From Homo videns to artificial intelligence: the value of human judgment

Introduction

Sometimes I am surprised by how easily we take for granted one of our most extraordinary capacities: our capacity for wonder. We can encounter something we do not fully understand and feel an almost spontaneous need to find out what is happening. An explanation fails to persuade us; something does not quite fit; and we begin asking questions that no one has instructed us to answer.

What fascinates me is that we do not need to be prompted to do so. Sometimes that curiosity leads us to discover that the real problem was not the one before us, but one we had not even noticed. Then something revealing happens: an idea clicks into place and everything begins to make sense. This is what we are like as human beings. Smile for a moment.

Aristotle already recognized wonder as the origin of philosophy (Metaphysics, I, 2, 982b12–21). When we marvel at what we do not know, we become aware of our ignorance and feel compelled to investigate. I am particularly interested in this willingness to continue searching, because knowledge often begins precisely when we recognize that our first explanation was not enough.

I have experienced something similar in my work with businesses. Sometimes a consultation involved preparing a document or reviewing a specific obligation. Yet when I examined the background, I might discover that the request arose from a deeper difficulty whose consequences had not yet been considered. At other times, of course, the initial solution was sufficient. I learned that understanding a problem also requires recognizing when there is no need to complicate it.

What moved me to investigate beyond what I had been asked to do? I believe that, to a considerable extent, it was genuine concern for the people who placed their trust in me. I wanted to protect them and spare them difficulties they might not yet have recognized. I did not always have the answers, nor was I necessarily right. But precisely because I cared about what might happen to them, I felt responsible for examining what I did not yet understand.

Over time I began wondering how much of that work artificial intelligence could perform. Today these tools can analyze situations at extraordinary speed, connect concepts, and even identify problems a professional might overlook. It would be difficult to deny the value of such progress. Nevertheless, these systems can also make mistakes, and their answers do not always make those mistakes easy to recognize.

This led me to a question I consider particularly important: what happens to our own judgment when an explanation is so convincing that we no longer feel the need to examine it?

In business, the question also has an economic dimension. Artificial intelligence can significantly reduce the cost of producing an answer without necessarily reducing, to the same degree, the cost of being wrong when we put it into practice. There may even be less obvious consequences, such as the gradual loss of certain professional capabilities when we stop exercising them.

The efficiency of a tool should therefore be assessed not only by the time it saves, but also by the consequences of using it, the risks it introduces, and the capabilities we will need to retain in order to evaluate its performance.

These considerations lead me to the central question of this essay:

If artificial intelligence can offer convincing answers, does the value of human judgment lie in finding the right answer, or in discerning whether we should act on it?

I do not intend to prove that human beings always judge better than machines, or that every automated answer should be distrusted. I want to examine what happens between receiving an answer and turning it into a decision, and why that distinction matters both philosophically and economically.

To do so, I will proceed in two parts. In the first, I examine the relationship between information, understanding, and practical judgment, drawing on Sartori, Aristotle, and the biblical image of a listening heart. In the second, I bring these reflections into the business context to consider the efficiency of automation, the economic value of decisions, and the possibility that delegating certain activities may weaken the capabilities needed to supervise them.

My purpose is not to set human intelligence against artificial intelligence, but to understand what we ought to preserve and what criteria we need in deciding when to delegate, when to supervise, and when to intervene.

Part One. Understanding and discernment: the foundations of human judgment

I. From Homo videns to artificial intelligence: receiving information is not the same as understanding

Some years ago, while I was at university, I wrote an essay about television's effects on our capacity to reason. One of the books that interested me most was Giovanni Sartori's Homo videns (1997), in which he warned of the risks of a culture dominated by images and their possible consequences for abstract thought.

Revisiting that work, I find a concern that still seems relevant: we can grow accustomed to receiving information without pausing to examine it. When we stop questioning what we see or hear, we risk accepting explanations without sufficiently considering their foundations.

Artificial intelligence introduces a different difficulty. We no longer receive only images or information about events; we can also obtain elaborate explanations that interpret what is happening. To understand what changes with this technology, it is useful to return first to Sartori's concern.

1. Abstraction and the limits of the visible

In Homo videns, Giovanni Sartori revisits the conception of the human being as a symbolic animal associated with Ernst Cassirer's thought. This idea helps us understand that our relationship with reality is not limited to what we perceive through the senses. We also interpret the world through symbols and concepts, and language allows us to think about realities whose meaning extends beyond what we can observe directly.

We can visually represent a mobile phone or computer and easily recognize the objects in question. But what image alone could convey the meaning of justice, freedom, or legitimacy? We can show situations related to those concepts, but understanding what they mean requires more than observing them: we must think about what defines them and distinguishes them from other realities.

This is the distinction that helps explain Sartori's concern. His criticism is not that images are harmful, but that their predominance may have consequences when it displaces other forms of thought. If we become accustomed to knowing reality chiefly through what is visible, we risk giving less room to the abstraction needed to understand what cannot immediately be seen.

It also helps explain why being informed does not necessarily mean understanding. We can know that an event occurred without understanding its causes or implications. Information brings us closer to what happened, but we have to interpret it to understand its significance.

Consider someone who witnesses the signing of a contract. They can identify the participants and describe what happens. But understanding the agreement's implications requires knowing the obligations it creates and the consequences it may produce. The scene is visible; much of what is at stake depends on interpreting concepts and circumstances that observation alone cannot reveal.

Artificial intelligence poses a different issue from the one Sartori explored. Unlike television, AI tools do not merely show us a reality: they can also explain it. That can help us understand complex matters, but it also raises a problem worth examining. A well-constructed answer may seem sufficient even when we have not examined its foundations or when the original question has omitted relevant aspects of the problem.

The issue is not whether we should distrust AI answers merely because they come from a technological tool. Nor should we assume an elaborate explanation is necessarily correct. What matters is establishing which problem it solves, what evidence it rests on, and how closely it responds to the specific circumstances.

Imagine a company that needs to amend its bylaws to bring in a new partner. An AI tool could prepare a legally appropriate document tailored to the request. Yet if the partners have not agreed on how to resolve possible disagreements over management, the document might leave a significant problem for the business unaddressed.

