Who Decides What Matters? AI, Attention and Human Judgement

Artificial intelligence is rapidly becoming part of how professionals make decisions. This immediately raises questions about responsibility. If an AI system analyses evidence, identifies patterns, evaluates alternatives and recommends a course of action, at what point does the decision cease to be meaningfully human?

The usual way of approaching this problem is to ask where assistance ends and delegation begins. We imagine a boundary somewhere along the decision-making process: AI may retrieve information, organise it, perhaps analyse alternatives and make recommendations, while a human being retains the authority to make the final choice.

But this way of drawing the boundary may already concede too much.

The influence of AI does not begin when it recommends a decision. It begins earlier, when it helps construct the representation of the situation upon which that decision will be based.

Before asking who makes the decision, perhaps we should therefore ask a different question:

Who decides what matters?

The decision before the decision

We tend to imagine decision-making as a relatively linear process. We gather information, interpret it, compare possible outcomes, exercise judgement and eventually decide. Within this model, it seems relatively easy to locate AI. It can assist with research, process large quantities of information, identify patterns, generate predictions or recommend possible courses of action. The human being remains at the end of the process, retaining the authority to accept or reject what the system proposes.

But this description may underestimate what happens earlier.

Imagine an executive asking an AI system to analyse hundreds of pages of documents before an important decision. The system produces a remarkably clear report. It identifies the principal issues, establishes relationships between events, distinguishes relevant information from background material and highlights patterns that would have been difficult for an individual reader to detect.

The system has not made the decision. But neither has it merely transmitted information.

It has organised a field of attention.

Some facts have become central while others have become peripheral. Certain relationships have been emphasised while others have disappeared from view. Events have been placed into a chronology. Categories have been created. What was initially an ambiguous and potentially overwhelming situation has acquired a structure.

The human decision-maker now encounters the problem through that structure.

We may therefore retain formal responsibility for a decision while progressively delegating parts of the cognitive process through which the situation itself becomes intelligible.

A summary, for example, is never simply a shorter version of reality. It distinguishes what is central from what can be omitted. A chronology places some events in meaningful proximity to one another. A comparison establishes the dimensions along which alternatives will be evaluated. A risk analysis directs attention towards certain possible futures rather than others.

None of this requires AI to tell us what to do. Yet each of these operations can influence the architecture within which judgement subsequently takes place.

This suggests a distinction between using AI to amplify perception and allowing it to allocate salience. In the first case, the technology enables us to perceive something we might otherwise have missed. In the second, it begins to influence what we consider worthy of attention in the first place.

The distinction is not absolute. Nor is allocating salience something peculiar to machines. Human beings do it continuously. Professional training teaches us what to notice. Organisations do it through reporting structures, KPIs, categories and incentives. Cultures provide narratives through which some events become meaningful while others remain almost invisible.

No decision-maker encounters an unfiltered reality.

The important question is whether we remain capable of noticing the frame through which reality is being presented to us.

AI can reveal anomalies we have missed, connect information that previously appeared unrelated and make patterns perceptible across quantities of data no individual could reasonably hold in mind. In that sense it genuinely enlarges the perceptual field available to a decision-maker. Yet the very coherence of the resulting picture creates another difficulty. A representation that is sufficiently comprehensive, persuasive and useful can cease to be experienced as a representation at all.

We begin to inhabit its distinctions. What it places in the foreground attracts our attention, while what it leaves outside the frame becomes progressively more difficult to notice.

The problem is therefore not exhausted by the possibility that AI might give us the wrong answer. It may influence the conditions under which some answers become thinkable in the first place.

When competence is not enough

What happens when an AI system performs its assigned task extremely well, but the task itself has been defined too narrowly?

Consider an AI agent used in procurement. The system is given an apparently reasonable objective: reduce purchasing costs. It analyses contracts and suppliers and negotiates more favourable terms. On its own metric, it performs impressively.

The difficulty may appear at another level. Continued optimisation could place increasing pressure on a strategically important supplier, eventually threatening the resilience of the supply chain and producing losses vastly greater than the savings the system generated.

The interesting point is that the system does not need to malfunction.

It may have done exactly what it was asked to do.

What failed was the relationship between the objective and the wider field within which the objective had meaning.

Reducing procurement costs makes sense within a particular horizon. Maintaining the resilience of a supply chain requires a wider one. Preserving strategically important relationships may require a wider horizon still.

The problem therefore concerns not only the quality of reasoning but the scale at which the situation is being perceived.

Something can be beneficial at one level and damaging at another. A department can optimise its own performance while creating costs elsewhere in an organisation. A manager can maximise a measurable target while gradually undermining trust. A company can increase short-term efficiency while weakening capabilities whose importance will become apparent only years later.

Consider an executive team deciding whether to close a business unit. AI may give them increasingly sophisticated financial forecasts, productivity comparisons, customer concentration data, market scenarios and predictions about the consequences of different choices. All of this can improve the quality of the decision.

Harder to represent may be the capabilities that disappear with the unit, the informal relationships connecting it to the rest of the organisation, knowledge that has never been codified, or what employees elsewhere will infer from the way the decision is taken.

