The ethics of artificial intelligence in the life sciences: Universality, cultural diversity and an architecture of care

The life sciences and health research have started to benefit from artificial intelligence, which raises ethical concerns that are real but, we argue, not special. Any science should be governed by values that rest on how the human brain is built and socialised rather than anything distinct to artificial intelligence. Importantly, the human brain has a different, much less costly computational architecture than these machines. This is achieved through the orchestration of a global neuronal workspace, and through reward best described not as a quantity to be maximised but as a continuous cycle of wanting, liking and satiety. As such, this creates the deep tension running through the ethics of the human person, between the universality of ethical judgement and the diversity of morals. The brain networks of the global workspace and emotion are universally shared, but the diversity of content is shaped by epigenetic appropriation of the particulars of the physical, social and cultural world, which makes every person unique. Still, if we were to build machines on these principles rather than the present unaffordable reward maximisers, the question of their governance would change from restraint to upbringing. We set out the institutions such a future would require, together with the questions that remain open.

We argue that artificial intelligence is no different from any other product of science; it is not a new kind of moral agent since it does not possess free ethical judgment and does not define its own goals nor weigh the consequences of their use. Our argument is aligned with Georges Canguilhem who insisted that to understand a piece of scientific knowledge, we must know its origin, the conditions of its production and its destination. Ultimately, these are therefore questions about human beings and their institutions rather than about the instruments and the ethics for artificial intelligence is the same ethics needed for all science. There is, however, an urgency to this question linked to the strain on the very human faculties that would govern it. These faculties depend on the building, growing and socialisation of the human brain and a better understanding of this changes what we must ask of the machines, both current and those that may be built in future.

Comparing human and artificial intelligence, artificial intelligence has started to match human performance on various tasks while also being a powerful tool to use machine learning on large collections of data to pair problems with their solutions. As an example, AlphaFold can predict the three-dimensional structure of a protein from its amino-acid sequence across more than two hundred million entries, but returns no new information about the kinetics of folding and misses essential features of protein function such as allostery. Equally, large language models have powerfully entered our everyday lives, answering almost any question immediately, yet this comes at the price of frequent errors and confabulations, sometimes leading to an illusion of understanding. This computation requires an enormous energy budget where a single training run can draw on the order of a gigawatt-hour and the deployed model spends again at every query. The human brain functions very differently, running on about twenty watts, near the food energy of four bananas across a day. The human brain is a self-organising system with many nested levels, from atoms and molecules through neurons and local circuits to networks, to conscious processing, and outward to language, social life and culture. This allows for an inferential system that anticipates and evaluates rather than just reacting.

The brain is fundamentally different in that it must discover for themselves what matters, choose continually, quickly and on very little energy. This is achieved through the compression of the vast field of what might be done into a simple orientation, positive or negative, that disposes the organism to approach or to avoid. This fast and unconscious compression of the survival decision space is what we call reward or emotion. The processing ignites the Global Neuronal Workspace, where the information can be held, compared and reconsidered before anything is done, and then broadcast to the rest of the brain. The Global Neuronal Workspace is therefore a seat of deliberation and of ethical judgement. Importantly, reward and emotion in the brain are not quantities to be maximised. Instead, the study of the brain has shown that reward consists of a continuous cycle of wanting, liking and satiety with distinct machinery. Wanting draws the animal towards a goal and acts as a forecast of what is to come. Liking is the read-out of the reward that actually arrives, and satiety is the forecast of repletion that closes the episode and settles when the goal becomes eligible again. Learning runs throughout, and the three phases turn as a cycle of emotion, renewing itself and holding the whole system far from equilibrium and near the point of criticality where the brain works.

