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

2.1 Life + artefact: how is symbiosis possible for machines?

At the heart of considering Symbiotic Artificial Intelligence (SAI) from a foundational perspective is the interplay between organic life and manufactured artefacts, traditionally viewed as diametric opposites. The essence of a living organism differs markedly from that of an artificial construct, yet in the context of SAI, these distinct entities are envisioned to coalesce. The historical backdrop for differentiating natural beings from artificial ones dates back to Aristotle’s Physics and extends through Darwin’s delineation between organic and inorganic realms [11], suggesting an initial clear division between two distinct realms of existence that, within the SAI acronym, are brought into conjunction.

According to this traditional paradigm of the life sciences that goes from Aristotle up to modern biology, an artefact or a machine (an entity that is a product of techne, according to Aristotle’s own definition) is not something alive. What does it mean to be alive for Aristotle? In De anima Aristotle says that to live means to have the capacity to nourish, grow and decay [1, p. 22]. These functions are all united under the sign of movement, specifically spontaneous movement. For Aristotle and then throughout the long course of Western biology, what distinguishes a living thing from an artefact is the ability to move spontaneously. The machine is not endowed with life, it is not ‘animated’ (ensouled) because it does not possess an autonomous principle of movement. We move the machine, whereas a living thing, an insect, or an animal, such as plants themselves (which have a nonlocal motion but a motion of growth), move autonomously. We do not drive them.

The modern era, especially with the rise of the mechanistic model, will reject such a view of life by coming to conceive even living beings as machines whose movement is always the result of a complex chain of external impacts (think of Descartes, Hobbes, and Spinoza). But a revival of Aristotelian conceptuality will occur in the nineteenth century, with Naturphilosophie and the emergence of an ‘organicistic’ paradigm, which gave birth to modern biology.

This conception, whereby a living being is one that possesses an autonomous principle of movement, a self-finality (entelechy), has been somewhat challenged by the emergence of robotics and computer science in the twentieth century. These sciences study the ways in which it is possible to build machines that can attain a certain level of autonomy. Obviously, it is not the autonomy that pertains to living beings in general and human beings less than ever. However, all artificial intelligence systems, especially Large Language Models (LLMs), have shown self-development capacity. Robotics and computer science constitute an exception to the old distinction between physis and techne, between life and artefact. Robots and instances of artificial intelligence are a special kind of machine: they confront us with ‘mechanical’ processes that nevertheless exhibit characteristics that somehow have to do with life and, thus, potentially also with symbiosis. Maturana and Varela proposed in the twentieth century the famous distinction between autopoietic and allopoietic machines: the former machines would be the living organisms, that is, machines whose process results in the preservation, reproduction, and development of the machines themselves, while allopoietic machines are machines that are operated and have something other than the machine itself because of their process [43].

One might ask: What distinction would remain if an artefact (a device or technology devoid of organic components) shared all characteristics typical to organisms composed of organic material? Would they both be deemed living entities? [44, p. 55; p. 56]. Is it possible for an artefact to transform into an autopoietic machine? This avenue has been explored since the mid-20th century (for instance, by Von Neumann with his self-replicating machines) and later during the 1980 s in the field of artificial life (A-life). A-Life operates on the functionalist premise that life can also be digitally synthesised, meaning the composition of the substrate, whether atomic or digital (bits), is irrelevant. What matters is that this substrate demonstrates specific interrelations and characteristics, such as self-preservation, self-reproduction, and autonomous movement for its own sake, among others. Numerous examples exist (Conway’s Game of Life, Polyworld, RepRap, Slugbot, etc.). These advancements appear to blur the traditional line between life and artefact, a matter of great significance for the question of founding SAI. In this respect, SAI seems to be an AI that potentially engages in a ‘life-to-life’ relationship. But accordingly, are robots ‘living’ entities? Are artificial intelligence such as ChatGPT entities that, although not “material,” exhibit some of the characteristics of life?