The tool might also identify that difficulty and recommend considering it before drafting the document. The point is not to attribute to people an exclusive capacity to identify problems that machines could never recognize. It is to note that an answer's usefulness depends in part on how the problem was framed and whether we have considered the circumstances needed to evaluate it.

Abstraction remains important, then, although the challenge is no longer merely to move beyond what is visible. We must also examine the interpretations we receive, understand their foundations, and recognize their limits. AI can help us do this, but the quality of its answer does not, by itself, prove that we have understood what we need to resolve.

2. Experience and our relationship with reality

Understanding depends not only on the concepts we use or the information available to us. Our experience shapes how we perceive a situation and recognize what deserves attention.

Henri Bergson addresses this issue in Matter and Memory (1896), where he examines the relationship between perception, memory, and action. His reflection invites us to consider that our relationship with the present is not simply a matter of receiving information: what we have lived through also enters into how we interpret what is happening.

Experience, however, is no guarantee of sound judgment. A situation may remind us of something we previously experienced and help us recognize a genuine problem, but it may equally cause us to overlook important differences. Having confronted similar difficulties can help us understand a new one, but does not exempt us from examining its particularities.

I do not aim here to develop all the implications of Bergson's philosophy. I wish to preserve one distinction: understanding a situation involves more than possessing background information, because it also involves how we interpret that information through experience. The relationship deserves further study, especially when we ask what value we attach to professional experience in comparison with the answers AI systems can provide.

Looking back at my earlier essay on Homo videns, I also find that I linked the loss of abstraction with our ability to distinguish the true from the false and the right from the wrong. Today I think that relationship needs to be made more precise.

Abstraction enables us to understand concepts that cannot be reduced to what is immediately visible, and reasoning helps us evaluate whether an explanation is well founded. Yet neither capacity, by itself, settles the question of how we should act. We may understand a problem, recognize its possible consequences, and even have sufficient information without knowing which alternative to choose.

Think again of the company seeking to introduce a new partner. Understanding the legal implications of the amendment is necessary, but might not be enough to determine whether to proceed immediately or first address disagreements about management. To decide, we must consider the specific circumstances and assess which alternative best responds to the situation.

The decision itself can also prompt us to revise what we thought we understood. Asking what each alternative would entail may reveal missing information or that we have given undue weight to one aspect of the problem. Understanding and deciding are connected, but they are not identical: in deliberating about what to do, we may discover that we need to understand the situation better.

Sartori's reflection raises questions about the place of thought in an increasingly technology-mediated culture; it does not prove that using AI necessarily weakens our capacity for abstraction. That possibility must be investigated, not taken for granted. Similarly, Bergson's reflection highlights the role of experience but cannot, by itself, establish that artificial systems are incapable of contributing to adequate understanding.

What interests me is the difficulty that emerges when we move from understanding a situation to deciding what to do about it. An answer may be well constructed and useful for completing a task, yet we still have to determine whether the action it enables is appropriate in the circumstances.

This takes us beyond information, abstraction, and experience. It asks us to examine how we deliberate about our actions, which ends we pursue, and what it means to judge well in a particular situation. Aristotle's distinction between knowing how to produce something and practical wisdom is particularly useful here.

II. From understanding to discernment: practical wisdom and the listening heart

We can know how to perform an activity without having determined whether we ought to perform it. That difference takes us from the problem of understanding to another question: how do we judge what we should do in a particular situation?

1. Aristotle: knowing how to make something is not the same as knowing how to act

In Book VI of the Nicomachean Ethics, Aristotle distinguishes technē, associated with the knowledge of how to produce, from phronēsis, usually translated as practical wisdom or prudence. The former concerns knowledge that enables us to make something; the latter concerns the capacity to deliberate well about what ought to be done in the particular circumstances of human life.

This distinction matters because making something correctly and acting well are not the same. A professional may master the procedures needed to draft a contract or execute a financial transaction, but that knowledge alone does not determine whether proceeding is advisable or whether the result serves the good being pursued.

Return to the company seeking to bring in a partner. Knowing how to amend its bylaws enables us to prepare the relevant document. But when we examine the reasons behind the request, we may discover unresolved disagreements over management. That fact changes how the problem should be approached.

Perhaps the right course is to prepare the requested document; perhaps the partners should first clarify their differences. The decision depends on the circumstances and on what the company is trying to resolve. In deliberating over what to do, we may discover aspects we had not considered and thereby revise our original understanding.

This is where I find the value of Aristotelian practical wisdom. Knowing the means to obtain a result is not enough: we must also judge which action is appropriate to the circumstances and toward which good it should be directed.

The distinction helps clarify the problem posed by artificial intelligence. A tool can help us find a technically correct solution and even identify difficulties we have overlooked. But technical correctness does not, by itself, settle every question involved in a decision. We still need to examine what we hope to achieve, which circumstances matter, and what acting in one way rather than another would entail.

This does not mean that people always judge better than technology. We too can make mistakes, overlook important facts, or defend harmful decisions. The point is to understand what practical judgment requires and what responsibility we assume when we decide to act.

2. Love as a disposition to pay attention

To explore the question further, I want to consider something that is not directly about what we know, but about our disposition toward what we are trying to understand: the importance we attach to a situation and to those who may be affected by our decisions.

When we genuinely care about someone, the way we attend to their circumstances may change. Perhaps we notice something we previously overlooked, or ask whether the solution that seemed sufficient really meets their needs.

I think love can play a part in this disposition. Not because it gives us correct answers or protects us from mistakes, but because it can move us to pay attention to matters we might otherwise have ignored.

Consider the difference between completing an assigned task and trying to ensure that the result is genuinely useful to the person receiving it. In the latter case, checking that we have fulfilled the request is not enough. We also need to ask whether our intervention responds to that person's circumstances.

Yet caring about someone does not guarantee that we know what is best for them. We may misinterpret their needs, project our expectations onto them, or insist on what we consider a beneficial solution without having understood their situation well enough.

Love is therefore no substitute for judgment. It may motivate us to understand better, but we must critically examine what we believe we have understood. Otherwise even genuinely good intentions can produce consequences we did not want.

This relationship between care and judgment brings to mind an image I find especially meaningful: Solomon's request for a heart capable of listening.

3. Solomon and the listening heart

In the First Book of Kings, Solomon asks God for a listening heart so that he may govern his people and discern between good and evil (1 Kings 3:9). What strikes me about the request is that Solomon does not simply ask to know more; he asks for the capacity to judge in order to fulfill a responsibility that affects other people.

