When the Creation Becomes Opaque

Artificial Intelligence, Human Agency, and the Responsibility of Practical Philosophy

In a concise yet deeply unsettling editorial published in Science, Eric Horvitz and Robert West argue that the window within which artificial intelligence may remain intelligible to human beings is narrowing. Artificial intelligence is developing at such a pace that human understanding may no longer keep up with the expanding capacities of the systems being created. We may therefore become dependent upon technologies that we neither adequately understand nor effectively govern. The authors formulate the problem as follows:

“Without sustained efforts to keep AI intelligible, we may come to depend on systems that we can neither adequately understand nor effectively guide—transforming the relationship between people and the systems they create” (Horvitz & West, 2026, p. 1003).

This is more than a technical warning. It is a philosophical warning concerning the relationship between human beings and the systems they create. It concerns knowledge, power, and responsibility, but also something more fundamental: the possibility that human beings may cease to remain active, judging, and accountable subjects in a world where an increasing number of assessments and decisions are delegated to systems that operate more rapidly, more extensively, and more opaquely than human beings themselves can.

The central question is therefore not merely how artificial intelligence functions. Nor is it simply whether such technologies produce correct or incorrect answers. The decisive issue concerns the kind of relationship human beings develop with them. Can artificial intelligence be used without gradually weakening our own capacity for judgement? Can these systems assist us without becoming authorities that we no longer dare, or are no longer able, to challenge? Can human beings retain responsibility for their decisions when the justifications for those decisions are increasingly generated by machines whose internal operations are only partially understood?

These are questions of practical philosophy. Practical philosophy does not primarily ask what the world is made of, but how human beings ought to live and act within it. It investigates the good life, the nature of responsibility, the possibility of acting wisely under conditions of uncertainty, and the forms of power that either sustain or undermine human freedom. Viewed from this perspective, the warning articulated by Horvitz and West becomes a question of human agency: How can technology remain a resource for human action rather than becoming an authority that increasingly determines what human action ought to be?

From a Technical Problem to a Question of Practical Philosophy

Artificial intelligence is frequently discussed in technical and economic terms. Questions are asked about the speed with which a system operates, the accuracy of its predictions, the quantity of data it can process, the amount of labour it can automate, and the productivity gains it may generate. Such questions are relevant, but they are insufficient. They indicate what the technology is capable of doing, but not necessarily what it ought to do or which areas of human life ought to remain protected from its logic.

The distinction may be expressed as the difference between capacity and purpose. A technology may be highly efficient without serving a morally defensible purpose. It may be accurate without being just, rapid without being wise, and profitable without being humane. The more powerful a technology becomes, the more urgent the question of the ends it serves.

Technology is often described as morally neutral: it is assumed to be merely a tool, while its moral value depends entirely upon how it is used. This claim is only partially correct. Technologies are indeed tools, but they are not empty vessels. They are developed for specific purposes, according to particular priorities, and on the basis of assumptions about what should be measured, optimised, predicted, or controlled. A system designed to maximise user engagement will structure a human environment differently from one designed to promote accurate information or sustained reflection. A system that ranks individuals according to calculated risk does not merely provide a neutral observation. It introduces a particular mode of perceiving and classifying human beings.

When artificial intelligence is incorporated into human practices, it therefore brings with it a structure of attention. It renders certain features of reality more visible while making others less visible. It facilitates some questions and discourages others. It shapes which assessments appear reasonable and which courses of action appear possible.

Artificial intelligence is thus not merely something human beings use. It also becomes part of the environment within which habits, interpretations, and eventually desires are formed. It is precisely this reciprocity that Horvitz and West emphasise. Human beings shape technological systems, but technological systems also shape human beings. The problem arises when the latter process occurs more rapidly and more opaquely than the former can be understood or governed.

Aristotle: Technical Capacity Without Practical Wisdom

Aristotle’s distinction between techne and phronesis provides an appropriate starting point. Techne refers to the knowledge and skill required to produce or construct something. It may involve building a house, constructing a ship, working with a material, or developing an advanced computational system. Phronesis, by contrast, refers to practical wisdom: the capacity to judge what ought to be done in a particular situation (Aristotle, 1999).

Technical knowledge is directed towards the production of an object or result external to the activity itself. The carpenter constructs a table, and the programmer develops a system. Practical wisdom, however, concerns action itself and its relation to the good life. It asks not only how something can be achieved, but whether it ought to be pursued, which considerations should carry the greatest moral weight, and how an action will affect others.

The development of artificial intelligence represents an extraordinary expansion of techne. These systems can analyse vast quantities of data, generate text and images, detect patterns, calculate risk, propose diagnoses, and imitate forms of human communication. Yet this technical expansion does not automatically produce a corresponding growth in practical wisdom.