Addressing robots, Lévy [38] suggests that, according to some broadly accepted biological benchmarks (such as those proposed by Koshland [34]), it is not entirely far-fetched to classify robots and, by extension, AI programs as life forms. De Collibus [13] expresses a contrary view regarding Large Language Models (LLMs). Several attributes typically associated with living beings are also identifiable in ChatGPT, like the ability to self-evolve and learn, a trait shared with humans, dogs, or fish. ChatGPT can capitalise on accumulated data like these organisms, enhancing its complexity and devising increasingly effective solutions over successive problem-solving attempts. However, these technologies lack a crucial aspect of life as defined by Spinoza’s concept of conatus and Schopenhauer’s Wille zum Leben, the instinct for self-preservation. None of these technologies operate driven by a desire to continue existing. Perhaps the self-purposefulness Aristotle initially identified as a primary feature of life is not fully present in AI: it is indeed capable of self-expansion or self-improvement, but to what end? It does not appear to do so in pursuit of its assertion or its own ‘good’. This could represent an insurmountable gap between living beings and machines, posing a significant challenge to applying biological categories such as symbiosis to AI technologies.

2.2 Intelligence and symbiosis: three patterns based on (dis)continuity

How do things stand with intelligent life? This question is essential for the foundation of an SAI since it is assumed that the two lives that should enter symbiosis are both intelligent lives. Indeed, in the case of the artificial symbiont, its life might coincide completely with its intelligence. SAI thus allows us to ask important questions about the relationship between life and intelligence.

The problem is tough since it depends on what kind of definition of intelligence we start with.

A spectrum of perspectives on defining intelligence, ranging from inclusive to exclusive, underpins discussions on the nature and potential of SAI. Starting with Schelling and Darwin, a minimal view emerges that intelligence essentially involves the capacity for organisation. This perspective posits that even simple biological entities, such as mushrooms or worms, demonstrate intelligence through their organised efforts to solve problems and adapt to their environment. This organisation principle extends to artificial intelligence, where software and programs exhibit a basic structure enabling them to reconfigure based on contextual inputs to meet specific goals [41, p. 21]. From this perspective, which could be called ‘homogeneous continuism’, an SAI is already at work. It has been for a long time [39]. Our devices, through which we interact with generative AI systems, are already in some sense forms of intelligence with which we live in symbiosis. Here we are dealing with a definition of intelligence that flattens discontinuities and reduces them to a mere difference in degree: it is enough for an entity to exhibit a specific computational strategy to be considered intelligent (from viruses to humans and AI systems). At this level, an SAI is conceivable as a form of integration between human beings and digital devices [35, 62], in a relationship that provides for their increasing autonomy, seamlessness and self-development. A perspective that has begun to be investigated in the White Papers on Symbiotic Autonomous Systems (SAS) [3, 49].

However, there are less inclusive definitions, such as those that restrict the field of intelligence to vertebrates with a minimum of brain activity [40]. In this case, calling that AI an ‘intelligent’ life would already become less noticeable. Floridi [26], for example, contends that the intelligence of current generative AI systems falls short of even that of a sheepdog, likening their cognitive capabilities to those of a dishwasher (see also [14], in which Dennett argues that AI is the result of “a slow, mindless process”). Accepting this premise implies that SAI cannot be based on our current interactions with technologies like smartphones or ChatGPT. Instead, envisioning SAI requires imagining a relationship between humans and a robot whose artificial intelligence parallels at least that of an animal. Under this scenario, true SAI would manifest primarily in an exosymbiotic, rather than endosymbiotic, framework, potentially side-lining wearable or prosthetic AI technologies from the SAI paradigm. However, the problem is that the realisation of ‘animal’ level AI remains highly problematic.

In opposition to continuism, there are various versions of intense or absolute discontinuism, whereby intelligence is a uniquely human capability. To define intelligence as any ability to solve problems or to calculate even the way animals do would be an equivocal way of talking about intelligence because proper intelligence is only our own. Indicative of ‘proper’ intelligence would only be capabilities such as intentionality, universalisation, creativity, spontaneity, self-consciousness, emotionality, etc.: all characteristics that are the exclusive preserve of human beings. In this case, to talk about SAI, we would have to approach science fiction or post-humanism or very distant in time perspectives that contemplate the possibility of achieving an Artificial General Intelligence (AGI) and thus some “Singularity” [7].