The image of a listening heart expresses a recognition of the limits of our own judgment. Before deciding, we must be willing to attend to a reality that we may not yet know completely and that may force us to revise what we initially believed.

Benedict XVI returned to this request in his address to the German Bundestag on 22 September 2011. Reflecting on the foundations of law, he argued for a reason open to recognizing what is just, beyond criteria such as success or immediate usefulness. Justice requires asking not only what we can do, but also what we ought to do.

Pope Francis, for his part, in his 2022 message Listening with the Ear of the Heart, emphasized listening as a fundamental dimension of our relationship with others. Listening does not consist merely in receiving words, but in paying attention to another person without assuming that we already understand what they need to express.

I find the relationship between listening, understanding, and discernment valuable. To listen is to acknowledge that our interpretation may be incomplete and that attending to someone else's circumstances may lead us to reconsider a decision.

Still, listening is no guarantee of being right. We may pay close attention and continue to misinterpret a situation; we may genuinely care for someone and hurt them because we have insufficiently considered the consequences of our intervention.

This is why I believe love and practical wisdom should be understood together. Love may move us to pay attention; prudence requires us to deliberate about what to do in light of what we have understood. Neither removes the possibility of error, but together they invite us to see judgment as more than the application of a procedure or the pursuit of an outcome.

The biblical image of a listening heart does not establish that human beings are infallible or that AI cannot contribute to a decision. Its importance here is different: it reminds us that discernment involves recognizing what is at stake for others and taking responsibility for our actions.

Recognizing that responsibility also means asking how to translate it into concrete decisions. In business, where time, resources, and risks have economic consequences, the question acquires another dimension: how do we determine when to delegate a task to artificial intelligence and when to intervene ourselves?

This question leads to the second part of the essay, where I examine the relationship between human judgment, automation, and the economic value of decisions.

4. Caring also means knowing when to stop

There is a difficulty I find particularly interesting: sometimes care requires us to investigate beyond what has been requested; at other times it consists precisely in recognizing that doing so is unnecessary.

In my work with businesses I have learned that going deeper does not always mean helping better. Examining background matters that initially seemed secondary may reveal a major difficulty. But we can also spend time and resources searching for problems that do not justify further intervention, or end up proposing more documents, procedures, and technology than a person actually needs.

I think caring requires us to recognize that difference. If our concern is genuine, we should care not only about solving the problem, but also about avoiding needless complications and considering the resources our intervention demands.

Sometimes this means acknowledging that the decision belongs to the person concerned. We can help them understand their alternatives and the foreseeable consequences of each without replacing their judgment or assuming that we know better what is right for them.

Prudence here has a concrete meaning. It does not consist in always intervening or investigating indefinitely, but in determining what is appropriate in each case. Sometimes we will need to go deeper; at others a simple solution will suffice. The difference lies not in how much we do, but in whether what we do actually responds to the problem.

This also changes how I understand the relationship between human judgment and AI. These tools can extend our analytical capacity, help us identify alternatives, and reduce the effort needed to perform certain tasks. If they enable us to obtain a reliable outcome using fewer resources, employing them can be perfectly reasonable.

But being able to produce a solution does not prove that it needs to be implemented. Nor is a tool's past success enough to conclude that delegation will always be appropriate. We must be able to review our decision when circumstances change, new evidence appears, or errors we did not previously know about come to light.

Prudence therefore also enters into the decision to delegate: it requires us to consider not only what a tool can do, but whether using it makes sense in the specific case and what degree of supervision is needed. This is not a matter of distrusting technology on principle. It is about preventing the ease of getting an answer from replacing scrutiny of its necessity and consequences.

So far I have considered how our decisions depend on understanding the problem, the ends we pursue, and our attention to those who might be affected. One more concrete question remains: what consequences does this exercise have for the results we obtain and the resources we use?

Investigating an additional difficulty can prevent an expensive error; it can also consume time without improving the decision. Delegating a task can lower costs, but doing so without adequate supervision may create consequences that outweigh the initial savings.

Here philosophical reflection meets economic analysis. The question is no longer only what decision is prudent, but how to evaluate the value of that judgment, the costs of exercising it, and the risks of ceasing to do so. We must also consider whether systematically delegating certain activities might, over time, weaken the capabilities we need to evaluate the results of that very delegation.

Part Two. The economics of judgment: from the efficiency of answers to the value of decisions

III. Automation efficiency and the limits of economic decision-making

If in the first part I have tried to show that obtaining an answer is not necessarily the same as understanding a problem, I now want to examine what that difference means when we must decide how to organize work in a business.

The question is no longer simply whether a tool can produce a correct answer, but whether using it helps solve the problem better, at what cost, and under what conditions. This distinction matters because technology can substantially reduce the time needed to complete an activity without bringing about a proportional improvement in the resulting decision.

Nor would it be rigorous to assume that human intervention adds value merely by being present. A professional might spend hours examining irrelevant information, miss a risk, or insist on a review that does not alter the outcome. AI, meanwhile, can perform certain tasks faster, at lower cost, and with greater accuracy than a person under comparable conditions.

The point, therefore, is not to defend either alternative in advance. It is to establish what each contributes and when that contribution justifies the resources it requires.

To do so, we should distinguish three questions. The first is economic: which alternative offers better results once costs and risks are taken into account? The second is practical: which action is appropriate to the specific circumstances? The third is moral: which ends should we pursue, and which boundaries should we not cross, even if doing so would be profitable?

These questions are related, but not interchangeable. Economic analysis allows us to compare alternatives and estimate their consequences; prudence, in the Aristotelian sense developed in the first part, concerns deliberation about what it is appropriate to do. And no economic advantage, by itself, is enough to justify an action that violates an obligation or unjustifiably harms others.

This distinction is important to the rest of the essay. I am not trying to demonstrate that human intervention is always superior to automation. I want to examine the conditions under which judgment improves a decision, when delegation is reasonable, and what we ought to consider if that delegation changes the capacities with which we will assess future decisions.

IV. Lowering the cost of an answer does not necessarily improve a decision

In Prediction, Judgment, and Complexity: A Theory of Decision-Making and Artificial Intelligence, Agrawal, Gans, and Goldfarb (2019) propose a particularly useful distinction for understanding the economic consequences of AI: the difference between prediction and judgment.