None of these considerations necessarily constitutes an argument against closing the business unit. They belong instead to the wider field within which the meaning of the decision has to be understood.

This is where I find the idea of attunement useful.

Competence describes, among other things, our capacity to perform effectively within a given domain or towards a given objective. Attunement adds sensitivity to the wider situation in which that competence is being exercised: context, relationships, changing conditions, different scales of consequence and signals that may not fit comfortably within the existing frame.

A system, team or individual can therefore be highly competent while poorly attuned.

AI makes this tension particularly visible because it can pursue specified objectives with extraordinary consistency. But the underlying problem is not technological. People and organisations have always been capable of becoming extremely effective at doing something that no longer makes sense.

The challenge is maintaining sufficient contact with the wider environment to notice when this is happening.

AI enters the ecology of attention

Leaders operate in environments containing vastly more information than they can possibly process. Their attention must therefore be selective. Some signals become salient and others remain in the background. Experience affects this selection, as do professional training, organisational culture, incentives, emotions, expectations and relationships.

AI is now entering this ecology of attention.

Increasingly, it will help determine which signals reach decision-makers, how those signals are organised and which relationships between them become visible.

There is enormous potential here. An executive working with AI may be able to examine a situation from more perspectives, test assumptions more rapidly, compare scenarios that would previously have required weeks of analysis and discover patterns that nobody in the organisation had noticed.

But more information does not necessarily produce a wider horizon of judgement.

We can know more while seeing through a narrower frame.

Leadership decisions rarely take place in situations where every relevant variable can be specified in advance. They unfold within overlapping systems of relationships, incentives, histories, expectations and values. Consequences appear at different times and different organisational levels. Some of the most consequential information may be explicit and measurable; some may reside in relationships, trust, organisational memory or tacit knowledge.

There is also a participatory dimension that is easy to overlook. A leader is not simply observing an organisation from outside and selecting the correct answer. The act of deciding changes the environment in which subsequent decisions will be made. People interpret what leaders attend to, what they ignore, which questions they ask and which forms of knowledge they treat as legitimate. Decisions alter trust, expectations and future behaviour.

The decision-maker is therefore part of the field being interpreted.

This makes judgement more than an exercise in information processing. It involves remaining sensitive to a situation while participating in it.

AI can contribute enormously to this process, but its increasing power makes it more, rather than less, important to ask how the field itself has been constructed.

Cognitive sovereignty without cognitive isolation

The alternative to cognitive delegation cannot simply be refusing to delegate.

Human cognition has never been isolated.

We think through language, writing, diagrams, institutions, professional practices and other people. A board meeting distributes cognition across several individuals. An organisation stores knowledge in procedures, documents, habits and relationships. Technologies have always extended what human beings can remember, calculate and perceive.

AI dramatically extends this process.

The relevant question is therefore not whether cognition is being distributed. It already is. The question is whether we remain capable of recognising, evaluating and governing that distribution.

When working with an AI-generated analysis, for example, it may be useful to retain some awareness of what appeared important before consulting the system and then notice what has changed afterwards. What has become more salient? What has receded? Which objective is the system implicitly optimising? What assumptions define its representation of the situation? At what level does its recommendation make sense, and what becomes visible if we move one level higher or lower?

We might even ask what we would notice differently if the system had reached the opposite conclusion.

These are not questions designed to protect human beings from AI. They are part of learning how to think with it without becoming unaware of how it is shaping our thinking.

The same principle applies beyond technology. Good judgement has always required the capacity to move between immersion in a frame and the ability to examine the frame itself.

What changes with AI is the speed, scale and persuasive coherence with which such frames can now be produced.

What remains human in human judgement?

As AI becomes more capable, genuinely human judgement may become more important rather than less. But this requires a richer account of what judgement actually involves.

Judgement is not simply the final moment at which somebody chooses option A rather than option B. It begins much earlier.

We sense a situation. Some features become salient. We interpret what they mean. We anticipate what may happen next. We relate possible consequences to objectives, values and responsibilities. Eventually we act, observe what happens and revise our understanding.

AI can participate in almost every stage of this process.

This is precisely why retaining the final click of approval is an inadequate definition of human agency.

Human responsibility cannot consist merely in placing a person ceremonially at the end of an increasingly automated cognitive chain. It also involves responsibility for how the problem has been framed, what has been included, what has become salient, what has disappeared into the background and at what level of the system a proposed solution makes sense.

There is also something more difficult to formalise.

Professional and leadership judgement frequently involves knowing that the available representation of a situation is incomplete. Trust matters. History matters. Timing matters. Relationships matter. Silence can matter. Sometimes the most significant signal is precisely the one that does not fit the existing model.

And occasionally the appropriate response is not to optimise more intelligently within the frame, but to recognise that the frame itself needs to change.

AI gives us access to forms of perception that would previously have been impossible: more information, more patterns, more comparisons and more possible futures. The question is what happens to our own attention as we increasingly begin to see through these systems.

This leaves me with a question that seems more important than whether a particular decision was ultimately made by a human or a machine:

How can we use artificial intelligence to enlarge the field of perception without silently allowing it to determine the field of relevance?

The answer, I suspect, has less to do with defending human beings against intelligent machines than with becoming more attentive to how judgement itself is formed.