3.1 SAI foundation vs. human-centred ethics

As previously discussed, the intricate nature of the ‘symbiosis’ concept encompasses various definitions and foundational viewpoints regarding the complex relationship between beings in general, with particular emphasis on humans and artificial agents. This relationship poses a significant challenge due to its dual investigative nature, where symbiosis is interpreted as a connection between life forms at one level and as a relationship between forms of intelligence at another [28].

However, what implications arise from such foundational variety when transitioning from speculation to ethical evaluation and normative orientation? Do we genuinely necessitate an ethics of SAIs? And if so, given the absence of a unified foundation, what type of normative orientation should we adopt?

In ethics, specifically, the normative ethics associated with AI [26], obligations are not necessarily imposed, but behavioural normative orientations are expressed. A normative orientation indicates a ‘should be’, reflecting a structured yet not always systematic framework containing justifiable and coherent moral normative statements [12]. However, it’s essential not to misconstrue the ‘should be’ as a forced adherence to a value or ideal; instead, adherence may be motivated by various factors, such as a consideration of consequences (consequentialist ethics), one’s character or behavioural disposition (virtue ethics), or a principle transcending consequential calculations (deontological ethics) [57]. Moreover, as observed in eudaimonic ethics, individuals may follow a rule to align with a personal or collective ‘should be’, even for the sake of happiness [12].

Therefore, the challenge lies in establishing a foundation for being and determining the normative orientation to govern the diverse aspects of moral ‘should-be’. Consequently, the question arises: to what normative orientation can SAI, already complicated by excessive variability at its foundation, be traced back?

The human-centric approach is the indispensable cornerstone of normative guidance within the FAIR project framework. It forms the bedrock of the European Commission’s ethos on AI ethics [21, 30] and resonates throughout various international normative doctrines [48, 63]. Yet, harmonising the intricate foundation of SAI with the imperative of AI human-centeredness presents formidable hurdles. Our inquiry reveals this reconciliation to be somewhat problematic. Upon scrutiny, a growing dichotomy emerges between these two paradigms, primarily due to irreconcilable disparities.

Beginning with SAI, the quandary lies in its excessive foundational variability, which defies attempts to embed it within the realm of human understanding. Consequently, its origins need a distinctly human-centric focus.

A second challenge arises concerning the ethical dimension of AI human-centeredness, revealing inherent vulnerabilities. Even before delving into specific technology interactions, criticism may be levelled at the human-centric approach for its latent speciesism [60]. Critics contend that human superiority is a flawed concept, advocating instead for acknowledging intrinsic value in non-human entities. The application of anthropocentrism to AI ethics thus risks being perceived as a biased global guiding principle, prompting scrutiny over its relevance in matters of international and environmental justice [8–10].

These challenges underscore the apparent discordance between an SAI needing a more explicit foundation and a notion of human-centeredness that may verge on abstraction or even AI colonialism [10]. If left unresolved, this dichotomy is poised to escalate. On the one hand, “techno symbiosis” is on the rise, albeit not exclusively in a biological context but also in a metaphorical and meta-semantic sense [29]. Conversely, substantive proposals to transcend mere principle-based definitions and actualise normative ideals into tangible, democratic social and political frameworks – such as navigating human existence in the algorithmic age – remain conspicuously absent [27].

3.2 Constructivism as an onto-epistemological framework of SAI

Could we transcend this apparent tension of irreconcilability? Such an avenue seems feasible only through an approach rooted in constructivism, emerging as the foremost plausible option for reinstating a human-centred focus in the governance of SAI.

