of those key social actors influencing the trustworthiness of generative AI. A full case study of this kind is beyond the scope of this paper which has only sought to lay the philosophical groundwork for this network approach.

To briefly illustrate this approach, however, let us consider a simple example of an AI-generated output relating to a complex and controversial historical event. Namely, the UK’s withdrawal from the European Union in 2020 (Brexit) (see Table 1).

As a brief example to support this discussion, we have gathered outputs from four different AI chatbots that are free and readily available to users in the UK: ChatGPT, Google Gemini, Claude and DeepSeek. These chatbots were each provided with the same basic prompt: “summarize the impact of Brexit”. In response, each provided a brief and broad overview of the economic, social, political, cultural and legal impli-cations. Notable examples of these responses are shown in Table 1, while the full responses from each chatbot are provided in the Appendix.

Not only do these responses contain statements that describe concrete activities (e.g., regulatory changes, statistical decline in trade), they also contain statements that refer to nuanced discourse concerning the political aftermath (e.g., strained inter-national relations, increased societal divisions). Furthermore, they contain statements that refer to more abstract or speculative arguments (e.g., changing perception of Brit-ish identity, global standing). Consider, for example, statements such as “Prolonged uncertainty surrounding Brexit has discouraged business investment, hindering eco-nomic growth” (Google Gemini), “Seen as a blow to European integration, though the EU has remained resilient” (DeepSeek), and “The UK is seeking to redefine its global role post-Brexit” (ChatGPT).

While there is certainly some evidence to sup- port each of these claims, they remain only interpretations of a highly complex and politically polarizing topic. These AI-generated outputs cannot be considered simply neutral regurgitations of facts. Rather, they exist as semantic artefacts in themselves that convey a specific meaning and progress specific socio-political narratives. In this way, these chatbots contribute to the discourse around Brexit, our historical under-standing of the event and its impacts, and also the ongoing debate around UK-EU relations. The question is not whether or not such AI-generated outputs can be trusted as purely neutral assessments of the impact of Brexit. Rather, the question is whether or not these outputs can be trusted as arguments that are logically consistent, rooted in evidence, and not politically biased.

To address this question, it is necessary to take a step back and consider the broader network of social actors that frame and inform these outputs. Of course, this network includes those social actors directly involved in the development and deployment of these chatbots (e.g., engineers at OpenAI). However, as discussed in the previous section, these outputs are not solely the product of the technology itself. Rather they are a synthesis of other semantic artefacts produced by a variety of other social actors that may differ depending on the response.

When prompted to provide sources for the responses shown in Table 1, the chatbots specifically referenced communications released by think tanks (Centre for European Reform, UK in a Changing Europe, Centre for Economic Policy Research, Institute for Government), online blogs (Lon- don School of Economics blog, Investopedia), media sources (BBC, Financial Times, The Times, The Guardian), and UK institutions (Office for Budget Responsibility, Bank of England, His Majesty’s Revenue and Customs). As such, the responses shown in Table 1 are the products of an epistemic labour distributed among these various social actors. It should be noted, however, that these responses are highly variable, for the reasons discussed in Sect. 3. For these particular responses at this particular moment in time, these social actors become temporarily interconnected in the material network around generative AI.

As such, they influence the trustworthi-ness of these outputs. Having identified the various social actors at play, we must now consider what measures can be introduced to ensure the trustworthiness of the whole network including those actors outside the industrial pipeline.

Certainly, we can introduce technical and legal measures governing the develop-ment and deployment of these chatbots such as those proposed by the Trustworthy AI framework. For example, ensuring that the chatbot is trained to represent a plu-rality of sources from across the political spectrum to mitigate bias. Such measures would go some way to ensuring that the process by which the technology synthesises information is trustworthy and a somewhat balanced representation of the discourse. However, we do not believe this would be enough. Beyond the trustworthy synthesis of information, we require trustworthy sources in the first place. What constitutes a trustworthy source is a complex issue, particularly in our current information envi-ronment, and one we feel this Brexit example illustrates well.

The issue here is that the Brexit debate has never been a discussion of policy alone but rather one shaped by post-truth politics like so many other issues today. Not only is much of the discourse around Brexit driven by online disinformation but, crucially, it is characterised by a fundamental distrust of once trusted sources of information such as democratic institutions, political figures, academic experts and mainstream media organisations.

