A human-centred approach to symbiotic AI: Questioning the ethical and conceptual foundation

The European Commission (EC) has long advocated the certainty of a human-centric approach to Artificial Intelligence (AI), emphasising communication between societal stakeholders and technology developers to establish a tangible foundation of trust in AI. This approach found notable expression in the “Ethics Guidelines for Trustworthy AI” [30] promoted by the EC High-level Expert Group on Artificial Intelligence (hereafter referred to as HLEGAI). Explicit references in this regard are also present in other European documents. Notably, as early as April 2018, the Commission had outlined the cornerstones of its strategic direction by emphasising the centrality of humans in AI development. The Commission’s communication to the European Parliament, the Council, the European Economic and Social Committee, and the Committee of the Regions, titled “Building Trust in Human-Centric Artificial Intelligence” [20], deserves mention. These communications have been incorporated not only in the guidelines but also in the “White Paper on AI” [21] and the proposed “Artificial Intelligence Act” [22].

Collectively, these documents underscore the European Union’s (EU) political agenda not only to coordinate an integrated approach for maximising AI opportunities and addressing associated challenges but, more importantly, to position itself as a global leader in the development and deployment of AI that normatively aligns with principles of sustainable technological innovation and respect for fundamental rights (already started with the GDPR [19], the so-called “Brussels effect”, see: [59]).

This approach underscores several pivotal strengths, notably encapsulated in the HLEGAI’s endeavour to leverage a dual normative function of fundamental rights. Indeed, these rights encompass both the legal protections enshrined in the constitutional frameworks of nations (actual legal rights) and the inherent rights of individuals grounded in the intrinsic moral status of human beings (ethical values that may not be legally binding but are pivotal for ensuring the trustworthiness of AI on a global scale).

Another salient strength is recognising that every normative orientation constitutes a structured, albeit not always systematic, array of moral assertions interlinked in diverse manners. In essence, all these assertions can be traced back to a primary moral principle (monistic ethics) or can stem from multiple principles (pluralistic ethics) [57]. The EU’s ethical framework for a “good AI society” [25], particularly as formulated within the HLEGAI, adopts a pluralistic ethical standpoint contingent upon interrelated principles for the betterment of society.

In this paper, we discuss the challenges posed by applying a human-centric normative orientation to Symbiotic AI (SAI) or human-AI symbiosis [28, 51]. Human-computer symbiosis is a relatively long-time metaphor [29], initially proposed by J.C.R. Licklider in his historical 1960 article [39]. However, his vision describes a symbiotic relationship between humans and computers, where humans initiate goals and hypotheses, and computers process data to test these hypotheses, with the anticipation of a networked system for communication. Despite the initial vision from sixty years ago, contemporary AI systems now surpass this model by discerning patterns in large datasets, making decisions, and even creating the data themselves, suggesting a transition towards “technosymbiosis” [29] or what we call symbiosis as collaborative teaming framework. SAI today aims to enhance human-machine collaboration in various domains, such as healthcare and manufacturing, where humans and AI systems collaborate for improved outcomes [32]. Machines are becoming agents of technology’s “in-betweenness” [24], and they can engineeringly (and not cognitively) generate informational meaning. Something dissimilar to human-like understanding and inferring human-like responsibility [26], but precisely because of this, it represents a new ethical challenge that highlights the sociotechnical interdependence between humans (and nonhumans) and computing devices, noting both benefits and risks. This is why fostering a human-AI symbiosis requires today’s design of AI systems according to a human-centred approach.

Against this backdrop, research at the University of Bari, conducted within the NRPP-funded Future AI Research (FAIR) project, focuses on designing new interaction paradigms to amplify human performance while ensuring system reliability, safety, and trustworthiness. Acceptability of SAI systems, particularly concerning value alignment between AI and humans, is a crucial focus of research within FAIR. This involves interdisciplinary studies to address issues such as privacy policies, security, and fundamental freedoms. Philosophical perspectives are also explored to understand the epistemological and ethical aspects of SAI.