Symbiosis, in essence, does not exist. It is not merely constructed within interactions between humans and machines – whether these interactions occur at a biological level (prosthetics) or within the realms of imagination (the metaverse and digital twins), creativity (ChatGPT, DALL-E), or as support for determinations, judgments, and courses of action in an organisation or business (DSS). Instead, it is constructed through interactions between humans and machines. In other words, symbiosis lies not solely in the algorithm and its learning model (supervised, unsupervised, reinforcement learning) but in the chain of choices, moral reasoning, and responsibility along the sociotechnical axes that comprise an SAI system. The sociotechnical nature of symbiosis [3, 45] primarily imparts the symbolic form to this concept [29], resembling what the German sociologist and philosopher Max Weber referred to as an “ideal type” [21, 33]. An ideal type functions as an abstract and hypothetical construct, not detached from reality but used in social sciences to dissect and comprehend complex social phenomena. It captures essential characteristics while transcending strict realism, as a conceptual scaffold for understanding the intricate social networking of technosystems [21, 33].

An ideal type’s comprehensive and explanatory power lies in the richness of the constructivist approach to the SAI foundation. It is not a proper foundation because, as we have said, symbiosis per se does not exist. Instead, symbiosis manifests as a set of processes rather than a brute fact in the interaction between humans and machines. Therefore, theoretical attention should shift toward understanding the dynamism of the processes through which such condensations exhibit some degree of symbiosis[8, 9].

New approaches are emerging around similar needs, challenging the link between machines and ethics understood solely in normative terms (compliance with laws and ethical principles) or engineering terms (machine ethics). The problem does not lie in the ontological definition of the type of symbiosis with machines but in extracting rational and social value from these definitions, interrogating them both as technological devices and as metaphors of “socio-material practices” [49] or “techno-scientific practices” [53].

Finally, such a repositioning of symbiosis in the domain of possibility and dynamics would render dialectical the old opposition between conceptual foundation and ethical evaluation of human-centeredness. In doing so, the foundational excess of SAI would not be a conceptual problem but an epistemic opportunity, a potential strength that constructivism would help to understand by extracting value and insights from how it can be defined. In this regard, it is precisely, onto-epistemic to use the terminology used by physicist-philosopher Karen Barad to point at the inseparability of ethics, ontology and epistemology when engaging in scientific knowledge production, with social and technological practices and with the world itself and its inhabitants – human and non-human beings that intra-actively co-constitute the world [2]. Our constructivist approach reshapes this agential realist perspective that underlines the intra-activity between the world and its forms of symbiosis because we foresee here different methodological advantaging premises:

Flexibility in definition. Artificial symbiosis evolves alongside technological advancements and societal shifts. A socio-technical constructivist approach recognises that the definition of artificial symbiosis can be problematic and should not be fixed in stone. It acknowledges that our understanding of this symbiosis depends on current knowledge, technology, and society. This approach allows for ongoing adaptation and refinement of the definition based on empirical testing and real-world implementation scenarios, ensuring it remains relevant and adaptable to changing circumstances.

Anthropocentric value review. The socio-technical constructivist approach acknowledges that the anthropocentric value of artificial intelligence is not absolute and specist but rather an ideal-typing perspective rooted in the social and historical dimensions of any realism, including values. This approach facilitates a more inclusive and open discussion about the roles and regulations of human-machine symbiosis. It allows us to incorporate diverse, intersectional perspectives within a framework that does not privilege any single viewpoint but recognises the value of coexistence and coevolution between humans and machines [50].

Adaptability to emerging AI-based technologies. As AI evolves rapidly, a socio-technical constructivist approach can accommodate the emergence of new technologies and their impact on the symbiotic relationship between humans and machines. It enables us to consider and adapt to unforeseen developments, ensuring our understanding and regulations remain relevant and practical. This flexibility is essential in an era where technology is advancing at an unprecedented pace.

A cautious approach to unbridled techno-ideology. Contemporary understanding underscores that technological potency is anything but neutral – it assumes a role in shaping and co-determining societal structures [18, 65]. Thus, embarking on a constructive exploration of the conditions underpinning human-machine symbiosis facilitates a precautionary and context-aware [56] stance against unchecked techno-ideological fervour, preventing succumbing to facile exuberance.

Evaluating a socio-technical system is notably more challenging than assessing an individual technology due to numerous variables, contextual variations, scenarios, and changing user dynamics. In our ongoing research as part of the FAIR project, we are developing a methodology for the human-centred impact assessment of SAI, which comprises several steps, each executable through diverse techniques and approaches.