Prediction allows us to estimate what may happen; judgment comes into play when we need to evaluate available alternatives and decide how to act on those estimates. Better predictions can expand the possibilities open to us, but do not automatically determine which alternative we should choose. For that, we need to consider each alternative's consequences and the goals we seek to achieve.

This economic distinction is not equivalent to Aristotelian prudence. Agrawal and his colleagues examine the relationship between prediction, decision, and economic value; Aristotle raises a broader question concerning action and the good toward which it should be directed. Nevertheless, both perspectives reveal a limit: having better means does not, in itself, determine what we should do with them.

Consider a business using AI to review contracts. The tool can compare versions, identify problematic clauses, and prepare comments in a fraction of the time a professional would require. If it provides adequate quality for the task, the time saving is a genuine advantage.

But imagine that the business evaluates the system solely by how quickly it produces reports. The documents may be correctly prepared and reach their recipients sooner, while a circumstance that would change the decision to accept, renegotiate, or reject the agreement remains unexamined.

The problem need not lie in an incorrect answer. It may lie in the fact that an answer, although appropriate to the question asked, fails to resolve the difficulty that truly matters.

The tool might also identify that difficulty. There is no reason to attribute to human beings an exclusive ability to spot problems that AI could never recognize. What we need to examine is whether the procedure has sufficient background information, considers the relevant risks, and includes appropriate criteria for evaluating the result.

Here an economically fundamental difference emerges: the cost of producing an answer is not necessarily the cost of arriving at a satisfactory decision. To the former we must add, as appropriate, implementation and maintenance expenditure, supervision, error correction, the losses caused by those errors, and the consequences of changing the organization of work.

That distinction led me to a question that became central to the development of my own methodology: how can we determine whether a solution is genuinely worthwhile, beyond its ability to produce a correct answer?

Before considering which parts of a procedure could be automated, I needed to understand what problem I was trying to solve, what criteria should guide the analysis, and how to assess whether the results were useful. Only then did it make sense to examine which tasks could be entrusted to AI and which called for other forms of intervention.

This order matters. If we start with what a technology can do, we risk reshaping the problem to fit the tool's capabilities. If we start with the problem, by contrast, we can evaluate technology according to what we actually need to accomplish.

That is why I did not develop my methodology by first asking how much work I could automate. I started by asking how to analyze a business situation and determine which solution best responded to its circumstances. Artificial intelligence came later, as a way of extending and carrying out parts of that work.

V. An economic criterion for comparing alternatives

As I tried to bring these reflections into the business world, I encountered a concrete difficulty: I needed a criterion for comparing alternatives without reducing the evaluation to the immediate cost of performing a task.

In my experience, an apparently simple solution may be entirely sufficient. On other occasions, examining the background reveals a risk that had not been considered, and preventing it justifies additional intervention. The opposite can also happen: pursuing a problem without sufficient reason consumes resources and does not improve the outcome.

I therefore needed a way to organize these considerations and make explicit the elements that enter into a decision. This is the origin of the expression I propose below:

Here, a represents each alternative under consideration, and V(a) its expected net economic value. The remaining terms distinguish four dimensions: expected operating benefits, B(a); implementation and operating costs, C(a); expected economic losses from risks remaining after the planned controls have been applied, R(a); and the future net economic effect associated with the organization's capabilities, K(a).

I am especially interested in including the last component because an automation decision can have consequences that do not become apparent when it is adopted. A tool may free up time and facilitate learning, but may also reduce opportunities to practice particular skills. If those skills are needed later, their loss could entail training costs, operational difficulties, or a reduced ability to respond to exceptional situations.

I do not assume that these effects are always negative. Technology may also help people develop new competencies or allow professionals to spend more time on more complex activities. That is why K(a) is a signed variable expressing the future net economic effect on capabilities: it is positive when future costs predominate and the effect is adverse to the organization, and negative when future benefits predominate and the effect is favorable. Because the formula subtracts K(a), a positive value reduces V(a), while a negative value increases it.

This expression is an analytical proposal I have developed to organize the comparison, not a universal economic formula or an empirically validated instrument. Nor does it assume that every relevant consequence can reliably be expressed in monetary terms. Its purpose is narrower: to prevent a decision from appearing attractive simply because it lowers immediate costs when it may also generate other economic effects that should be taken into account.

Two precautions are indispensable if the expression is to be applied rigorously. The first is to avoid double counting: if the loss of a skill increases the likelihood that an error will go undetected and that effect is already captured in R(a), the same loss must not be deducted again through K(a). The second is to use a common time horizon. When alternatives have costs and benefits spread across several years, future flows should be converted to present values using an explicit discount rate.

To calculate present value, each estimated annual flow of benefits, costs, risks, and capability effects is divided by (1 + r) raised to the power t, where r is the discount rate and t the number of years until the flow occurs. The discounted flows are then summed for each component, using the same time horizon for all alternatives. When sufficient data are available, we should also estimate explicitly how weakened supervision changes the probability of failing to detect errors and hence expected losses from residual risks, without counting that same effect again as a capability cost.

The formula does not eliminate uncertainty or replace deliberation. Rather, it helps me ask a more complete economic question: which alternative offers better outcomes when we consider not only the resources it consumes today, but also the risks and future consequences we can reasonably identify?

An illustrative application

To illustrate how this criterion might be used, imagine a business that needs to review its contracts and is considering three alternatives: automating most of the process, combining AI with professional supervision, or retaining a review carried out mainly by people.

The following figures are hypothetical, expressed in thousands of monetary units. They are working assumptions, not results observed in an actual business.

Under these assumptions, the alternative combining AI with professional supervision has the highest expected net value. Although its costs exceed those of full automation, the difference is offset by higher estimated benefits and lower expected losses from residual risks.

I do not, however, believe that this result proves professional supervision is always preferable. The ranking depends on the estimates and their underlying assumptions. If full automation were more reliable than expected, or professional intervention made little difference to the outcome, another alternative might be preferable.

This is why I consider it important to distinguish illustrating a criterion from demonstrating a conclusion. The example shows how a comparison changes when more economic dimensions are incorporated; it does not establish that its figures are representative of a real business or that one allocation of work is superior in every setting.