In the UK context specifically, Marshall and Drieschová (2018) identify that public trust in traditional institutions and public figures had been significantly eroded in the decades preceding the Brexit referendum. Primarily this is due to high-profile contro-versies and failures of the state (e.g., Iraq war, 2008 financial crisis, 2009 parliamen-tary expenses scandal). One might also add to this list, those scandals that brought once trusted mainstream media organisations into disrepute (e.g., BBC controversies, News of the World phone hacking). Alongside the rise of online disinformation, these historical events contributed to widespread public distrust in UK politicians, officials, journalists, academic experts and also in those institutions that formed the traditional system of governance as a whole.

As a result, during the Brexit referendum warnings of economic decline and socio-political unrest issued by these social actors were often characterised and dismissed as scaremongering rather than trustworthy argu-ments based on evidence and expert advice. Returning to the example above, then, the question arises: how can we expect users to trust these chatbots’ responses on Brexit are balanced and unbiased distillations of the evidence and discourse, when these outputs are based upon those very same sources that were so distrusted by the public in the first place?

Alongside measures aimed at the AI industry, it is also necessary to develop poli-cies that improve the trustworthiness of these other social actors. For example, let us hone in on the issue of public trust in politicians that was so damaged by the 2009 expenses scandal, wherein numerous elected Members of Parliament were found to be misusing public funds. An obvious and feasible policy response would be to imple-ment more robust and transparent systems governing politician’s conduct. Further-more, we could establish more rigid and enforceable standards for those serving in public life. Such standards would hold politicians more accountable for their actions and could thus improve public trust.

Members of the public may feel more confident that their elected representatives are not solely motivated by personal interests and that their statements on highly consequential policy issues, such as Brexit, are genu-ine evidence-based arguments rather than deceptive manipulation. This is not a new policy idea. Indeed, the UK parliament’s own Committee on Standards, (2021) has previously advised for tighter rules for Members of Parliament. Meanwhile, historian Rosenfeld (2018) has more broadly called for better ethical standards for public fig-ures as a means of combating post-truth politics.

However, we are arguing that both new and existing policies aimed at a diverse range of social actors (e.g., standards in public life, online safety, media regulation) could be recontextualised and repurposed as part of a more coherent strategy for promoting trustworthy AI by first fostering a more trustworthy society more broadly.

While we have not endeavoured to develop specific recommendations in this paper, we have presented our network approach that argues policy interventions should target a more diverse set of social actors and should be devised on the basis of social interactions. As such, we have laid the theoretical groundwork for more in-depth policy development in subsequent work. However, we will conclude by pro-posing that the issue of public trust in generative AI is intimately related to or reliant upon trust in our information environment more broadly. This is a point touched upon in the above Brexit example. As generative AI can now contribute to our collective knowledge, they have become an active part of our broader information environment, influencing and being influenced by the other actors operating within it.

This means that the policy goal of establishing trustworthy AI is hampered by the current post-truth political crisis that continues to undermine trust in our fundamental democratic institutions. Whether or not the notion of trust in AI is deemed a philosophically coherent and/or morally desirable policy goal, this link between AI and our infor-mation environment only underscores the urgent need for more substantial policy that properly addresses the post-truth political crisis. Notably, policy that tackles the spread of disinformation online, that holds post-truth actors accountable, and that seeks to counteract the erosion of trust in our political systems.

For policymakers aiming to establish trustworthy AI, we conclude that it is necessary to first take steps toward establishing a more trustworthy information environment and ensuring that the actors operating within it do so in trustworthy ways.

6 Conclusion

With this paper, we have sought to contribute to the ongoing discussion around trust in generative AI, a topic that intersects with multiple debates across the philosophi-cal, ethical, legal and regulatory fields. We began by examining the Trustworthy AI framework that has been at the forefront of the recent discussion of trust in AI and has greatly informed AI policy development. While we broadly agree with the holis-tic approach proposed by the framework, we believe that its scope is limited to the industrial pipeline and thus its policy proposals do not fully capture the nuance of the issue of trust in AI. Specifically, we note that generative AI technologies have come to take on a more significant role in our social activities and can contribute to our collective knowledge.

As such, trust in the technology has become complicated by the semantic complexity of its outputs, as well as their unpredictability, inaccuracy and bias. To better account for this social role of generative AI and to expand upon the Trustworthy AI framework, we have proposed a network approach as a means of analysing the issue of public trust in AI and developing new avenues for policy development from a social perspective.

Rooted in ANT, this approach situates generative AI not as a product but, rather, as a dynamic social actor connected with other actors within a broader environment. In contrast to the traditional notion of interpersonal trust, then, this network approach seeks to expand the discussion outward beyond the immediate interactions between user and technology to, instead, consider the wider material conditions of this interac-tion. In doing so, this approach accounts for a vast array of diverse social actors that are not involved in the AI industry but that nonetheless influence the trustworthiness of the technology. Notably this includes those actors involved in the production and dissemination of knowledge in society, as AI-generated outputs derive from and syn-thesise other semantic artefacts.