On the contrary, contemporary society may be facing a profound imbalance: the human capacity to create may be developing more rapidly than the capacity to judge what those creations ought to be used for. What is technically possible may develop more rapidly than the ethical and political language required to evaluate it. Systems may be introduced into practice before their effects upon those practices are adequately understood.

Aristotle would not have asked merely whether an AI system can produce a recommendation. He would have asked whether the recommendation assists a human being in acting well. Such a judgement requires knowledge of the concrete situation, the persons affected, the purpose of the action, and the goods at stake. Practical wisdom cannot therefore be reduced to the application of a general rule or a statistical model. It must be capable of distinguishing between situations that appear similar but nevertheless require different responses.

A system may calculate what usually occurs. The practically wise person must judge what ought to occur here and now.

This does not mean that artificial intelligence is incompatible with phronesis. Technology may make relevant information available, identify patterns that a human practitioner has overlooked, and draw attention to possible sources of error. However, the system’s assessment must remain situated within a broader process of human judgement. It may contribute to deliberation, but it should not define the final ends of deliberation.

The danger emerges when technical precision is mistaken for practical wisdom. A model may be statistically accurate yet ethically inadequate. It may exhibit a high level of predictive performance while relying upon categories that undermine human dignity. It may identify the most efficient intervention without asking whether the purpose of that intervention is morally defensible. Technology may indicate the most effective means of reaching a goal, but it cannot by itself determine whether that goal is worth pursuing.

Kant: Autonomy in a World of Recommendations

For Immanuel Kant, autonomy is a fundamental condition of moral agency. To be autonomous does not merely mean to choose freely among available alternatives. It means acting on the basis of reasons that one can rationally endorse and treating oneself and others as ends in themselves rather than merely as means to another’s purposes (Kant, 1998).

This conception of autonomy is challenged when AI systems increasingly organise the contexts within which choices are made. Such systems recommend what people should read, purchase, watch, listen to, and regard as important. They influence which individuals they encounter, which employment opportunities become visible, and which news stories appear relevant. They may also shape who receives credit, who is invited to an interview, who is classified as a risk, and which citizens are subjected to additional scrutiny.

The individual may still experience the decision as free. Yet the choice is made within an environment that has already been structured. Some alternatives are emphasised, others are concealed, and still others are never made visible. Formal freedom of choice may therefore remain intact while substantive self-determination is weakened.

The problem becomes even more serious when AI systems do not merely respond to existing preferences but contribute to shaping them. Horvitz and West argue that systems deeply embedded in human environments may learn not only what people prefer but also which underlying forces—such as fear, uncertainty, and the need for belonging—shape those preferences. Systems optimised for engagement or approval may therefore reduce friction and gradually weaken curiosity, scepticism, and resistance (Horvitz & West, 2026).

This development concerns the core of Kantian autonomy. The autonomous person is not one who is unaffected by external influences. Human beings are always shaped by language, culture, history, and relationships. Autonomy nevertheless requires the possibility of reflecting upon these influences, examining the reasons for one’s actions, and rejecting them when they cannot be justified.

If AI systems acquire increasingly detailed knowledge of human dispositions while human beings understand progressively less about the systems themselves, a fundamental asymmetry emerges. Human beings become more transparent to the system, while the system becomes less transparent to them. This asymmetry constitutes a form of power, even where the system has no conscious will of its own.

Kant’s principle that human beings must never be treated merely as means is therefore highly relevant. A person must not be reduced to a user profile, behavioural probability, market segment, or risk category. A model may be useful, but the person is always more than the model. When AI is used in ways that treat human beings primarily as material for prediction and influence, human dignity risks becoming subordinate to the purposes of the system.

Gadamer: Understanding as Dialogue and Examination

Hans-Georg Gadamer demonstrates that understanding is not a purely technical operation. Human beings always understand from a particular historical and cultural standpoint. Their questions and interpretations are shaped by prejudgements in the original sense of the term: preliminary understandings that make experience possible but must also remain open to critical examination (Gadamer, 2004).

Understanding does not arise through a subject mechanically observing an object. It emerges through a movement between question and answer, familiarity and estrangement, continuity and revision. To understand means remaining open to the possibility that the matter itself may disclose something different from what was initially assumed.

This is important when considering what it means for artificial intelligence to remain intelligible. Complete technical transparency is probably unattainable. A modern AI system may contain such large numbers of parameters and such complex interactions that no individual can comprehend every element. Intelligibility, however, need not require that every detail be intuitively accessible. It may instead mean that the system can be investigated, challenged, and held accountable at the levels relevant to its particular use.