In this paper— continuing the work presented at the workshop BEWARE@AIxIA 2023 [6]— we contribute to the understanding of human-centred symbiosis with AI by introducing a methodology for assessing the impact of SAI systems using a human-centred approach.

In Chapter 2, we address some foundational and philosophical challenges of SAI by analysing first the relationship between life and artefact (Section 2.1) and then the one between intelligence and symbiosis (Section 2.2).

Chapter 3 will explicitly address the potential of a constructivist approach in redefining the foundation of SAI and its alignment with human-centred ethics (Section 3.1), advocating for flexibility and context awareness. We will propose a constructivist framework for understanding (Section 3.2) and evaluating (Section 3.3) SAI systems, delineating the outline of an evolving methodology on which we are working in the project FAIR – constructivism is a social sciences theoretical perspective asserting that knowledge and meaning are constructed through social interactions and experiences, emphasising the role of culture and context in shaping human understanding [2, 53].

In Chapter 4, we conclude the paper with final remarks and outline directions for future research endeavours.

A last clarification before proceeding: We use the terms ‘human-centric’ and ‘human-centred’ synonymously, referring both to design and research approaches that prioritise human needs, behaviours, and experiences as the central focus, ensuring that the resulting systems or solutions are optimally aligned with user requirements and usability principles.

The concept of ‘Symbiotic Artificial Intelligence’ introduces an additional layer of complexity to the already problematic notion of ‘Artificial Intelligence.’ Incorporating the adjective ‘symbiotic’ further complicates the scenario yet represents one of the newest boundaries in AI research. The term ‘symbiosis’ inherently suggests the coexistence of two or more entities that collaborate or, at the very least, derive mutual benefit from their association. Thus, a genuinely Symbiotic AI would foster an interaction between humans and machines that transcends the traditional dynamic of controller and control, evolving into a reciprocal relationship between two agents potentially equal in decision-making, capable of mutually influencing one another. These agents, though not identical, are both intelligent, with the former ‘controller’ – the human – entering a partnership where the machine may assume control by adjusting, amending, or directly intervening in decisions in real-time towards a shared objective, leveraging a vast and precise understanding of data and procedures necessary for the task. It’s important to acknowledge that despite the potential asymmetry in this interaction between intelligent entities, a degree of control always remains, exercised mutually, ensuring the symbiotic relationship’s continuity.

A visionary example could be Elon Musk’s Neuralink. Let’s imagine that one day, some of us, for both therapeutic and creative purposes, could equip ourselves with brain implants designed to allow us to use our neural signals to control external technologies. Naturally, the AI guiding such implants would be fitted with an ethics by-design that – at the very least – requires it to adhere to Asimov’s laws. However, it is also true that, despite the significant imbalance in the symbiotic relationship between humans and the implant, the AI would also have its capacity for increasing self-learning. Paradoxically, this tendency to autonomy would grow by feeding on our neural processes, thereby evidently limiting our absolute control over the machine and, consequently, over ourselves. This model of human-machine interaction represents an extreme and radical variant of what Floridi defines as a “third-order technology” [24]. In the Italian philosopher’s idea, a third-order technology removes or replaces humanity from the loop, eliminating the H2M interface in favour of M2M protocols. Delving into nanotechnologies and biotechnologies, AI is not simply reengineering the world with its protocols, but re-ontologizing it. Through the re-ontologising of our society, the digital realm is also redefining, from an epistemological standpoint, the rationality underlying our society, namely many of our established conceptions and ideas [26].

Against this backdrop, envisioning SAI raises questions about the nature of the technology required and whether current forms of human-machine interaction hint at a future symbiotic relationship. Kai-Fu Lee and Chen Qiufan [37] speculate about a near future where smartphone applications, through real-time data on our actions and behaviours, could suggest changes to our decisions or behaviours based on an intricate network of app interactions. Is this merely a matter of data availability and technological capability, or is something more profound at play in achieving a symbiosis between natural and artificial intelligence?