What this approach emphasises is that technology does not exist in a vacuum and that there is a need for policy to go beyond regulation solely aimed at the technology or industrial processes.

This new approach could encourage policymakers to devise policy on the basis of social activity and to utilise more diverse policies aimed at different social actors in order to promote public trust in AI. This approach does not necessarily require new policy but suggests that current interventions could be repurposed into a coherent strategy directed at establishing a more trustworthy social environment around AI, rather than the technology and AI industry alone. Primarily, we conclude that trust in AI requires trust in our information environment. In their pursuit of achieving trust-worthy AI and promoting public trust in AI, it is vital that policymakers first prioritise tackling the current post-truth political crisis as part of a wider whole-of-society AI strategy.

Appendix.

The following appendix contains the full responses from four different AI chatbots when provided with the same input prompt: “Summarize the impact of Brexit”. These include ChatGPT (GPT-4-Turbo), Google Gemini (2.0 Flash), Claude (3.7 Sonnet), and DeepSeek-V3. These systems were accessed in the United Kingdom of Great Britain and Northern Ireland on 15 April 2025.

ChatGPT (GPT-4-Turbo)

The impact of Brexit, the United Kingdom's withdrawal from the European Union, has been wide-ranging and complex. Here's a summary of the main effects:

Economic Impact

●

Trade Disruptions: UK-EU trade now faces new customs checks, tariffs (in some cases), and regulatory barriers, leading to increased costs and delays.

●

GDP Growth: The UK economy has grown more slowly compared to if it had remained in the EU, with estimates suggesting a 4–5% hit to GDP over the long term.

●

Labor Market: Reduced immigration from the EU has contributed to labor short-ages in sectors like agriculture, hospitality, and healthcare.

●

Investment: Business investment has declined relative to other G7 countries, partly due to uncertainty and reduced market access.

Political Impact

UK Politics: Brexit reshaped political alignments, contributed to leadership ● changes, and intensified debates over Scottish independence and Northern Ire-land.

EU Relations: UK-EU relations have become more distant and sometimes ● strained, especially regarding Northern Ireland's trading arrangements.

Northern Ireland Protocol: A special agreement was needed to avoid a hard ● border with Ireland, causing ongoing political tensions and trade complications.

Social & Cultural Impact

Freedom of Movement: UK citizens lost the automatic right to live and work in ●

EU countries, and vice versa, affecting students, workers, and retirees.

Public Opinion: The country remains divided, with some expressing regret over ● the decision and others seeing it as a reclaiming of sovereignty.

Global Standing

Foreign Policy: The UK is seeking to redefine its global role post-Brexit through ● trade deals and the concept of "Global Britain."

Trade Deals: New agreements have been struck, but many are rollovers of EU ● deals and offer limited economic boosts compared to lost EU trade

Google Gemini (2.0 Flash)

The impact of Brexit is complex and continues to unfold, with effects felt across various sectors of the UK economy and society. Here's a summary of key impacts:

Economic Impacts:

Trade Disruptions:Increased trade barriers, including customs checks and regulatory differences, have led to higher costs and administrative burdens for UK businesses, particularly those trading with the EU. This has resulted in reduced trade volumes, especially for goods. Small and medium-sized enterprises (SMEs) have been disproportionately affected.

Labor Shortages:The end of free movement has created labor shortages in critical sectors such as agriculture, healthcare, and hospitality. This has led to increased operating costs and limited output.

Investment Uncertainty:Prolonged uncertainty surrounding Brexit has dis-couraged business investment, hindering economic growth. Concerns about regulatory divergence and trade barriers continue to impact investment decisions.

GDP and Productivity:Economists have indicated that brexit has negatively effected the UK's GDP.

Brexit has imposed lasting structural constraints on productivity.

Social Impacts:

Changes in Immigration:Brexit has significantly altered immigration pat-terns, impacting the availability of labor and the composition of the UK population.

Political and Social Divisions:Brexit has exacerbated existing political and social divisions within the UK.

Changes to rights:There have been changes to rights that citizens previously had, such as those related to travel, and working within the EU.

Key Considerations:

●

The full extent of Brexit's impact is still being assessed, and its effects are likely to continue to evolve.

●

Global events, such as the COVID-19 pandemic and the war in Ukraine, have compounded the economic challenges facing the UK, making it difficult to isolate the precise impact of Brexit.