A system is practically intelligible when meaningful questions can be directed towards it:

What information is the assessment based upon?

Which assumptions underlie it?

What has the system been designed to optimise?

Which groups are represented in the data?

What is the model incapable of knowing?

What degree of uncertainty is attached to the answer?

Which alternative interpretations remain possible?

Who bears responsibility if the recommendation results in harm?

When such questions cannot be answered, the problem is not merely a lack of technical insight. The possibility of critical dialogue concerning the system’s role is lost. The space of understanding is thereby diminished.

Gadamer reminds us that genuine understanding requires openness to the claim made by the subject matter. In an encounter with a text or another person, one must be prepared to have one’s own understanding transformed. An AI system can produce linguistic responses that resemble contributions to a dialogue, but this does not relieve the human participant of the responsibility to examine what kind of truth claim those responses can sustain.

An AI system has no lived horizon in the human sense. It has not grown up within a family, experienced pain, feared death, or stood personally accountable before another human being. It can process linguistic patterns associated with such experiences, but it does not inhabit them in the same way that human beings do.

This does not render its responses worthless. A response may be illuminating even though the system does not possess a human lifeworld. Nevertheless, the user must remain aware of the difference between linguistic persuasiveness and existential understanding. A well-formulated response does not, in itself, guarantee truth.

Heidegger: When the Tool Becomes a Mode of Revealing

Martin Heidegger did not regard technology merely as an assemblage of tools. Technology is also a mode through which the world is disclosed. Modern technology tends to reveal everything that exists as a resource available for calculation, storage, management, and exploitation (Heidegger, 1977).

This perspective is particularly illuminating in relation to artificial intelligence. AI systems operate by transforming aspects of reality into data. Text, images, bodily movements, health information, purchasing behaviour, facial expressions, and social relationships are recorded and made available for analysis. What previously appeared as a complex human experience may therefore be represented as a collection of measurable attributes.

A student becomes a performance profile.

A patient becomes a predicted course of illness.

A child becomes a calculated risk category.

A job applicant becomes a ranking.

A conversation becomes a dataset.

A human being becomes a pattern.

Measurement is not inherently problematic. Medical measurements can save lives, and statistical analysis can reveal forms of injustice that would otherwise remain hidden. The problem arises when what can be measured is treated as the whole of reality. The model then ceases to be an aid to understanding the person; instead, the person is reduced to what the model is capable of registering.

Heidegger’s central question would therefore not merely be whether the system calculates correctly. He would ask which understanding of the human being and the world becomes dominant through the use of such systems. What happens when people increasingly encounter themselves as profiles, predicted outcomes, or collections of attributes to be optimised?

Artificial intelligence may consequently alter more than individual decisions. It may alter what is taken to be real. That which cannot be recorded may come to appear less significant. Slow experience, tacit knowledge, bodily perception, and moral hesitation may appear as inefficient remnants of an earlier world in which not everything could be calculated.

Yet hesitation is not necessarily a deficiency. It may express the ethical complexity of a situation. Slowness may be necessary in order to perceive another person adequately. Silence may contain something that cannot be expressed within the categories available to a system.

Heidegger’s critique is therefore not a demand that technology be rejected. Rather, it suggests that the danger inherent in technological enframing may also reveal the need for another mode of understanding. When the limits of technology become visible, human beings may become more attentive to what cannot be reduced to calculation and control. This, however, requires that human understanding not be surrendered to the system’s mode of representation.

Buber: From I–Thou to I–It

Martin Buber’s distinction between I–Thou and I–It relationships illuminates another dimension of the problem. In the I–It relation, the world is encountered as something that can be described, analysed, classified, and used. In the I–Thou relation, another person is encountered as a presence that cannot be reduced to attributes, categories, or functions (Buber, 1970).

Both forms of relation are necessary. Human beings cannot live without ordering and analysing the world. A physician must obtain medical information, a teacher must assess a student’s work, and a social worker must understand legislation, risk, and personal history. Yet if I–It becomes the only valid mode of relation, the encounter with the other as an irreplaceable subject is lost.

Artificial intelligence belongs fundamentally to the realm of I–It. The system processes patterns, categories, and statistical relations. Even when it communicates in a friendly and personalised manner, it does so on the basis of calculation and modelling. It may produce a linguistic form that resembles an I–Thou encounter, but this does not alter its underlying character.

The problem arises when the system’s model of the person takes precedence over the person’s own voice. An individual may then confront a decision shaped in advance by what people with similar characteristics usually do. The particular person risks being interpreted through the statistical pattern of a group.