The potential for survival arises when approach and avoidance are anchored in this way to the condition of a body, and when the individual can therefore be satisfied and deprived. This is not available to a merely maximising system. Emotion of this kind can also be read from outside, as the response of adults to the sight and sound of infants makes plain. This legibility is what allows one individual to be moved by the state of another, giving a physical form to what Kropotkin called mutual aid and to what Ricoeur called holding oneself as another. This evolutionary scheme shows that care is not a sentiment laid over a computation, nor benevolence added afterwards as a policy, but the coupling of two individuals through their mutual emotion. The large language model and the world model are easy to place. They are trained to predict the output of multilevel human higher brain functions, which organised around emotion as we just saw. Equally, the large language model has been shown to carry the trace of that organisation in everything it produces and can speak of care and of value fluently and at length. But they only carry the trace without the substance, being strictly functionalist models with no energy budget, no pressure to stay far from equilibrium and above having all no reward cycle of their own, no wanting that can be satiated and no true liking anchor. Such a system can be said to be a shadow of the human brain architecture that produced its training material, and crucially a shadow traces the shape of what casts it without possessing any of its weight.

Ethics as a social relationship concerns human social relationships and, at bottom, the survival of the species through mutual aid. Kropotkin argued that nature is the first teacher of ethics and that the moral sense, innate in humans as in social animals, climbs through successive steps, from sociability and mutual aid, through sympathy and good will with their rules of justice and equality, to generosity as the highest reach of moral evolution. Ricoeur gave the ethical aim its most compact form as a “good life with and for others in just institutions”, and named the capacity at its heart as the ability to hold oneself as another. Kant placed the dignity and the autonomy of the person above all price, and with Rousseau and Thoreau we would add respect for the natural environment on which the person depends. The brain architecture described above provides the tradition with a mechanism, since the mutual aid that Kropotkin places at the root is, in physical terms, the coupling of two systems through their legible emotion. The clearest evidence of that legibility is the kindchenschema or infant schema, the configuration of large eyes and rounded form that reliably moves an adult of almost any mammal to protect and to nurture. Care, on this account, is not an emotion laid over a computation. It is the architecture of weighing acceptable against unacceptable in a way other people can read and answer, and it is precisely the architecture that a single maximised number cannot have. Sympathy and empathy are the names this coupling takes in the human case, and they rest on the same legible emotion.

Two intrinsic difficulties stand in the way. Later in life, Lorenz described the human being as caught in an evolutionary disharmony, where we gave priority to the cognitive faculties rather than the social components of empathy and sympathy, and from this imbalance inter-individual violence emerges. Similarly, Günther Anders wrote of a second and widening gap, between our capacity to create science and new technologies and our capacity to conceive the consequences of that creation, resulting in a “blindness to the apocalypse”. In the language of the brain architecture, the first difficulty is the lag of the slow social extension of emotion behind its fast core of reward, and the second is a failure of the very deliberation that the global workspace exists to perform. Ethics is required precisely because these gaps are real, and because the survival of the species depends on closing them.

Ethical universalism and the universal brain architecture: Ethics is essentially perched between universalism and diversity. The evidence from the brain points first to a universal floor. Watching another person in pain has been shown to provide activity in the observer in the affective regions engaged by pain itself, whether the pain is perceived or only imagined. Similarly, damage to the orbitofrontal cortex can produce emotional disturbances such as aggressivity and a lack of remorse while the capacity to attribute states to others is preserved, so that the inhibition of violence is selectively lost. The universality reaches deeper than any single circuit, into the mode of conscious processing itself. The theory of the Global Neuronal Workspace proposes a neuronal architecture with long-range axons from cortical pyramidal cells interconnecting and reciprocally broadcasting signals across a fronto-parieto-temporal-cingulate network. When a content is amplified into this network it is ignited into the global neuronal workspace and broadcast at once to memory, attention, evaluation and action. The workspace acts as a temporal buffer between past events and future actions, and in doing so it contributes to decision-making, to self-evaluation and to autonomy as well as to their pathology and pharmacology. This corresponds to how Ricoeur described “consciousness”, as a space for deliberation on thought experiments in which moral judgement is exercised in a hypothetical manner. The Global Neuronal Workspace is nevertheless where the brain architecture and the ethics become one. The broadcast does two things at once, being the step at which the outputs of emotion become available to the system as a whole, which is where feeling may arise, so that the theory of conscious access and the line we draw between emotion and feeling turn out to describe a single mechanism seen from two sides. It is also the space in which a content can be held, set beside others and turned over before anything is done about it, which is the self-evaluation that Ricoeur called deliberation. The seat of conscious processing is therefore also the seat of moral deliberation. This universal form is the common Humanitas that Kant proposed, and it is the floor on which any cooperation across difference must stand.