To consider AI in this context is to imagine machines or software capable of not just ‘live’ learning and response, as seen with the IoT, but also of discerning behavioural regularities, intervening in real-time with a specific ‘authority’ to alter human conducts that would otherwise occur differently. Symbiosis, in this sense, implies a learning-driven agents’ relationship where both entities coexist and learn from one another: the machine from observing human actions and the human from insights provided by the machine based on its observations. In a recent project attempting to use symbiosis between machine and human beings to build AI systems capable of understanding the nuances of the real world, precisely this point is emphasised: “While the computer’s role has shifted from being a passive ‘executor’ to an active learner, the role of the human is still that of an actively involved teacher, because it is the humans that create clean, consumable data for the computer to analyse and hence train itself. However, what if there was a way in which the computer’s role remains that of an active learner, but we humans can passively sit back and go on with our lives while the computer learns from our actions?” [52, p. 4].

The shift here is from instructing machines to learning from them, not just acquiring new information but receiving guidance on enhancing our actions to achieve desired outcomes. This form of HCI is widely discussed in scientific literature, but the accuracy and implications of applying ‘symbiosis’ to AI warrant further examination and pose foundational philosophical questions. This section addresses these questions, exploring the definition and feasibility of SAI and its compatibility with a human-centric approach.

Suppose we delve into the historical divide between organic life and artificial constructs, questioning whether machines can exhibit characteristics traditionally associated with living organisms. In that case, speaking and envisaging a foundational hypothesis of SAI takes a lot of work. While some argue for a continuum of intelligence across biological and artificial entities, others contend that true symbiosis with machines necessitates AI with capabilities rivalling at least those of animals. Accordingly, from the kinds of human-machine interaction known today, it is pretty impossible to find ‘symbiosis’ by reflecting on the nature of intelligence, the potential for symbiotic relationships between humans and machines, and the boundaries between living and non-living entities.

Based on these premises, in this part of the paper, we will argue in favour of a possible alternative pathway to the foundation of SAI. We will pursue a constructivist approach, a theoretical perspective asserting that knowledge and meaning are constructed through social interactions and experiences, emphasising the role of culture and context in shaping human understanding.

From a deterministic and theoretically firm perspective, we know that a constructivist approach may not appear as the most rationally justified path to speak of ‘foundation’, as nothing is constructed without first founding something. However, given the highly sociotechnical nature of SAI, there is no possibility of a foundation other than the ‘weak’ foundation of a constructivism that maintains a flexible, adaptable, cautious, and context-aware direction of thought. This position is supported not only by the foundational dilemmas and doubts we have shown in Chapter 2 but also by the argument that a constructivist approach is the only possible one if we want to reconcile an explainable justification of SAIs with ethics of AI based on the human-centred and trustworthy design of machines, as demanded by European institutions [21, 30] and significant international agencies [48, 63]. To demonstrate this, in Section 3.1, we will highlight the critical points of this discrepant opposition between the foundation of SAIs and the ethics of AI, such as human-centeredness.

Having highlighted this discrepancy in Section 3.2, we will argue in favour of a constructivist approach as the only possible way to reconsider the relationship between SAI and human-centeredness in dialectical terms rather than opposition. Of this approach – which we are trying to refine in the FAIR project – in this paper, we present its methodological outline consisting of a theoretical onto-epistemological framework (3.2) and a preliminary evaluation framework (3.3) to identify the main sociotechnical characteristics of an SAI system and, from there, proceed to the subsequent steps – which we are still refining in our FAIR team – for the ethical assessment of the human-centeredness of SAI.

A constructivist approach to symbiosis presents itself as the most viable alternative for reinstating human-centredness in the governance of SAI. Should we persist in adhering to rigid, deterministic foundational methods, we risk gradually divorcing ethics from being a mere adherence to moral principles (such as autonomy, beneficence, non-maleficence, justice, etc. [21, 30]) to being a compelling aspect of machine design, and more broadly, in shaping the digital society we aspire to inhabit. Indeed, the notion of sustainable, fair, and trustworthy algorithms – AI ethics – cannot be solely relegated to compliance with ethical principles. Ethics, in this vein, remains overly abstract and unnecessary [9], fails to serve stakeholders by providing clarity or guidance in development and commercialisation [46], and is at risk of being exploited maliciously [26] or even weaponised as a tool of biopower and colonisation [10].