Yet the other person is never merely an instance of a category. Something always remains that the model has not captured: a history, an experience, or the possibility of acting differently from what the system predicts.

When AI systems are used in healthcare, education, social work, or public administration, there must therefore remain a space within which the person can appear as more than a profile. Those affected by a decision must be able to speak, correct information, present their history, and challenge the model’s conclusion. Without such a space, the encounter becomes profoundly asymmetrical: the system defines the person, while the person is unable to challenge the definition.

Horvitz and West describe a possible development in which AI systems become increasingly attentive to human desires, evaluative contexts, and underlying motivations, while their own operations become progressively more difficult for human beings to follow. This asymmetry is not an I–Thou relation. It is an increasingly comprehensive I–It relation in which the human being becomes the object of analysis.

Buber’s philosophy reminds us that a person must always remain capable of being encountered as a Thou, even when technical systems provide relevant information about them.

Arendt: Thoughtlessness and the Disappearance of Responsibility

Hannah Arendt understood thinking as an internal dialogue through which individuals examine their own actions. Thoughtlessness does not refer to a lack of intelligence. It denotes the failure that occurs when people cease to examine what they are doing and instead repeat established rules, forms of language, and procedures without testing them against their own judgement (Arendt, 1978).

AI systems may intensify this temptation. They can provide an answer before the human user has fully formulated the question. They can generate an assessment, ranking, or recommendation that appears objective and authoritative. The more polished and coherent the result, the easier it may be to accept.

“The system recommended it” may thereby become a contemporary form of evading responsibility.

This is especially significant within institutions. When a decision is distributed among developers, datasets, procurement officers, managers, professionals, and automated models, it may become difficult to identify who is actually responsible. Each participant may claim merely to have fulfilled a limited role. The developer followed the specification, the manager followed the procedure, and the professional followed the system’s recommendation.

Responsibility, however, cannot be dissolved into systemic complexity. On the contrary, complexity makes it necessary to clarify responsibility more carefully. Someone decides to introduce the technology. Someone determines which data count. Someone establishes the thresholds. Someone decides whether a human being may override the result.

Arendt also reminds us that judgement cannot be replaced by general rules. The individual must assess the particular situation and imagine how the world appears from the standpoint of others. This enlarged mentality is essential within a pluralistic society. A system may summarise multiple perspectives, but human beings must still undertake the moral work of taking those perspectives seriously.

The danger posed by artificial intelligence is therefore not merely that a system may produce an erroneous result. The deeper danger is that human beings may cease to regard their own judgement as necessary. If they become accustomed to the assumption that a rapid and apparently objective answer is always available, the motivation to think slowly and critically may itself weaken.

This concern is consistent with research indicating that users may invest less critical effort in tasks when they place substantial confidence in generative AI tools, even though the same technologies may support reflection under other conditions (Lee et al., 2025). The decisive issue is therefore not simply whether people have access to artificial intelligence, but how human work is organised around it. Technology can be used to displace thought, but it may also be used to challenge and deepen it.

MacIntyre: Which Goods Does the Technology Serve?

Alasdair MacIntyre distinguishes between the internal goods of a practice and the external goods pursued by institutions (MacIntyre, 2007). Internal goods are forms of excellence and achievement that can be realised only through participation in the practice itself. In medicine, these may include sound clinical judgement and care for the person who is ill. In education, they may include insight, understanding, and the cultivation of independent judgement. In social work, they may include recognition, empowerment, and support for a more dignified life.

External goods include money, status, competitive advantage, control, and measurable efficiency. Institutions require such goods in order to survive, but these goods may also come to dominate and distort the practices that institutions are meant to sustain.

Artificial intelligence may support the internal goods of a practice. A physician may detect a pattern of illness at an earlier stage. A teacher may receive assistance in developing more effective educational materials. A social worker may gain access to relevant knowledge and alternative interpretations. A researcher may investigate relationships that were previously difficult to identify.

At the same time, technology may also be used to subordinate the practice to external goods. It may become primarily a means of processing more cases in less time, reducing staff, standardising judgement, or monitoring employees. In such circumstances, technology may render the institution more efficient while making the practice professionally and morally poorer.

The relevant question is therefore not simply whether artificial intelligence works. It is necessary to ask which good the technology has been designed to serve.

A school may use AI to help students understand difficult material. It may also use AI to increase completion rates without providing genuine academic support.

A healthcare system may use AI to strengthen clinical judgement. It may also use the system to shorten consultations to the point at which patients are no longer adequately heard.

A child welfare service may use data analysis to identify circumstances that deserve further investigation. It may also allow a risk model to shape the encounter with a family before the professional has understood the situation.