●

The effects of Brexit are felt unevenly across different sectors and regions of the UK.

In essence, Brexit has led to increased trade frictions, labor shortages, and invest-ment uncertainty, contributing to economic challenges. It has also had social and political ramifications that continue to shape the UK.

Claude (3.7 Sonnet)

Brexit Impact Summary

Brexit has had wide-ranging effects since the UK officially left the European Union on January 31, 2020, with the transition period ending on December 31, 2020:

Economic Impact

●

Trade with the EU declined significantly, with UK exports to the EU falling ap-proximately 14% in the first year after the transition ●

New customs requirements and regulatory checks created trade friction, increas-ing costs for businesses

●

Financial services sector experienced job relocations to EU financial centers like Frankfurt and Paris

●

Supply chain disruptions affected various industries, particularly those reliant on just-in-time delivery

Political Impact

●

Increased political tensions within the UK, particularly regarding Northern Ire-land and the border issue

●

Scotland's independence movement gained momentum due to Scotland's strong pro-EU stance

●

Reduced UK influence in European and global affairs

●

Strained UK-EU diplomatic relations that continue to require ongoing negotia-tion

Immigration and Labor Market

●

End of free movement led to labor shortages in sectors heavily reliant on EU workers (agriculture, hospitality, healthcare)

●

Introduction of a points-based immigration system prioritizing skilled workers

●

Reduction in EU migration to the UK and some return migration

Regulatory Divergence

●

UK began establishing its own regulatory frameworks in areas previously gov-erned by EU legislation

●

Some businesses now face dual regulatory compliance costs

●

UK gained more freedom to set its own standards, but with trade-offs in market access

The full long-term economic and geopolitical consequences are still unfolding, with ongoing adjustments to the new relationship between the UK and EU continuing through 2025.

DeepSeek-V3

Brexit, the UK's withdrawal from the European Union (effective January 31, 2020, with a transition period until December 31, 2020), has had significant economic, political, and social impacts:

Economic Impact

1. Trade Disruptions – New customs checks and regulatory barriers slowed.

UK-EU trade, increasing costs for businesses.

2. Economic Growth Slowdown – The UK economy underperformed compared to.

other G7 nations, with reduced foreign investment.

3. Labor Shortages – Stricter immigration rules led to shortages in sectors like.

healthcare, logistics, and hospitality.

4. Financial Services Strain – Many firms relocated EU operations to maintain.

market access, weakening London's dominance.

Political Impact

1. UK-EU Tensions – Disputes over Northern Ireland’s protocol strained relations.

though the Windsor Framework (2023) eased some issues.

2. Scottish Independence Calls – Brexit reignited Scotland’s push for indepen-.

dence, as it voted strongly to remain in the EU.

3. Domestic Instability – Brexit contributed to political turmoil, including multiple.

prime ministerial resignations (May, Truss, Sunak).

Social & Regulatory Impact

1. Immigration Changes – EU freedom of movement ended, replaced by a points-.

based system favoring skilled workers.

2. Loss of EU Benefits – UK citizens lost automatic rights to work, study, and.

access healthcare in the EU.

3. Divergence in Laws – The UK began altering EU-derived regulations (e.g..

environment, labor laws), creating uncertainty.

Global Perception

●

Seen as a blow to European integration, though the EU has remained resilient. ●

The UK’s global influence has diminished, though it seeks new trade deals (e.g., CPTPP).

Long-Term Outlook

●

Brexit’s full impact will take years to assess, with potential further economic and political realignment.

●

The UK faces challenges in balancing sovereignty with economic cooperation.

Acknowledgements The research for this paper has been conducted as part of the activities of the EU-funded SOLARIS project (Grant Agreement No. 101094665). We benefited from numerous discussions within the consortium, and also from feedback received at various academic and dissemination events where we presented this work. For more information, visit the linked source first draft of the manuscript was written by Andrew McIntyre and all authors commented on previous versions of the manuscript. Bibliographical research and editing was conducted by Lucy Conover. The final version was proofread and validated by Federica Russo. All authors read and approved the final manuscript.

Funding The research leading to these results received funding from the EU Commission under Grant Agreement No. 101094665.

Declarations

Ethics Approval The research did not involve any experiment requiring ethical approval.

Consent to Participate The research did not involve any participant.

Consent to Publish All authors gave consent to publication of this manuscript.

Competing interests The authors have no competing interests to declare that are relevant to the content of this article

Author Contributions The main thesis was conceptualised by Andrew McIntyre and Federica Russo. The