Why Science Still Needs Human Beings

Artificial Intelligence, Practical Wisdom, and the Human Responsibility of Research

In May 2026, the journal Nature posed a question that might at first appear surprising: Can artificial intelligence conduct good science without human beings? Behind this question lies the development of so-called “AI scientists” — systems of cooperating artificial-intelligence agents capable of reading research literature, formulating hypotheses, analysing data, debating possible explanations, and proposing new experiments. These are no longer merely programs that assist researchers with calculations or literature searches. Such systems are beginning to enter parts of the research process that have traditionally been understood as expressions of human intellectual activity.

The results are impressive. The AI system Robin was used to investigate possible treatments for dry age-related macular degeneration. The system reviewed the research literature, developed a therapeutic strategy, identified relevant molecules, and proposed experiments. Another system, Co-Scientist, investigated, among other things, whether existing drugs could be repurposed to treat leukaemia and liver fibrosis. It also formulated within a few days a hypothesis concerning antibiotic resistance that a group of researchers had been working on for nearly ten years.

Yet behind the narrative of artificial scientists lies a less dramatic, but philosophically more interesting, reality. The systems did not work alone. Human beings formulated the problems, selected relevant goals, carried out the experiments, evaluated the results, and corrected the systems’ errors. What is presented as AI-driven research was, in reality, a new form of collaboration between human beings and machines.

This is not a weakness of the technology. It is the very condition under which the technology can become part of responsible scientific practice.

The question, therefore, is not merely what artificial intelligence can do. The decisive question is what kind of activity science actually is. Is science primarily an efficient method for producing knowledge? Or is science also a human practice requiring experience, judgement, responsibility, curiosity, courage, and understanding?

Viewed through the lens of practical philosophy, the debate concerns more than technology. It concerns the relationship between knowledge and wisdom, between efficiency and insight, between calculation and understanding — and ultimately between what we are able to do and what we ought to do.

When Efficiency Becomes the Goal of Science

Perhaps the most striking statement in the Nature editorial is the following: we do not yet know whether greater efficiency also entails greater insight.

Artificial intelligence can make certain research processes considerably faster. The Robin system is reported to have reduced the time required for a particular research process by a factor of approximately 200 compared with a conventional human workflow. Figures of this kind seem almost irresistible. When research funding is limited, diseases require treatment, and scientific problems become increasingly complex, it can appear irresponsible not to make use of technologies that accelerate the work.

Yet faster research is not necessarily better research.

Here we encounter a fundamental tension already described by Aristotle. He distinguished between different forms of knowledge. Epistēmē referred to theoretical and scientific knowledge of what is universal and necessary. Technē referred to practical knowledge of how something can be produced or performed. Human action, however, also required phronēsis — practical wisdom.

Practical wisdom is not merely the ability to identify the most effective means to a given end. It also involves the capacity to judge whether the end itself is good. A technically rational actor asks: How can this be done as quickly and precisely as possible? The practically wise person also asks: Why should this be done? Who will be affected? Which considerations are in conflict? What would be right in this particular situation?

Artificial intelligence can optimise a process when the objective and the criteria have already been defined. Yet the selection of the objective is not merely a computational problem. Why should society prioritise one disease rather than another? How much risk can be justified in an experiment? When is uncertainty so great that research ought to be stopped? Who should have access to an expensive treatment if it proves successful?

These are not questions that can be resolved simply by processing more information. They require practical judgement.

The Aristotelian insight therefore demonstrates why efficiency can never be the sole aim of science. Efficiency tells us how quickly or accurately we are approaching an objective. It does not tell us whether the objective is worth pursuing.

Science as a Human Practice

Alasdair MacIntyre developed Aristotle’s thought further by describing human activities as practices. A practice is a socially established and complex form of activity that possesses its own internal goods and standards of excellence. Medicine, education, art, and science are examples of such practices.

The internal goods of science cannot be reduced to the number of publications, patents, citations, or commercial products. They include an increasingly profound understanding of the world, methodological integrity, intellectual honesty, openness to criticism, collegial scrutiny, and the ability to acknowledge error. These goods can be fully understood only by those who participate in the practice over time.

An AI system may contribute to the practice, but it is not automatically a participant in MacIntyre’s sense of the term. It has no scientific biography. It has not experienced doubt following a failed experiment, resistance from colleagues, or the gradual transformation of its own understanding. It does not stand within an academic community in which it must defend its claims and assume responsibility for their consequences.

A system can produce a hypothesis, but it has no personal or moral relationship to the truth of that hypothesis.

This does not mean that machine-generated hypotheses are without value. A hypothesis should not be judged according to who or what formulated it, but according to how well it can be justified and tested. Nevertheless, science is more than the production of testable propositions. It is also a practice in which human beings learn to be attentive, critical, persistent, and truthful.