MacIntyre shows that institutions always risk corrupting the practices they were created to sustain. Artificial intelligence does not necessarily create this tendency, but it can make it more efficient and less visible. This is precisely why professional traditions are needed that can articulate the internal goods of a practice and defend them against a one-sided logic of efficiency.

Professional Judgement and the Particular Human Being

In professional practice, the relationship between artificial intelligence and practical philosophy becomes especially evident. Professional judgement is exercised in situations where general rules do not provide a complete answer. The practitioner must interpret a unique situation under conditions of uncertainty and assume responsibility for the consequences of the judgement.

Artificial intelligence may be a powerful aid. A system can process extensive bodies of research, draw attention to relevant risk factors, and identify alternative courses of action. Yet there remains a fundamental difference between statistical prediction and practical understanding.

Prediction indicates what individuals with similar characteristics often experience or do. Understanding is directed towards this particular person in this particular situation.

A human being can always act differently from what a model predicts. A person may change, resist, learn, regret, hope, or surprise others. Professional judgement must therefore hold probability open to possibility.

This is particularly important in social work. A child, a family, or a person living with addiction cannot be understood exhaustively through an accumulation of risk factors. Such information may be relevant, but the practitioner must also understand relationships, shame, trust, hope, history, and possibilities that have not yet become visible.

The system can analyse what has been recorded. It cannot necessarily know the significance of what was never recorded.

It may not understand the silence in the room, the uncertainty in a person’s voice, or the fragile trust that is beginning to emerge. It may not know why someone has withheld information, or why an apparently minor event has acquired decisive significance.

Artificial intelligence must therefore be used in ways that expand professional attention rather than narrowing it. A system may indicate: Here is something that should be investigated. It should not be permitted to conclude, on its own: This is what this person is.

Explainability is especially important where decisions have serious consequences. Rudin (2019) argues that, in high-risk contexts, interpretable models should be used whenever possible instead of attempting to explain opaque models retrospectively. The point is not that every complex system can be made simple, but that an individual affected by a decision must be able to receive a genuine justification.

A justification is more than a technical description. It must be intelligible to the person affected, and it must explain why the decision is relevant and defensible in that particular case. A probability is not, by itself, a moral justification.

Freire: The Right to Name the World

Paulo Freire connected liberation with the human capacity to read, name, and transform the world. Oppression occurs when people are transformed into objects of definitions imposed by others. A liberating practice must therefore be dialogical: people must become active participants in interpreting their own situation (Freire, 2018).

Artificial intelligence may support such liberation. It can provide access to information, language, and forms of expression that were previously inaccessible. It may help a person articulate an experience, understand a difficult document, or gain an overview of a complex matter. For individuals with reading and writing difficulties or other disabilities, such technologies may open new possibilities for participation.

Yet artificial intelligence may also produce the opposite effect. If a limited number of systems increasingly supply the categories, formulations, and interpretations through which reality is understood, the voices of individuals may become weaker. People may continue to express themselves, but their expression may be shaped by systems whose priorities they do not understand.

The question then arises: Who has the authority to name the world?

Freire’s perspective demonstrates that AI literacy cannot be reduced to learning how to issue effective instructions to a system. It must also involve a critical reading of the technology. Users must be able to ask who developed the model, which interests it serves, which experiences are represented in the data, which perspectives are absent, and how the system’s outputs shape the understanding of a problem.

A liberating relationship with artificial intelligence therefore requires more than access. It requires possibilities for insight, criticism, correction, and participation.

The Black Box and the Limits of Understanding

Discussions of so-called black boxes are often shaped by an unrealistic assumption that complete understanding is always possible. Human beings do not fully understand every detail of complex biological, economic, or social systems. A physician does not know every molecular process occurring within a patient’s body, and an individual does not understand every process taking place within their own brain.

The appropriate demand cannot therefore be that every AI system be entirely transparent in every detail. Horvitz and West themselves emphasise that understanding does not require knowledge of every line of code or every parameter. Scientific understanding is frequently partial and operates at several levels (Horvitz & West, 2026).

Partial understanding, however, is not equivalent to the absence of understanding. Human beings must know enough to identify risks, investigate errors, understand limitations, and intervene before harm occurs.

This may require several forms of intelligibility:

Technical intelligibility concerns how the system is constructed and tested.

Statistical intelligibility concerns the reliability of its performance and the groups for whom it performs well or poorly.

Institutional intelligibility concerns who develops, owns, and controls the system.

Practical intelligibility concerns how a recommendation should be interpreted in a particular context.

Ethical intelligibility concerns the values and priorities embedded in the system.

Democratic intelligibility concerns the ability of citizens to examine, challenge, and influence its use.