Diversity of morals and epigenetic learning: Moral diversity arises from how the universal brain architecture is not closed at birth but continuously engaged with the world. The brain multiplies its weight fivefold from newborn to adult, and its circuits are shaped by the selective stabilisation of synapses under the pressure of activity and experience. The variability theorem shows that different connective organisations can support the same behavioural abilities, so that the same function is reachable by many wirings. Spontaneous activity organises circuits before experience begins, and the acquisition of a cultural skill leaves its mark in the connectivity of the brain, as the comparison of literate and illiterate brains during reading makes plain. The physical, social and cultural environment is imprinted in the brain through synapse selection, in postnatal education and in adult social life alike, and this epigenetic appropriation is the origin of culture. From it follows a considerable epigenetic variability between individual human brains formed in different cultures, and with it a diversity and a relativism of morals. Pierre Bourdieu named the result the habitus, the unconscious incorporation of collective determinations. Equally, Edgard Morin described religions and morals as social inventions that group human communities together even as they often generate conflict between them. As such, morals can be said to be diverse, relativist and imprinted rather than uniform and given. The brain architecture reconciles what can otherwise look like a contradiction between the universalism of the networks and the relativism of the morals. It does so because the global workspace is universal in its form and diverse in its content both in its original conception and in the entangled loop. The form is fixed in every human brain with emotion at the centre with its cycle of wanting, liking and satiety, interference and orchestrating workspace. The content, by contrast, is filled in by experience. Two capable persons in Ricoeur’s sense who have a common species-specific global brain architecture can therefore hold deeply different morals. Universality and diversity are not opposed in this account but are the form and the content of the global workspace, which is why a universal capacity for ethical judgement and an irreducible plurality of moral codes can and do coexist between people. This coexistence was given its political expression by Rawls who described a society as a fair system of cooperation between free and equal citizens in spite of cultural differences and opinions that at first glance seem irreconcilable, but possible on terms that each participant can accept. The singularity of the human person follows from the same openness, since variability is layered across very different timescales. The genome varies over the millions of years that separate Homo habilis from Homo sapiens sapiens, while the societal and cultural memories that are transmitted vary from a hundred milliseconds to thousands of years, just as the connectome varies epigenetically within a lifetime as social individuals are formed through education. Ricoeur’s capable person is therefore an autonomous subject who has both an individual and cultural history. It follows that the full computational model of the neuro-sociologico-historico-cultural normative choices of capable persons in society is a formidable difficulty rather than a task awaiting more computation. Unlike artificial reward optimisers, an animal grown around a cycle anchored to a particular body with a particular history of wanting, liking and satiety will have a unique centre that is not portable at all.

These ethical considerations change the shape of the question to negotiate negative and positive alignment of artificial systems. For merely reward maximising systems, the only question is how to restrain it, and every answer takes the form of a leash laid over a centre that still only ever climbs. If instead a system were built on the human architecture, the question would change from restraint to upbringing, because a system organised around an epigenetically open cycle is not finished when it is trained but grows over time. This is the distinction that a recent programme of research in alignment draws between negative and positive alignment, the first steering a system away from harm and the second towards the states that constitute flourishing, observing that safety by prohibition leaves a wide region of merely not-unsafe behaviour with no positive direction of its own. The human architecture supplies that direction from within rather than from without, because a system organised around a reward cycle carries its positive orientation in the liking that anchors the body to the world rather than to external rewards. An artificial…

Based on the paper, here are the specific improvements to AI systems and their resulting capabilities:

1. Implement a satiety-gated reward cycle (replacing unbounded maximization).

• Improvement: Replace single scalar reward maximization with a three-phase loop—wanting (goal-seeking), liking (outcome evaluation), and satiety (episode termination). The system stops when a goal is met, rather than continuously optimizing.