We must relinquish the grip of determinism – the persistent desire to control centrally – and instead embrace a distributed, contextual, relational, and dialectical structure to embody the ecosystemic essence of human-centeredness truly. While a constructivist approach may initially appear to shift responsibility from humans to machines, it, in fact, addresses the issue of AI responsibility not by fixating on who holds control (humans or machines) but by methodically analysing and assessing the chain of responsibility. In essence, transitioning from the question of “what” to an examination of the technology in action [49] entails grappling with the question of “how” [46]. Given its sociotechnical nature, symbiosis with machines prompts a re-evaluation of responsibility, focusing less on its ontological core – who bears this faculty? – and more on its inherent interplay between the social and the technological within everyday organisational life – what has also been called “socio-material practices” [49] or “techno-scientific practices” [53].

In Chapter 2, we highlighted the main issues concerning the possible foundation of SAI. Crucial philosophical questions were raised about the possibility of thinking about a form of AI capable of entering symbiosis with humans, starting from the specificity that such a technology should exhibit compared to more traditional AI. In Section 2.1, we questioned the application of biological categories – such as symbiosis – to artefacts. While robotics and AI challenge the conventional distinction between living and non-living based on the capacity of organisms to move autonomously, self-develop and reproduce, some differences remain that currently appear insurmountable. Also, in Section 2.2, we offered a scan into three patterns of SAI based on as many definitions of intelligent life: the kind of SAI we can conceive of strictly depends on how we think about the continuity between organisation and intelligence. The less continuist our position, the less chance we have of putting down a foundation of SAI in a strong sense, namely from a biological or ontological point of view.

Chapter 3 explicitly addresses the potential of a constructivist approach in redefining the foundation of SAI and its alignment with human-centred ethics. We advocate for adopting a constructivist perspective to navigate the intricate landscape of SAI, placing particular emphasis on attributes such as flexibility, adaptability, and context awareness. The discourse extensively explores the multifaceted challenges arising from the diverse foundational perspectives of SAI, proposing a constructivist framework aimed at comprehensively understanding and evaluating SAI systems. This section delineates the fundamental aspects of a theoretical (onto-epistemological) framework and an evolving evaluative methodology designed to assess the sociotechnical characteristics and human-centred attributes of SAI systems.

This method that we are working on will require further refinements as the FAIR project progresses. For this reason, in the future, the research will involve jurists to address the theme of acceptability from the ethical and legal viewpoints. Then, the legal and ethical acceptability of SAI will need to go through an operationalisation process to be of practical relevance for the design and implementation of SAI applications. High-level principles will be, therefore, turned into operational definitions that pave the way to technical solutions, e.g., for (partially) automated compliance testing. Operationalisation will be accompanied by appropriate modelling of the SAI application in hand. Notably, socio-technical systems (STS) are widely recognised as a valuable approach to complex organisational work design that stresses the interaction between people and technology in workplaces [3]. Also, multi-agent systems (MAS) [15] are promising from the practical viewpoint since they enable the simulation of possible scenarios and the experimentation of different operational definitions of the legal and ethical acceptability of SAI in a controlled environment. A starting point might be the MAS prototype presented in [16] for the moral evaluation and monitoring of dialogue systems. Finally, compliance tests might be reformulated as problems that can be addressed with automated reasoning techniques and formal methods. Overall, along the direction already explored, e.g., [17], logic will play a prominent role in implementing the computational solutions for many of the problems in our research on SAI within FAIR.

This work was partially supported by the project FAIR— Future AI Research (PE00000013), which is part of the NRRP MUR program funded by NextGenerationEU.