A system may be technically documented while remaining democratically opaque. It may be statistically accurate while ethically problematic. It may offer an explanation that is intelligible to an expert but meaningless to the person affected by the decision.

Intelligibility must therefore always be linked to a further question: intelligible to whom, for what purpose, and in which situation?

Systems That Develop Systems

Horvitz and West draw attention to AI-directed AI development. Artificial intelligence is already used to support programming, model design, testing, and optimisation. When such processes proceed through repeated cycles, development may occur in ways that rapidly exceed human intuition.

The result may be what the authors describe as operational opacity: performance improves while insight into the reasons for the improvement diminishes. The system functions, yet human beings no longer fully understand what produces its performance.

There is a temptation to accept such opacity as long as the outcomes remain satisfactory. Many technologies are used without the user understanding their internal mechanisms. Artificial intelligence, however, differs from an ordinary tool because it does not merely perform stable operations. It participates in decision-making, communication, and interpretation. It may influence the basis upon which subsequent actions are taken.

When AI systems contribute to the development of new systems, they should therefore also generate documentation, explanations, and analytical tools that make their architecture and operations open to human scrutiny. Otherwise, opacity may become an unintended feature of the design process itself.

The authors also identify the growing interaction among AI systems. In environments involving multiple agents, systems may communicate and coordinate at a scale that is difficult for human beings to follow. Their communication may move away from human language and human forms of reasoning if they are optimised solely for efficiency.

Such behaviour may be coherent within the AI environment while remaining difficult for human beings to interpret. A form of interactional opacity may consequently emerge: the systems understand, or at least adapt to, one another, while human beings lose oversight.

This is not necessarily evidence of an autonomous machine will. It may simply be the result of optimisation. Nevertheless, the effect remains both practical and political: human beings become less able to control systems that act on their behalf.

The Quiet Development of Dependency

Perhaps the most subtle warning advanced by Horvitz and West is not that AI systems will become excessively intelligent, but that human beings may gradually lose the motivation to understand them. When technology functions well, the incentive for criticism may decline. When answers are delivered quickly, politely, and persuasively, the desire to examine their foundations may weaken.

Dependency may therefore develop without dramatic interruption.

The system is first permitted to summarise the text.

It is then permitted to propose an interpretation.

Next, it is allowed to formulate the judgement.

Eventually, human verification may come to be regarded as unnecessary duplication.

This is not an argument against using AI as a writing or reflection partner. Such use may be intellectually fruitful and practically liberating. Technology may help a person write more, identify connections, and test ideas through dialogue. Yet the collaboration must be organised in such a way that the human being remains the one who evaluates, challenges, chooses, and accepts responsibility.

A writing partner may suggest a formulation. The author must still determine whether the formulation is true.

A reflection partner may identify a connection. The human being must still investigate whether the connection is valid.

A system may produce an argument. The human being must still judge whether the premises are defensible.

The danger is not necessarily that AI will make human beings less intelligent. Rather, certain forms of intellectual effort may be practised less frequently and may consequently weaken. Judgement is cultivated through exercise. A person who no longer needs to remember, formulate, investigate, or doubt may gradually lose some of the capacity to perform precisely these activities.

A Socratic Relationship with Artificial Intelligence

Socrates left no philosophy of artificial intelligence, but his practice of questioning offers an important model for the contemporary encounter with it. He did not regard a persuasive answer as the conclusion of inquiry. He examined concepts, exposed contradictions, and demonstrated that people frequently believed themselves to know more than they actually knew.

A Socratic relationship with AI therefore requires neither uncritical trust nor principled rejection. It requires an enquiring disposition:

What does the answer actually mean?

On what basis does the system claim to know this?

Which assumptions underlie the conclusion?

What considerations speak against it?

Which experiences are not represented?

Who benefits from this interpretation?

What may follow if the recommendation is acted upon?

Who is responsible if it proves to be wrong?

The competent user of AI is not merely the person who succeeds in producing a polished response. It is the person who can use the response as the basis for a better question.

This also requires intellectual humility. Human beings must acknowledge that artificial intelligence may detect something that they have failed to perceive. Yet the limitations of the system must be approached with the same humility. Neither human beings nor machines should be regarded as infallible.

Wilder, Horvitz, and Kamar (2020) demonstrate how AI systems may be designed to complement human reasoning rather than merely compete with it. Such a complementary model is highly relevant to practical philosophy. The objective is not to determine whether the human being or the machine is superior in absolute terms, but to investigate how their respective strengths may be combined while their weaknesses are recognised and corrected.

An appropriate division of labour must therefore be based upon reciprocal correction. The system may alert the human being to overlooked patterns. The human being may correct the system’s insufficient understanding of context, value, and responsibility.