MacIntyre warns that the internal goods of a practice can be displaced by external goods such as money, prestige, and power. In today’s academic system, efficiency can become such an external good. When universities measure research through publication volume, competition for funding, and international rankings, artificial intelligence may be used to intensify an already instrumental system. It may make research faster without making it wiser.

The challenge, therefore, is not merely to integrate artificial intelligence into existing research practices. We must also ask what kind of practice the technology is being integrated into. If the research system rewards speed and production rather than reflection and quality, artificial intelligence may accelerate precisely those developments about which we ought to be most concerned.

Gadamer and the Historical Character of Understanding

Hans-Georg Gadamer can help us understand why scientific insight cannot be reduced to information processing. For Gadamer, understanding is never a purely technical operation. We always encounter the world from a particular standpoint, shaped by language, history, experience, and tradition.

The researcher’s pre-understanding is therefore not merely a source of error to be eliminated. It is also the condition that enables something to appear as a meaningful question in the first place.

A dataset does not speak for itself. What is identified as a relevant pattern depends on which questions are asked, which concepts are employed, and within which scholarly context the data are interpreted. Even highly advanced AI systems work with materials, categories, and objectives that have already been shaped by human research and human interests.

Artificial intelligence does not emerge outside history. It is trained on texts, data, and images produced by human beings. It therefore inherits both the knowledge of science and its blind spots. If certain population groups are poorly represented in medical datasets, a more efficient analytical system will not automatically correct the imbalance. On the contrary, it may lend the imbalance an appearance of objectivity.

Gadamer’s hermeneutics also reminds us that understanding involves placing one’s own pre-understanding at risk. In a genuine encounter with a text, an experience, or another human being, we must remain open to the possibility that what we encounter may change us.

An AI system can update its calculations in response to new data. It is more difficult, however, to say that it places itself at risk. It has no identity or understanding of life that might be transformed. It can revise a conclusion, but it does not experience the intellectual humiliation that may follow from discovering that a conviction on which one has built one’s work was mistaken.

For the researcher, such an experience can be decisive. It teaches us that knowledge grows not only through confirmation, but also through rupture, surprise, and defeat.

Arendt and the Difference Between Labour and Action

Hannah Arendt distinguished between labour, work, and action. Labour is connected to the necessities and repetitions of life. Work creates lasting objects and technical products. Action arises between human beings in a shared world. It cannot be fully predicted or controlled because every human being possesses the capacity to begin something new.

Science contains all these dimensions. Much laboratory work is routine. Instruments must be calibrated, literature reviewed, and data organised. In these areas, artificial intelligence can relieve researchers of burdensome tasks and free time for other work.

Science, however, is also action. The researcher appears before others with a claim about the world. The claim is criticised, interpreted, and situated within an academic and social community. It may alter how human beings understand nature, illness, or themselves.

This public dimension cannot be reduced to information processing. Someone must come forward and say: This is what we have found. These are our reasons. This is the uncertainty involved. These are the consequences we can foresee. And this is the responsibility we are prepared to assume.

An AI system can generate a report, but it cannot appear as a responsible subject. It cannot be held morally accountable in the same way as a researcher, a university, or a company. When an erroneous conclusion causes harm, the system cannot feel guilt, offer a sincere apology, or commit itself to acting differently in the future.

Arendt’s philosophy thus reveals a decisive distinction between producing a result and assuming responsibility for that result’s place within a shared world.

Kant and the Human Being as an End

Immanuel Kant’s moral philosophy offers another perspective on why human responsibility cannot be replaced by machine efficiency. In one formulation of the categorical imperative, Kant argues that human beings must always be treated as ends in themselves and never merely as means.

This principle is particularly important in medical research. Participants in clinical trials are not merely data points. Patients are not simply carriers of biological mechanisms. They are persons with dignity, experiences, hopes, and fears.

An AI system may identify statistically optimal strategies. Yet it cannot, by its own power, recognise human dignity. Such recognition presupposes a normative relationship between persons. It requires that the other person is not perceived merely as a source of information or a means of producing knowledge.

This does not mean that artificial intelligence necessarily treats human beings worse than researchers do. History contains numerous examples of human researchers violating research participants, neglecting vulnerable groups, and allowing scientific ambition to eclipse moral boundaries. Yet such violations can be criticised precisely because human beings are capable of understanding and justifying normative principles.

Technology does not therefore release us from moral responsibility. The greater the power for action that technology gives us, the greater the need for someone to be answerable for how it is used.

Jonas and Responsibility for the Future

Hans Jonas developed his ethics of responsibility in response to the growing power of modern technology. Traditional ethics had primarily concerned actions between human beings living at the same time. Modern technology makes it possible to affect individuals, societies, and ecological systems far into the future.