The initial step of this methodology entails electing socio-technical constructivism, which we have delineated as the onto-epistemological foundation [53] upon which to ground our method. Naturally, this foundation is ‘non-foundational’, as it is determined by the levels of abstraction and participation implicated in techno-scientific practices.

3.3 Sociotechnical features for a preliminary ethical evaluation of human-centred SAI

How can we effectively evaluate the human-centredness of an SAI system? What criteria should we consider when identifying the level of symbiosis within the system?

Indeed, establishing an evaluation framework is the second step of the methodology we are working on. This framework scrutinises and weighs the ethical robustness concerning the design and implementation of SAI systems in various contexts. Its primary purpose is to describe the application of an SAI system deployed in a specific project and sociotechnical context. This step allows the method to initialise and serve: (a) to map out the general socio-technical features of the SAI system under analysis; (b) to provide a set of dimensions to contextualise the application of the SAI system under examination; (c) (For each dimension) to offer a series of questions, the answers to which facilitate the screening of the respective SAI system.

To be succinct, this article will not delve into the questions but rather aims to account for the five dimensions along which screening occurs in our method.

AI model description. First and foremost, it’s essential to define the main features that identify the AI system being designed and implemented or that we want to consider in our analysis. In this initial impact assessment stage, developers are asked to describe the AI system under evaluation and specify the symbiosis’s operating context. The entire developers’ team should collaborate on completing this passage, explaining (a) the scope and users of the AI system, (b) the dependency on other AI systems, and (c) the algorithm model description [48, 63].

Symbiosis description. The elucidation of symbiotic levels holds pivotal significance in constructing the framework for the analysed SAI system. We have categorised the description into criteria and sub-criteria to address a diverse spectrum of symbiotic levels and dimensions comprehensively: (a) biological and corporeal symbiosis; (b) symbiosis on task(s) to perform; (c) symbiosis on combining tasks and actions into multi-task, composite systems; symbiosis as autonomy level.

Data description. Data encompasses the information and expert insights an AI model utilises to construct a representation of the context or environment. Expert input typically involves codifying human knowledge into rules. Key characteristics within this domain encompass the origin of data and information, the method of collection by either machine or human, data structure and format, and inherent data properties. These characteristics are pertinent to the data employed in training an AI system (‘in the lab’) and the data utilised in production (‘in the field’). The following table provides a summary of these distinct criteria and sub-criteria: (a) Data collection; (b) rights and identifiability; (c) structure and format; (d) quality and appropriateness [48].

Proportionality description. In this framework descriptive phase, it is also crucial to comprehend the degree of proportionality between the AI system’s promoted (or promised) services and utilities and whether the intended purpose justifies its utilisation, considering the technology’s risks, uncertainties, and drawbacks. This examination allows practitioners and experts to balance the means and the intended goal. It is a process to justify the necessity of utilising a specific method or AI system and to demonstrate its suitability. This ensures that procedures related to, or a part of, the AI system do not exceed what is necessary to achieve legitimate objectives [63].

Governance description. Conducting this screening is pivotal as it allows for clearly identifying critical roles and responsibilities, ensuring transparency and accountability within development teams. Indeed, in our method, framing the governance means paying attention to two intertwined aspects. On the one hand, the identification of the responsibilities (top-down). Given the inherent risks associated with AI, it becomes imperative to delineate clear lines of responsibility for each facet of the AI system. Establishing this governance framework not only safeguards against ambiguity in accountability but also enhances trust, ethical adherence, and effective decision-making within the development and deployment processes of SAI.

On the other hand, it is identifying stakeholders’ value chain (bottom-up). Engaging diverse stakeholders is imperative for comprehensively evaluating the impacts associated with this SAI system. Consequently, the developers’ team should devise a detailed stakeholder engagement strategy in the initial phases of system design. This strategy will enable the team to establish their engagement goals, which should be periodically reviewed. This ensures that stakeholder involvement is not merely a procedural task but becomes an integral and transformative element of the decision-making process.