In an actual application, I would test the assumptions against available data, examine unfavorable scenarios, and determine how far each estimate would need to change before the choice changed. Where evidence was insufficient, I would prefer ranges and sensitivity analysis over a single figure with a precision the available information cannot justify.

A simple sensitivity analysis illustrates the point. In the table, the supervised option has an expected net value of 600, against 350 for full automation, a difference of 250 thousand monetary units. Holding every other figure constant, if the supervised option's total costs rose from 500 to 750, the two options would tie; above 750, full automation would have the higher net value of the two. The same reversal would occur if expected residual-risk losses for the supervised option increased from 100 to more than 350; at 350, the alternatives would tie. These thresholds are not predictions, but a way of showing how much the assumptions would have to change to alter the result.

There is, moreover, a limit I regard as essential: expected economic value cannot be the sole criterion of admissibility. An alternative does not become acceptable merely because it promises higher returns if it violates legal obligations, infringes rights, or fails to meet essential safety and quality requirements. Those boundaries must be set before comparing the economic merits of the available options.

The proposed expression is thus useful because it makes the reasons for a choice more explicit. It is not intended to decide for us, but to help us recognize what we are valuing, which consequences are included, and what may be missing from the analysis.

VI. Productivity, evidence, and the limits of collaboration between people and artificial intelligence

So far I have proposed a criterion for comparing alternatives. But the criterion will only be useful if we can check our estimates, at least in part, against the evidence available.

Empirical research offers important findings, but it does not establish a universal rule about which tasks should be automated or how much value human supervision adds.

Brynjolfsson, Li, and Raymond (2025), in a study published in The Quarterly Journal of Economics, examined the introduction of an AI-powered conversational assistant among 5,172 customer-service workers. They found an average 15% increase in productivity, measured by issues resolved per hour. The effects differed among workers: those with less experience benefited substantially, while the speed gains among more experienced workers were more modest and in some cases accompanied by reductions in quality. The study also found evidence consistent with a learning effect.

These findings matter because they show that AI may do more than accelerate a task: under certain conditions, it can help workers improve their performance. But they concern a particular activity, tool, and population. They do not establish that all kinds of automation will produce equivalent benefits.

Nor should we confuse productivity with total economic value. Resolving more issues per hour is an operational improvement, but to assess its economic effect we need to consider adoption, integration, maintenance, and supervision costs, as well as service quality and losses that errors may cause.

The same demand for evidence should apply to collaboration between people and AI. It would be inconsistent to demand empirical support when evaluating a tool and then assume, without testing, that adding a person to the process automatically improves performance.

Vaccaro, Almaatouq, and Malone (2024), through a systematic review and meta-analysis of 106 experiments, compared the performance of people, AI systems, and combinations of both. On average, human–AI combinations performed worse than the best-performing individual component. The authors also found differences by task: collaboration tended to be less favorable in decision-making tasks and more favorable in content-creation tasks.

That finding calls into question a common intuition: bringing human and artificial capabilities together does not guarantee that their combined performance will exceed the best result either can produce independently.

Collaboration can create value when each party contributes relevant capabilities and the contributions are integrated through a suitable process. It can also introduce costs, encourage uncritical acceptance of a recommendation, or make mistakes harder to identify.

It is therefore insufficient simply to add a human review at the end of an automated process. We need to establish what that review actually contributes, which errors it detects, and whether the improvement justifies the additional cost.

The meta-analysis does not show that human–AI collaboration is generally ineffective. Its findings depend on the tasks, systems, and experimental conditions included. Its more precise contribution is to demonstrate that the superiority of a combination must be verified for the activity and outcome we are seeking to improve.

Putting these studies together, I reach a conclusion fundamental to the argument of this essay: we cannot determine in advance that automation, human intervention, or a combination of the two will always yield the best result.

AI can improve productivity and support learning under certain conditions. Working with a person may improve an outcome, but may also introduce costs or reduce performance compared with the better-performing alternative on its own. The suitability of each option must therefore be examined in relation to the task, the results we want, and the specific conditions in which the work is performed.

This conclusion also obliges us to assess human intervention with the same rigor we demand of technology. It would be incoherent to require evidence before trusting an automated system while assuming that human involvement necessarily improves a process.

My interest is not in defending human supervision as a universal requirement, but in identifying when it provides benefits that justify its costs and which capabilities we must retain for that supervision to be effective. That leads to a question different from immediate performance: the consequences an automation decision may have for the organization's future ability to assess its own tools.

Economic evaluation, however, does not exhaust the issue. An organization could reduce costs by transferring serious risks to its customers or exploiting information it should not use. The savings would be real, but they would not be sufficient to justify the decision.

When an activity affects third parties, we must also consider the consequences they bear and respect boundaries that should not be crossed for economic advantage.

So far we have compared available alternatives and examined what results they produce. But there is a further issue that an initial evaluation may overlook: an option that works well today can change the conditions under which it can be used reliably tomorrow.

VII. The paradox of delegating judgment: learning and future capabilities

Imagine a business that introduces artificial intelligence into its analytical and decision-making processes.

At first, the results are favorable. Its professionals save time, automate routine tasks, and devote more attention to other activities. Productivity rises, and management decides to expand the tool's use.

Over time the process becomes routine. Recommendations are accepted because they are satisfactory in most cases. Some activities that once allowed professionals to compare alternatives, practice particular skills, and recognize mistakes cease to form part of their everyday work.

There is no reason to suppose that this will have negative consequences in every case. The tool might continue to work correctly, and the organization might use the time saved to develop other capabilities. But some of the skills needed to evaluate the system may also cease to be exercised.

The problem might become visible in an exceptional situation. The system then produces an inappropriate recommendation, perhaps because the circumstances differ from those under which its performance was tested. Professionals need to intervene, but they no longer have the same practice identifying the difficulty and weighing alternatives.

The business still has an efficient tool. What it may have lost, at least in part, is the capacity to recognize when that tool needs correcting or when it is advisable to stop using it.

What interests me in this scenario is not the claim that automation necessarily leads to a loss of capabilities. That would be a conclusion too broad for the evidence available. I want to examine a more precise possibility: an organization may secure immediate improvements by delegating certain activities while gradually reducing opportunities to practice some of the skills needed to evaluate that delegation.