Jonas therefore formulated a new imperative: we must act in such a way that the effects of our actions are compatible with the continued existence of genuinely human life.

Artificial intelligence in science brings this idea into renewed relevance. New medicines, genetic interventions, and biological technologies may have consequences that are difficult to foresee. When the research process accelerates, the distance between discovery and application may also become shorter. What can technically be developed may quickly come to appear as something that ought to be implemented.

Greater speed, however, can reduce the time available for ethical and social reflection.

Jonas emphasises the importance of caution when the consequences may be far-reaching and irreversible. This does not mean that research should cease whenever uncertainty exists. All research proceeds under conditions of uncertainty. It does mean, however, that uncertainty itself must be taken morally seriously.

An AI system may calculate probabilities on the basis of available data. It cannot by itself determine what risks human beings have the right to impose on others or on future generations. This judgement requires responsible actors who can be held accountable both for what they did and for what they failed to do.

The Detours of Curiosity

The Nature editorial draws attention to something that often disappears in the image of science as an efficient production process: human messiness, curiosity, and playfulness have contributed to countless discoveries.

The history of science is full of detours. Experiments fail. Instruments malfunction. Researchers follow paths that turn out to be blind alleys. Results appear that do not fit the original hypothesis. Conversations in a corridor or chance encounters between disciplines may open entirely new perspectives.

From a purely efficiency-oriented perspective, such events may appear wasteful. Yet the detours may constitute the learning space of knowledge.

John Dewey described thinking as an activity of inquiry arising when human beings encounter an uncertain or problematic situation. We act, experience the consequences, and revise our understanding. Knowledge emerges through an interaction between the human being and the environment.

Research is therefore not simply the application of an already established method. The method itself may need to change when reality resists it. An experienced researcher may notice that something “does not fit” even before the discrepancy can be precisely articulated. This sensitivity is not mysterious. It is formed through prolonged participation in a practice.

Michael Polanyi called this tacit knowledge. We can know more than we are able to state explicitly. A researcher may recognise an unusual pattern, understand that an instrument is behaving differently from normal, or sense that an apparently negative result may nevertheless be significant. Such knowledge is embodied, situated, and grounded in experience.

AI systems may undoubtedly identify patterns that human beings overlook. Human experience is not thereby rendered obsolete. On the contrary, the encounter between machine pattern recognition and human tacit knowledge may become one of the most fruitful areas of future research.

The Place of Empathy in Science

It may seem unusual to claim that science requires empathy. The ideal of science is often associated with distance and objectivity. Results should not depend on the researcher’s emotions.

Empathy, however, does not mean that feelings should replace evidence. Empathy makes it possible to understand what knowledge means in human lives.

A medical researcher must know not only whether a treatment reduces a symptom. The researcher must also ask how the treatment affects the patient’s everyday life, dignity, relationships, and sense of self. What is statistically significant is not always humanly meaningful. And what matters most to a person is not always easily measured.

Martha Nussbaum has shown how emotions may contain judgements about what has value. Compassion arises when we understand that another person’s suffering is serious, undeserved, and connected to a vulnerability that we ourselves share.

Scientific empathy is therefore not sentimentality. It is an expansion of attention. It helps the researcher perceive who is affected by a problem, who is not being heard, and which experiences disappear when reality is reduced to measurable variables.

A system may be trained to recognise expressions of pain or preferences in patient data. Yet it does not share human vulnerability. It does not know what it means to wait for a diagnosis, to lose functional capacity, or to fear that a treatment will not work.

Artificial intelligence may therefore contribute to research on suffering, but it cannot by itself determine what suffering requires of us.

The False Opposition

Debates about artificial intelligence are often trapped between two extremes. On the one hand, the technology is portrayed as an almost autonomous intelligence that will soon solve problems with which human beings have struggled for centuries. On the other hand, it is described as a threat to scientific quality, researchers’ employment, and the human place in the world.

Both perspectives contain an element of truth, but the opposition is misleading.

In 1989, Max Perutz asked whether science was the noblest pursuit of the human mind or a sorcerer’s broom that threatened humanity with destruction. In our own time, the question can be reformulated: Is artificial intelligence the liberation of science or its dehumanisation?

Practical philosophy teaches us to be cautious about such either–or questions. Technology has no unambiguous moral significance independently of the practices and institutions into which it is incorporated. It can relieve burdens or increase control, liberate or marginalise, open new questions or narrow our attention.

The decisive issue is not whether artificial intelligence should be used in research. It is already part of research and will become increasingly important. The question is how it should be used, which tasks it should assume, and which tasks human beings must continue to bear responsibility for.

