Introduction

Once a technology is available and fully socially integrated, it becomes difficult to find strategies to modify its range of possible uses1. Therefore, the emergence of a potentially impactful technological development calls for reflections on its possible ethical and social implications. For example, the digital revolution has had an unforeseen, yet extremely major, influence on human subjectivation2,3, the continuous and adaptive process by which each human individual becomes and remains the subject of their life, and thus one of the key elements that constitutes human nature. The current advent of advanced neural interfaces incorporating biomimetic artificial neural networks makes it possible to envision, in the near future, devices that allow a bidirectional connection at the level of individual neurons between the central nervous system (CNS) and artificial spiking neural networks (SNNs) that can learn and produce activity autonomously. The CNS is in essence the physical substrate of consciousness and of the faculty of being subject, therefore, here we would like to question the concept of subjectivation in light of the advent of bidirectional neuromorphic brain interfaces.

Emerging neuromorphic brain interfaces

Brain interfaces have seen an extraordinary development over the past decades. The most efficient interfaces make use of microelectrode arrays to record neural activity at the level of multiple single neurons. These microsystems are the cornerstone of intracortical brain–computer interfaces (BCIs) for the rehabilitation or compensation of lost functions in people with severe motor or speech disabilities. Intracortical BCIs are based on decoding intended actions from brain activity and, consistently across studies, their performance improves when the number of recorded cells increases. Therefore, efforts are ongoing to build probes with thousands of microcontacts to record from thousands of individual neurons simultaneously4.

Beyond these technological improvements of neural interfacing microsystems, the methods used to decode brain activity also undergo important changes. In particular, two lines of improvement are currently emerging: adaptive decoders that automatically self-optimize the processing of neural activity, and deep neural networks that provide more efficient predictions of intended behaviour than original linear decoders. Thus, brain implants are on the way to incorporate some form of intelligence to allow better automatic online processing and decoding of brain activity and better rehabilitation solutions for patients to regain autonomy. Yet, dense neural probes suffer from a major drawback with the very large amount of data that they produce, which needs to be processed in real-time. In particular, it is not possible to envision the embedding of efficient algorithms based on von Neumann architectures inside brain implants to automatically handle hundreds of thousands of neural data signals without power dissipation beyond biocompatible limits preserving the surrounding tissue.

To mitigate this limitation, new generation algorithms are being developed based on SNNs. SNNs are event-based artificial neural networks with local plasticity rules allowing autonomous learning distributed throughout the network without transfer in memory of large amounts of data. This makes SNNs conceptually highly energy efficient and embeddable in very-low-power neuromorphic analogue circuits.

In the field of neuroscience, SNNs embedded in very-low-power neuromorphic hardware have already been used to detect pathological intracranial neural activity events at a regional level5. SNNs have also been used in a fully unsupervised way to extract the individual activity of multiple neurons from neural recordings6. In this latter design, biological spiking activity is automatically converted into artificial spiking activity in a one-to-one mapping from biological to artificial neurons: one artificial neuron reproduces the activity of one of the neurons recorded by the microelectrode array. This biological-to-artificial conversion allows for biological spiking activity to be further integrated into artificial SNNs for downstream processing. Conversely, artificial SNNs can also be used to control extracellular electrical neural stimulations. For example, at a regional level, intraspinal microstimulations could elicit locomotor-like activity in live spinal cords under the control of an artificial SNN mimicking a central pattern generator7. Achieving an artificial-to-biological conversion of spiking activity in a one-to-one neuron mapping could theoretically be possible with arrays of intracellular electrodes. Such devices have been demonstrated in vitro8. Once available for in vivo interfacing, they would enable individual artificial spiking neurons to drive the activity of individual biological neurons in the brain. We are thus at the dawn of bidirectional neuromorphic brain interfacing in which biological and artificial neural networks could cooperate to enable a function that each network would be unable to achieve alone.

Implications on human subjectivation

The intrinsic functioning features of bidirectional neuromorphic brain interfaces could open a new way of how a human being interacts with an artificial system, in which technical objects would no longer exist as pure executants of human commands. Rather, their mode of existence could participate in a co-construction of the set of norms that guide the actions of a human being through an open and reciprocal relationship established between the artificial and biological networks, each of which having its activity nurtured by that of the other one. In such a scheme, one might thus be invited to consider human subjectivation under a different angle. For example, if an implantable neural prosthesis could help a person to recover some autonomy so that they could take a decision based on the reciprocal cooperation between their biological brain activity and the artificial network activity, could, paradoxically, this person still be held as the author, or in other words, the subject, of this decision?

To address this question, it is necessary to specify what is meant by ‘being subject’. The concept of subjectivation has a long philosophical history. First, it can be said to mean being capable of self-reflection, self-awareness, as well as being responsible; in other words, being able to explain the reasons and develop justifications for one’s thoughts, decisions (or choices) and actions. These capacities participate in making each human individual a subject by helping them to be in control of their existence: to become conscious of their environment, of themself9, and of the relationship between themself and their environment. Second, an important feature of subjectivation is that it is not something given nor acquired once for all, but rather a continuous and dynamic construction process based on the experiences that an individual makes during their existence. Therefore, being subject is not a pre-established but an ever-evolving state continuously undergoing transformations, for example, in the form of a heteronomy, when a person’s will gets influenced by a third-party’s will.

In the case of neuromorphic brain interfaces, introducing a bidirectional relationship between the biological and the artificial co-influencing each other in real time, the major challenge regarding subjectivation would be to find the technical conditions ensuring an ethical relationship between the user and the technology, under which the technology does not subdue its user. To ethically respect the subject’s integrity, the development of neuromorphic brain interfaces would thus have no other choice but to think, well upstream along their conception, the technological conditions ensuring an access to the raisons d'être of decisional recommendations (principle of explicability), leaving the subject sovereign over themself. In other words, leaving the person with a moral autonomy preserved from any algorithmic heteronomy. If, in some situations, these conditions cannot be met, the neuromorphic prosthesis could then be no longer considered as a tool with a strictly utilitarian mode of existence, but possibly as an artificial social agent acting alongside the biological social agent, further bringing neuromorphic interfaces into the current large and complex ethical reflection on the juridical and moral status of certain non-human entities. For example, two independent legal decisions have recognized the Argentinian Chimpanzee ‘Cecilia’ and the New-Zealand’s Whanganui River as subjects with rights. In both cases, the interiority and/or complexity of a non-human entity was appreciated to rival those of humans. Along this line, when artificial intelligence (AI) systems or other so-called ‘intelligent’ machines are used to resolve moral dilemmas, they have been considered to possess a beginning of autonomy, responsibility or rationality, at different degrees, underlying a juridical debate of whether they deserve to be considered as subjects. If the status of ‘electronic person’10 would eventually become recognized for an autonomous AI entity, then what would become its status when integrated with brain networks of a human individual to become the functional substrate underlying their subjectivation?

For all these reasons, how to preserve human subjectivation constitutes a major ethical issue in the development of neuromorphic brain interfaces, which cannot be apprehended after, but all along the design of these technologies.