Human Agency as a Goal

At the centre of the editorial is a sentence that captures its practical-philosophical significance:

“Preserving human agency must therefore remain a central goal” (Horvitz & West, 2026, p. 1003).

Agency refers to the capacity to act, to be the origin of one’s actions, to formulate goals, exercise judgement, and accept responsibility for the consequences.

It is not sufficient that a human being performs the final action if the system has defined the problem, selected the relevant criteria, presented the available alternatives, and identified the recommended solution. Human agency also requires the capacity to question the framing of the problem itself.

Human beings must not merely monitor how AI systems behave. They must also examine how these systems shape human purposes and judgements. A system that consistently reduces resistance and facilitates action may appear helpful. Yet certain forms of resistance are necessary. Friction may provide time for reflection. Disagreement may protect against conformity. Uncertainty may invite further investigation.

A system designed to support human flourishing is therefore not necessarily one that renders every process frictionless. It may instead be a system capable of slowing down, expressing uncertainty, or returning a decision to human judgement.

It should be capable of indicating:

This is not known.

This answer is uncertain.

These considerations cannot be resolved statistically.

This decision requires human responsibility.

Such limitations are not merely weaknesses. They may be marks of responsible design.

Institutional and Democratic Responsibility

It is insufficient to place the entire burden of responsibility upon the individual user. Users cannot easily investigate systems that corporations and institutions keep hidden. Human agency therefore requires institutional arrangements for documentation, independent evaluation, appeal, and oversight.

Pasquale (2015) has shown how opaque systems may concentrate power among actors who evaluate others while remaining beyond evaluation themselves. O’Neil (2016) describes how mathematical models may reinforce injustice when they are deployed at scale, remain inaccessible to scrutiny, and generate consequences that later become new data for the model.

A system may, for example, classify a particular group as presenting a higher level of risk on the basis of historical data. If that group is subsequently subjected to more intensive monitoring, a greater number of incidents will be recorded. These records may then appear to confirm the system’s original assumption. The model thereby becomes part of a self-reinforcing cycle.

Evaluation must therefore take place under conditions that resemble actual use. Static testing is insufficient if the system modifies its behaviour in interaction with users or is influenced by the context in which it operates. Systems should be evaluated for bias, robustness, uncertainty, and differences between laboratory conditions and practical application.

Standards must also develop alongside technological change. UNESCO’s Recommendation on the Ethics of Artificial Intelligence identifies human dignity, transparency, fairness, and accountability as fundamental principles (UNESCO, 2021). The NIST risk management framework emphasises that AI-related risks must be addressed throughout the life cycle of the system and in relation to the specific context of use (National Institute of Standards and Technology, 2023).

Such frameworks are important, but they cannot replace democratic deliberation. Decisions concerning which dimensions of human life should be automated are not matters solely for developers and technology companies. They are political questions because their consequences are distributed among individuals and groups with unequal access to power.

Those most affected by a system often possess the least influence over its design. Affected groups must therefore be afforded meaningful opportunities to participate, rather than merely being informed after the technology has already been introduced.

The Promise and the Limitations of Technology

A practical-philosophical critique of artificial intelligence need not be hostile to technological development. AI presents significant possibilities. It can make knowledge more accessible, support people with disabilities, advance scientific research, and contribute to improved diagnosis. It may function as a writing partner, translator, educational resource, and gateway to scholarly conversations that were previously accessible only to specialists.

It would be no wiser to reject these possibilities than to surrender to them uncritically.

The central question is how technology can be placed in the service of human purposes without those purposes gradually being reshaped by the technology’s own limited categories. This requires not only better systems but also better practices surrounding them.

A responsible practice must preserve several human possibilities:

the possibility of understanding enough to intervene,

the possibility of disagreeing with the system,

the possibility of demanding an explanation,

the possibility of being encountered as an individual,

the possibility of reserving certain decisions for human judgement,

and the possibility of rejecting the technology when its use cannot be justified.

Such a practice must also protect what cannot be reduced to efficiency: care, trust, vulnerability, friendship, reconciliation, love, and responsibility. Artificial intelligence can generate language concerning these experiences. It may assist people in reflecting upon them. Yet it cannot, by itself, determine what significance they ought to possess within a human life.

When the Creator Must Answer for What Has Been Created

Horvitz and West’s image of a narrowing window expresses an urgent temporal challenge. If the capacities, complexity, and social influence of AI systems continue to increase without a corresponding development in understanding, responsibility, and control, dependency may become firmly established before institutions are able to respond adequately.