A responsible division of labour may be based on a simple principle: machines may assist with calculation, searching, simulation, and pattern recognition. Human beings must continue to assume responsibility for problem formulation, interpretation, justification, ethical judgement, and public accountability.

The boundary will not always be clear. AI systems can already contribute to interpretation and hypothesis formation, while human beings have long delegated parts of their judgement to standardised procedures. Yet the lack of clarity does not make the question less important. It makes continuous critical reflection necessary.

The Human Researcher After Artificial Intelligence

Artificial intelligence will also transform what it means to educate a researcher. If literature searches, data analysis, and article drafts are increasingly performed by machines, young researchers may lose some of the slow experiences through which scholarly judgement was formerly developed.

A doctoral student does not learn only by arriving at correct answers. Learning also occurs through reading texts that prove less relevant than expected, formulating poor hypotheses, misunderstanding data, and receiving criticism. Efficient tools may remove unnecessary work, but they may also remove parts of the learning process.

This creates an educational paradox. We wish to use artificial intelligence because it makes research faster. Yet if researchers no longer learn to perform the work that the systems take over, they may gradually lose the ability to evaluate the systems’ results.

Human oversight cannot merely mean that a human being presses an approval button. Genuine oversight presupposes understanding. The researcher must possess sufficient disciplinary knowledge to identify when the system is operating on false premises, misusing data, or producing a convincing but erroneous explanation.

The education of future researchers must therefore teach students not only how to use artificial intelligence, but also when they should not trust it.

Science’s Need for Humanity

Can artificial intelligence conduct good science without human beings?

The answer depends on what we mean by science. If science is understood as the production of hypotheses, the analysis of data, and the identification of patterns, artificial intelligence can already perform substantial parts of the work. In some areas, it can do so faster and on a greater scale than any human being.

If, however, science is also a practice in which human beings attempt to understand the world, justify their claims, acknowledge uncertainty, and assume responsibility for the consequences of knowledge, the human being remains indispensable.

Aristotle reminds us that technical skill is not the same as practical wisdom. Kant reminds us that the human being must never be treated merely as a means. Gadamer demonstrates that understanding is always historical and interpretative. Arendt makes visible the difference between producing a result and acting responsibly within a shared world. Jonas teaches us that technological power requires responsibility for future consequences. MacIntyre reminds us that science is a practice possessing internal goods that may be undermined when efficiency, prestige, and production become dominant.

Artificial intelligence does not merely challenge the human place in science. It compels us to articulate more clearly what is valuable about human research.

What is valuable is not that human beings always think faster or calculate more precisely. It also lies in our capacity to wonder, doubt, regret, learn from defeat, understand the suffering of others, and question our own objectives. We can be held responsible because we are capable of answering for what we do.

Good science will therefore probably be neither purely human nor purely machine-based. It will be a collaboration in which technology expands human possibilities without displacing human responsibility.

Science needs artificial intelligence because reality is complex and the amount of knowledge exceeds what any individual can master. Yet science still needs human beings because knowledge is never only a matter of finding an answer.

It is also a matter of understanding what the answer means — and of assuming responsibility for what we do with it.

References

Arendt, H. (1958). The human condition. University of Chicago Press.

Aristotle. (2009). The Nicomachean ethics (D. Ross, Trans.; L. Brown, Rev.). Oxford University Press.

Dewey, J. (1938). Logic: The theory of inquiry. Henry Holt.

Gadamer, H.-G. (2004). Truth and method (J. Weinsheimer & D. G. Marshall, Trans.; 2nd rev. ed.). Continuum. (Original work published 1960)

Ghareeb, A. E., et al. (2026). A multi-agent system for automating scientific discovery. Nature. https://doi.org/10.1038/s41586-026-10652-y

Gottweis, J., et al. (2026). Accelerating scientific discovery with Co-Scientist. Nature. https://doi.org/10.1038/s41586-026-10644-y

Jonas, H. (1984). The imperative of responsibility: In search of an ethics for the technological age. University of Chicago Press.

Kant, I. (2012). Groundwork of the metaphysics of morals (M. Gregor & J. Timmermann, Eds. & Trans.; Rev. ed.). Cambridge University Press. (Original work published 1785)

MacIntyre, A. (2007). After virtue: A study in moral theory (3rd ed.). University of Notre Dame Press.

Nature. (2026). Why AI cannot do good science without humans. Nature, 653, 650. https://doi.org/10.1038/d41586-026-01551-3

Nussbaum, M. C. (2001). Upheavals of thought: The intelligence of emotions. Cambridge University Press.

Perutz, M. F. (1989). Is science necessary? Essays on science and scientists. Oxford University Press.

Polanyi, M. (1966). The tacit dimension. Doubleday.