If that happens, the difficulty would go beyond losing the ability to perform a task. It might also affect the capacity to recognize when an automated process is no longer reliable, when circumstances call for renewed scrutiny, or when it would be appropriate to change how responsibilities are distributed.

This possibility is especially relevant because it introduces a time dimension into economic decisions. An alternative yielding better results today may not remain the most suitable if it changes the capabilities with which an organization must assess its performance tomorrow.

I am not claiming that such deterioration always occurs, or that manual execution is the only way to avoid it. An organization can maintain competent supervision through training, independent tests, reviews of exceptional cases, and other controls. The essential point is to verify that those mechanisms work, rather than assuming they are effective merely because a person has formally been assigned to supervise the process.

The question I propose investigating is therefore more specific than a general defense of human intervention: which capabilities must be retained if an organization is to continue evaluating delegated decisions reliably?

1. Automation and the supervision paradox

This concern has precedents in the literature on automation.

In Ironies of Automation (1983), Lisanne Bainbridge examined difficulties that can arise when automated systems take over routine operations and people are left to supervise their functioning and intervene in abnormal situations.

The difficulty lies in the division of work. People must retain the capacity to intervene when something goes wrong, yet the opportunities to practice the necessary skills may diminish precisely because the system performs ordinary operations automatically.

Bainbridge studied industrial automation. Her work does not demonstrate that current generative AI systems necessarily produce the same effects. It does, however, give us reason to examine an enduring tension: we may formally preserve responsibility for supervising a system while reducing opportunities to develop the capabilities needed to fulfill it.

Athey, Bryan, and Gans (2020) offer another way to examine the problem through an economic model of how decision authority is allocated between people and AI. They analyze how that distribution may alter the incentives of a human agent to invest effort in learning the consequences of decisions.

Under particular assumptions, retaining decision authority may encourage learning, while delegating it may reduce the effort devoted to acquiring such knowledge. This is a theoretical result that depends on the model's assumptions, not evidence that every act of delegation causes skills to deteriorate.

Its relevance to this essay lies in showing that the decision about who performs an activity affects more than the immediate outcome. It can also change the incentives and capabilities of those who will need to make decisions later.

Here I find a connection with the concern that prompted my university essay on Homo videns. Sartori considered the risks that the predominance of images might pose to the exercise of certain intellectual capacities. He did not study generative AI, and his conclusions cannot be attributed to today's systems.

His reflection nevertheless allows us to ask a contemporary question: what happens when we stop receiving information alone and start receiving analyses, arguments, and recommendations that we previously had to work out for ourselves?

The answer need not be negative. The study by Brynjolfsson, Li, and Raymond provides evidence consistent with the possibility that AI can facilitate learning. These tools can help us test hypotheses, understand procedures, and recognize errors.

But they may also encourage us to accept answers we have not sufficiently examined. The outcome will depend, among other things, on how the tools are used, which opportunities for learning remain, and what mechanisms exist to test the quality of their answers.

It would therefore be unwarranted to claim that AI necessarily strengthens or weakens our capabilities. We should investigate which ways of using it develop those capabilities, which preserve them, and which may undermine them.

2. The cost of losing the capacity to supervise

In a business, the loss of a technical skill may create training expenses or make it difficult to restore a procedure that was once performed internally. But when that skill also supports supervision of an automated system, its deterioration can affect the reliability of the entire process.

We should distinguish two capabilities: performing a task and evaluating when the procedure entrusted with that task ought to be questioned.

The two may be related, but they are not identical. An organization can stop carrying out some activities manually while maintaining competent supervision through independent tests, training, quality controls, and contingency procedures. Automation does not necessarily require keeping every former execution skill intact.

The difficulty arises when delegation weakens precisely those capabilities on which supervision depends, and the organization lacks sufficiently reliable alternatives.

From an economic standpoint, this possibility needs to be included in the evaluation of automation, even if it cannot always be quantified precisely.

Under the proposed criterion, the cost of restoring skills may form part of K(a). If weaker supervisory capabilities increase the chance that significant errors will remain undetected, that effect may be reflected in R(a), provided the same consequence is not counted twice.

We should neither assign arbitrary numbers to these effects nor present a loss of skills as proven when it has not been observed. But it would be equally unjustified to exclude that loss from consideration simply because it is difficult to measure.

An alternative that appears economically superior today might cease to be so if its controls become less effective over time and the cost of responding to exceptional situations rises.

The question is therefore not only how much work a tool can perform. We must also ask under what conditions we will be able to continue evaluating its performance and what resources will be necessary to correct it when it fails.

Aristotelian prudence reappears here. Deliberating well does not mean performing every task personally. Delegation may be reasonable and in some cases preferable. But we must examine the circumstances that justify it and recognize when they have changed enough to call for reconsideration.

The ability to revisit our own decision to delegate forms part of the practical problem we are trying to solve.

This clarifies what is worth preserving in human judgment: not necessarily manual execution of every activity, but the capabilities that allow us to understand a process, assess its consequences, and recognize when we need to act differently.

3. Care as an orientation of attention and purposes

The paradox of delegation also brings us back to a question raised in the first part: the place of care in our decisions.

There I suggested that concern for others may lead us to examine a request beyond its original terms. I do not wish to turn that idea into the claim that love necessarily makes us more intelligent or that a solution developed by a person is inherently superior to one proposed by AI.

I am interested in a more precise possibility: what matters to us may influence the questions we ask, the problems we identify, and the consequences we consider before deciding.

Grant and Berry (2011) studied the relationship among intrinsic motivation, prosocial motivation, perspective taking, and creativity. Their results suggest that considering other people's needs can favor the production of ideas that are both novel and useful.

That study does not prove that love guarantees better decisions or that its findings apply automatically to every business decision. It does, however, offer reasons to consider that our orientation toward others may shape how we identify problems and develop solutions.

The connection with the previous argument is important. Before comparing the cost of different alternatives, we must identify which problem deserves to be solved and which consequences matter to the people involved. Otherwise we might assess with great precision options that address no real need.

I recognize that difficulty in my professional experience. When a company asks for a document, a review, or a technological tool, completing the request correctly may be enough. But examining the background may also reveal that the real difficulty differs from what we were initially asked to resolve.