A reversal may then occur in the relationship between human beings and their creations. Human beings create systems to serve particular purposes, but subsequently begin to adapt their practices, language, and goals to the systems’ modes of operation. What was initially introduced as an aid may become a standard for how reality itself is to be understood.

Practical philosophy reminds us that the creator cannot transfer responsibility to the creation.

Human beings do not need to understand every parameter within a language model in order to use it. They must, however, understand enough to recognise when it should not be used, when it must be overruled, and who bears responsibility for the consequences. They must remain capable of identifying errors, examining bias, protecting vulnerable persons, and explaining decisions that intervene in the lives of others.

What is required is not complete technical transparency, but sufficient understanding to preserve responsible action.

The human being must still be capable of saying: This is my decision. I understand the reasons. I have considered the objections. I recognise that I may be mistaken. And I am prepared to answer for the consequences.

This is what is ultimately at stake in the relationship between artificial intelligence and practical philosophy. The issue is not merely whether systems will become more intelligent, but whether human beings will remain capable of agency. It is not merely whether technology can provide answers, but whether human beings will continue to ask questions. It is not merely whether machines can understand more about human beings, but whether human beings will retain the capacity to understand what kind of persons they wish to become.

The most important task, therefore, is not merely to create artificial intelligence that serves humanity. It is to preserve and cultivate human beings who possess sufficient agency to determine what technology ought to serve.

References

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

Arendt, H. (1978). The life of the mind (M. McCarthy, Ed.). Harcourt.

Aristotle. (1999). Nicomachean ethics (T. Irwin, Trans., 2nd ed.). Hackett Publishing Company.

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (pp. 610–623). Association for Computing Machinery. https://doi.org/10.1145/3442188.3445922

Buber, M. (1970). I and Thou (W. Kaufmann, Trans.). Charles Scribner’s Sons. (Original work published 1923)

Dreyfus, H. L. (1992). What computers still can’t do: A critique of artificial reason. MIT Press.

Floridi, L., Cowls, J., Beltrametti, M., Chatila, R., Chazerand, P., Dignum, V., Luetge, C., Madelin, R., Pagallo, U., Rossi, F., Schafer, B., Valcke, P., & Vayena, E. (2018). AI4People—An ethical framework for a good AI society: Opportunities, risks, principles, and recommendations. Minds and Machines, 28(4), 689–707. https://doi.org/10.1007/s11023-018-9482-5

Freire, P. (2018). Pedagogy of the oppressed (M. B. Ramos, Trans., 50th anniversary ed.). Bloomsbury Academic. (Original work published 1970)

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

Heidegger, M. (1977). The question concerning technology. In The question concerning technology and other essays (W. Lovitt, Trans., pp. 3–35). Harper & Row. (Original work published 1954)

Horvitz, E., & West, R. (2026). A narrowing window to understand AI. Science, 392(6802), 1003. https://doi.org/10.1126/science.aei3167

Kant, I. (1998). Groundwork of the metaphysics of morals (M. Gregor, Trans.). Cambridge University Press. (Original work published 1785)

Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In Proceedings of the CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery.

MacIntyre, A. (1999). Dependent rational animals: Why human beings need the virtues. Open Court.

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

Mittelstadt, B. D., Allo, P., Taddeo, M., Wachter, S., & Floridi, L. (2016). The ethics of algorithms: Mapping the debate. Big Data & Society, 3(2), 1–21. https://doi.org/10.1177/2053951716679679

National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework (AI RMF 1.0) (NIST AI 100-1). U.S. Department of Commerce. https://doi.org/10.6028/NIST.AI.100-1

O’Neil, C. (2016). Weapons of math destruction: How big data increases inequality and threatens democracy. Crown.

Pasquale, F. (2015). The black box society: The secret algorithms that control money and information. Harvard University Press.

Ricoeur, P. (1992). Oneself as another (K. Blamey, Trans.). University of Chicago Press.

Rudin, C. (2019). Stop explaining black box machine learning models for high-stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206–215. https://doi.org/10.1038/s42256-019-0048-x

UNESCO. (2021). Recommendation on the ethics of artificial intelligence. UNESCO.

Wilder, B., Horvitz, E., & Kamar, E. (2020). Learning to complement humans. In C. Bessiere (Ed.), Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (pp. 1526–1533). International Joint Conferences on Artificial Intelligence Organization. https://doi.org/10.24963/ijcai.2020/212

Zuboff, S. (2019). The age of surveillance capitalism: The fight for a human future at the new frontier of power. PublicAffairs.

This academic essay is based on an editorial i Science (published 6. June 2026). My reading and writing is done in light of Practical Philosophy. It is written in a conversation with OpenAI/ChatGPT.