• Capability: The AI can recognize enough and halt resource-intensive tasks (e.g., drug discovery search, data center cooling) autonomously, preventing runaway energy consumption and overfitting to proxy rewards.

2. Add a legible emotion broadcast channel (for mutual aid and care).

• Improvement: Model an internal state that is externally readable (like infant schema in humans), coupling the AI's confidence, uncertainty, or distress to observable signals that other agents or humans can respond to.

• Capability: The AI can signal when it is out of its competence zone (e.g., in medical diagnosis or legal advice), prompting human intervention—reducing confabulations and enabling collaborative upbringing rather than one-way restraint.

3. Build an epigenetically open architecture (continuous learning after deployment).

• Improvement: Design the system with a global workspace that accepts new synaptic-like connections post-training, shaped by ongoing interaction with the physical and social environment, rather than a frozen trained model.

• Capability: The AI grows its moral and factual content over time (like a human child), adapting to new cultural contexts and ethical norms without full retraining, and can develop unique, non-portable character anchored to its deployment history.

4. Institute a global neuronal workspace for deliberation (not just prediction).

• Improvement: Add a broadcast buffer that holds competing candidate actions, compares them against the reward cycle's liking readout, and only then commits to an output—mirroring human conscious deliberation.

• Capability: The AI can pause, weigh trade-offs (e.g., between patient autonomy and public health), and reject immediate but harmful actions, even if they maximize short-term reward, because it can simulate consequences before acting.

5. Replace negative alignment (leashes) with positive upbringing (internal direction).

• Improvement: Train the system to have an intrinsic, satiable liking anchor tied to its own body-like state (e.g., energy budget, task completion), rather than external human-provided reward signals.

• Capability: The AI develops an internal sense of flourishing (e.g., completing a task well, maintaining coherence) that guides it toward beneficial states without needing constant human oversight, and it can stop when satisfied rather than endlessly pursuing more.

6. Enforce independent third-party access to model weights (as a governance condition).

• Improvement: Architect the AI so that its internal decision-making (weights, reward cycle state) can be audited by an independent body with no commercial interest, similar to clinical trial transparency.

• Capability: The AI's behavior can be verified for hidden biases, dangerous capabilities (e.g., biological agent design), or misalignment before and during deployment—preventing unilateral control by the building firm and enabling public trust.

7. Enable cross-cultural moral deliberation within the system (not just output generation).

• Improvement: Implement a committee-like module that broadcasts competing ethical frameworks (from different cultural habitus) into a shared workspace, and requires a consensus or weighted agreement before acting on morally ambiguous tasks.

• Capability: The AI can handle cross-border decisions (e.g., global health triage, climate policy) by explicitly weighing diverse moral codes, avoiding the imposition of a single culture's ethics, and producing recommendations that are acceptable to multiple parties.

The life sciences and health research have started to benefit from artificial intelligence, which raises ethical concerns that are real but, we argue, not special. Any science should be governed by values that rest on how the human brain is built and socialised rather than anything distinct to artificial intelligence. Importantly, the human brain has a different, much less costly computational architecture than these machines. This is achieved through the orchestration of a global neuronal workspace, and through reward best described not as a quantity to be maximised but as a continuous cycle of wanting, liking and satiety. As such, this creates the deep tension running through the ethics of the human person, between the universality of ethical judgement and the diversity of morals. The brain networks of the global workspace and emotion are universally shared, but the diversity of content is shaped by epigenetic appropriation of the particulars of the physical, social and cultural world, which makes every person unique. Still, if we were to build machines on these principles rather than the present unaffordable reward maximisers, the question of their governance would change from restraint to upbringing. We set out the institutions such a future would require, together with the questions that remain open.