Recognizing that difference may help us identify a risk, avoid an unnecessary investment, or conclude that there is no need to develop a new solution.

The same disposition may matter when deciding how much to delegate. If we care about the outcome our clients will receive, it is not enough to know that a tool works under normal conditions. We must also ask what would happen if it failed, who would bear the consequences, and whether the organization could recognize the problem in time.

Yet identifying a need does not guarantee that we know how to respond appropriately. Genuine concern may lead us to intervene unnecessarily, complicate a straightforward situation, or override the judgment of the person we are trying to help.

Care should therefore not be measured by the quantity of extra work we do. We need to test our interpretation against the circumstances and recognize when to stop.

If an automated solution reliably completes a task at a lower cost, with suitable mechanisms for detecting and correcting errors, insisting on additional human review may be unnecessary. If, by contrast, the decision turns on relevant circumstances excluded from the process, we need to examine which information, controls, and capabilities are required to address that gap.

Care does not replace economic analysis. It helps us recognize which needs and consequences deserve attention; economic analysis helps us assess whether the proposed intervention creates benefits sufficient to justify its costs. Moral boundaries, meanwhile, prevent us from reducing every decision to a monetary comparison.

These perspectives serve different functions. Their value lies in the possibility of combining them without confusing them.

VIII. When to automate, when to supervise, and when to intervene

The preceding considerations allow us to move from general reflection to a decision-making procedure. If we want to establish which tasks should be automated, supervised, or kept chiefly under human control, we must compare concrete alternatives without beginning from a preference for technology or for human work.

The paradox of delegation adds another requirement: it is not enough to determine which process works best when introduced. We must also consider how the capacity to assess its suitability will be maintained throughout its use.

I propose organizing this evaluation around four dimensions.

This framework is an analytical proposal within the essay, not an empirically validated quantitative instrument. Its role is to organize the questions that should guide the economic and practical evaluation of each alternative.

Significance matters because not all errors have the same consequences. An easily corrected administrative error may have a limited cost; a contractual, financial, or safety error may cause major losses and affect third parties. The seriousness of the consequences should shape the controls we establish, without implying that human intervention is always the most reliable safeguard.

Uncertainty requires us to examine available information and the conditions under which the tool has been evaluated. A system may perform well in ordinary cases but be less reliable when circumstances change. Obtaining additional information may improve a decision, but also has a cost. We need to determine whether the missing information could alter the outcome and whether acquiring it is worthwhile.

Reversibility distinguishes decisions that are easy to correct from those whose effects are difficult to repair. Where an action is reversible and its results can be detected promptly, automation with suitable monitoring may be reasonable. Where consequences are hard to reverse, a more demanding preliminary evaluation and authorization and contingency procedures may be justified.

Learning extends the analysis into the future. Automation can free time for more complex activities, but it can also reduce opportunities to practice the skills needed to detect errors. When those skills matter for an organization's continuity, we should consider how they will be retained and tested.

These dimensions do not amount to an automatic rule. A decision with serious consequences may benefit more from a properly validated automated system than from unreliable human review. The determining consideration is which procedure delivers better results within the applicable legal, ethical, and quality boundaries.

A procedure for deciding

On the basis of these dimensions, I propose five steps.

1. Define the objective and limits. Specify the problem to be solved, what counts as a satisfactory outcome, and which legal, ethical, and quality requirements every alternative must meet.
2. Identify the alternatives. Compare full automation, a combination of AI and human supervision, and predominantly human intervention, insofar as each is a viable option for the activity.
3. Estimate the economic components. Assess expected benefits B(a), total costs C(a), residual-risk losses R(a), and future economic effects on capabilities K(a). Avoid double counting and use a shared time horizon.
4. Test assumptions against evidence. Use historical data, experiments, or pilot projects to assess the quality of outcomes, examine unfavorable scenarios, and determine whether additional intervention actually improves performance.
5. Choose, supervise, and review. Adopt the admissible alternative with the best expected results, establish indicators to detect deviations, and revisit the decision when circumstances change.

I would add a cross-cutting check to these five steps: periodically verify that the organization retains the effective ability to recognize relevant errors and change or suspend the process when necessary.

Depending on the activity, this might include testing with known errors, examining exceptional situations, evaluating professionals' ability to detect incorrect recommendations, or conducting contingency exercises. Not every process needs the same controls, nor is it essential to preserve every manual execution skill. What must be verified is that supervision mechanisms remain appropriate for the risks they are meant to control.

A business might formally employ professionals to review a system's recommendations and nevertheless discover that the review no longer detects certain errors. The mere existence of human supervision would not prove that the process was adequately controlled.

This ties back to the paradox examined earlier: if delegation can alter the capabilities of those responsible for evaluating it, the quality of supervision must also be tested over time.

It will not always be possible to assign a reliable monetary value to every variable. When evidence is lacking, it is better to work with ranges, scenarios, or sensitivity analyses than to invent figures that create an appearance of precision. Minimum quality requirements and risk limits can also be set before economic comparisons are made.

Nor must we find a perfect solution before acting. A business can begin with a limited trial, compare outcomes against a reference process, identify errors, and adjust controls in light of the evidence obtained.

The goal is not to eliminate all uncertainty, but to make decisions proportionate to the risks, available resources, and information we genuinely possess.

IX. The value of judgment depends on what it achieves and what we must consider

Bringing these considerations together, I think we can identify more precisely where the value of judgment lies in business decisions.

It does not necessarily lie in personally performing every task. Human intervention can be economically valuable when it improves outcomes, identifies a relevant problem, reduces risks, or preserves capabilities needed for future action. If it does not produce a sufficient improvement to justify its costs, maintaining it as a matter of principle would lack an economic basis.

Aristotelian prudence nevertheless raises a question that cannot be exhausted by that comparison: deliberating over means does not, by itself, determine what ends we ought to pursue. Care extends our attention to the needs and consequences affecting others, while economic analysis helps us examine the resources and outcomes associated with the alternatives available.

These perspectives serve different functions. Economic evaluation allows us to compare admissible options; prudence invites us to consider which action is appropriate in the particular circumstances; and legal and moral limits prevent any result from being justified solely by profitability.

Artificial intelligence can contribute at every stage through information analysis, comparison of alternatives, and identification of potential difficulties. In some circumstances delegation may be the best decision. In others, we will need additional controls or to reconsider whether delegating is advisable.

In my view, the paradox of delegation adds a requirement we should not overlook: it is not enough to verify that a tool produces satisfactory results. We must also know whether we retain the capacity to recognize when those results cease to be reliable and to revise our decision when circumstances require it.

This does not mean preserving human judgment as a kind of privilege over technology. It means treating judgment as a capacity that must demonstrate its usefulness, be maintained, and be developed wherever it is needed.

Even after identifying the most economically attractive option among admissible alternatives, a question remains that calculation cannot settle on its own: what ends should we pursue, and what consequences are we prepared to accept in achieving them?

It is in this relationship among evaluating means, deliberating about ends, and accepting responsibility for consequences that I locate the value of human judgment in the age of artificial intelligence.

Conclusion

I began this essay by asking what happens when an answer is so convincing that we no longer feel the need to examine it.

Returning to Homo videns, I found a concern already present in my university reflections: the importance of preserving our capacity for abstraction and not confusing access to information with a sufficient understanding of reality.

Artificial intelligence profoundly changes the terms of that question. We can now receive not only information but also interpretations, arguments, and recommendations that make our work easier in extraordinary ways. Yet that possibility does not remove the need to examine the problem we are trying to solve or guarantee that the answer provided is sufficient to address it.

Aristotle takes us to an additional difficulty: understanding a situation does not necessarily determine what ought to be done. Practical wisdom requires us to deliberate about our actions, taking account of particular circumstances and the good we are trying to achieve.

Along the way, I have also wanted to emphasize the place of love and care. I believe that when someone matters to us, our disposition to understand can change. We may attend to circumstances we had not previously considered or feel the need to question a solution that appears sufficient.

That does not make us infallible. Precisely because other people's well-being matters, we need to examine our interpretations, test what we believe we understand, and recognize when an intervention may be unnecessary. Caring does not mean always doing more; it means trying to ensure that what we do actually responds to the circumstances.

Brought into the business world, these reflections acquire an economic dimension. AI can lower the cost of producing answers and improve productivity, but whether using it is worthwhile depends on more than speed. We must consider the outcomes, total costs, residual risks, and effects automation may have on the organization's future capabilities.

The evidence considered does not support a universal rule. In some tasks, AI improves performance and helps people learn. In others, combining it with human intervention does not guarantee better outcomes. Both automation and supervision must therefore be evaluated according to their actual effects.

The paradox of delegation introduces a further difficulty: we can benefit from a tool while also reducing opportunities to practice some of the skills needed to evaluate its operation. This does not demonstrate that all automation weakens judgment, but it does justify examining how supervisory capabilities will be retained and how the decision to delegate will be revisited as circumstances change.

My conclusion is not that we should reserve to people every task a machine can perform, nor that we should embrace automation whenever it promises savings. I propose evaluating each alternative in light of the problem it is meant to solve, its consequences, and the boundaries we must respect.

Perhaps the real challenge is learning to delegate without giving up on understanding what we are delegating; taking advantage of technology's capabilities without confusing operational efficiency with the wisdom of our decisions.

I return, then, to the wonder with which I began. The curiosity that moves us to ask why something happens can be the beginning of deeper understanding, but it can also force us to acknowledge that we do not yet know enough to decide.

AI can help us in that search. It can expand our possibilities, reveal connections we had overlooked, and challenge our first interpretations. What matters will be how we use those contributions, which criteria we employ to evaluate them, and whether we are prepared to reconsider our conclusions when reality does not fit them.

I would therefore not locate the value of human judgment in an obligation always to find an answer technology cannot provide. I would locate it in our responsibility to examine problems, deliberate about ends, and answer for the decisions we make.

We do not need to prove that human beings always judge better than machines. We need to understand what makes a decision appropriate, what we must consider when making it, and how to preserve the ability to recognize when we have been wrong.

Bibliographical references

Agrawal, A., Gans, J. S., and Goldfarb, A. (2019). “Prediction, Judgment, and Complexity: A Theory of Decision-Making and Artificial Intelligence.” In The Economics of Artificial Intelligence: An Agenda (pp. 89–110). University of Chicago Press. https://doi.org/10.7208/chicago/9780226613475.003.0003

Aristotle. Nicomachean Ethics, Book VI, particularly chapters 4–5.

Aristotle. Metaphysics, Book I, chapter 2, 982b12–21.

Athey, S. C., Bryan, K. A., and Gans, J. S. (2020). “The Allocation of Decision Authority to Human and Artificial Intelligence.” AEA Papers and Proceedings, 110, 80–84. https://doi.org/10.1257/pandp.20201034

Bainbridge, L. (1983). “Ironies of Automation.” Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8

Benedict XVI. (22 September 2011). Address to the German Bundestag, Reichstag, Berlin. Holy See. https://www.vatican.va/content/benedict-xvi/en/speeches/2011/september/documents/hf_ben-xvi_spe_20110922_reichstag-berlin.html

Bergson, H. (1896). Matière et mémoire. Essai sur la relation du corps à l’esprit. Félix Alcan.

Brynjolfsson, E., Li, D., and Raymond, L. (2025). “Generative AI at Work.” The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044

Francis. (24 January 2022). Listening with the Ear of the Heart. Message for the 56th World Communications Day. Holy See. https://www.vatican.va/content/francesco/en/messages/communications/documents/20220124-messaggio-comunicazioni-sociali.html

Grant, A. M., and Berry, J. W. (2011). “The Necessity of Others Is the Mother of Invention: Intrinsic and Prosocial Motivations, Perspective Taking, and Creativity.” Academy of Management Journal, 54(1), 73–96. https://doi.org/10.5465/AMJ.2011.59215085

Sartori, G. (1997). Homo videns. Laterza. Spanish edition: Homo videns. La sociedad teledirigida. Taurus, 1998.

Vaccaro, M., Almaatouq, A., and Malone, T. (2024). “When Combinations of Humans and AI Are Useful: A Systematic Review and Meta-Analysis.” Nature Human Behaviour, 8, 2293–2303. https://doi.org/10.1038/s41562-024-02024-1

The Holy Bible. First Book of Kings, 3:9. The interpretation developed in this essay also draws on the texts by Benedict XVI (2011) and Francis (2022) cited above.