Being a Selfish Generation

Every generation believes it is living through the most consequential shift in human history. Most are wrong. But every so often, something changes not just in what humans can do, but in what they are asked to be. We may be at one of those moments, and the strange thing is, it has arrived not in the form of a revolution, but an economics textbook.

The book is Prediction Machines, by Agrawal, Gans and Goldfarb. Its central claim is deceptively modest: AI is not magic, it is not consciousness, it is a prediction technology, a tool for filling in missing information. What happens when a diagnosis has to be made, with a road but not yet a safe path, a person but not yet a decision about whether to trust them with a loan, AI predicts the missing piece. And it does so, now, very cheaply.

That cheapness is the philosophical event. Because when something foundational becomes cheap, it does not merely change what we can afford. It changes what we are forced to become. The book frames this as an economics argument. I want to read it as something older and stranger a question about the self, about judgment, and about what it means to be human when the machines have taken over our guesswork.

I. The oldest trick in the mind’s book

We have always predicted. Every time you reach for a cup of coffee without watching your hand, every time you finish someone’s sentence, every time you slow down before a bend you can’t yet see you are predicting. The brain, Jeff Hawkins has argued, is fundamentally a prediction organ. It does not wait for the world to arrive. It builds a continuous model of what should arrive next, notices anomalies, and updates accordingly.

In this light, what AI is doing is not alien to us. It is, in some sense, a mirror a mirror that does one thing we do, scaled to industrial volume and stripped of everything else.

“Prediction is the process of filling in missing information. It takes information you have and uses it to generate information you don’t have.”

The authors are careful here. They do not say AI thinks. They say it predicts. And prediction when reframed this way turns out to encompass a surprising amount: diagnosing a disease, flagging a fraudulent transaction, translating a sentence, recognising a face, anticipating what a driver would do at an unmarked intersection. These are not forecasts about the future. They are the filling of missing labels in the present. And that distinction matters enormously.

Because once you see it that way, you begin to see prediction everywhere. Not as a specialised act reserved for meteorologists and market analysts, but as the constant, invisible scaffolding of cognition itself. And the moment you see it everywhere, you must confront the question: what happens when that scaffolding is no longer yours to build?

* * *

II. From forecasting to judgment ; the great handover

The most significant shift that cheap prediction brings is not in what machines can do. It is in what we will stop doing ourselves.

For most of human history, prediction and judgment were performed together as a single mental act. You assessed the situation, guessed the outcome, weighed the options, and decided. all in one continuous movement of thought. The guess and the choice were inseparable. We were, in the language of economics, bundled.

AI unbundles them.

When a machine can tell you with reliable confidence what is likely to happen, which customer will churn, which loan will default, which road condition requires braking the question of what will happen is no longer the hard part. The hard part becomes what do I want to happen. That is judgment. And the authors are clear: judgment remains a human act, at least for now.

We will spend less time guessing what will happen and more time deciding what we want to happen.

This is not a trivial rearrangement. Judgment, in this framework, is the process of determining the reward function ; the articulation of values, priorities and acceptable trade-offs that gives a prediction its direction. A machine can tell you there is a twenty percent chance of rain. It cannot tell you whether you should care more about arriving dry or travelling light. That choice requires a self. It requires preference. It requires, in the deepest sense, a point of view.

And here is where the title of this essay begins to earn itself. We are a generation that has grown up being asked to be objective, to be data-driven, to suppress the self in favour of evidence. And now the machine arrives and says: I will handle the evidence. What do you actually want? It forces a return to the self not as a failure of rationality, but as its necessary complement.

We are not abandoning thought. We are being asked to do only the part of thought that cannot be automated: the part that requires knowing who you are.

* * *

III. The end of good enough

There is a concept in behavioural economics called satisficing a portmanteau of “satisfying” and “sufficing.” It describes how humans make decisions under cognitive load: not by finding the best possible option, but by finding the first option that clears an acceptable threshold. We satisfice because thinking is expensive. Every additional scenario we model, every additional branch of the decision tree we explore, costs something; attention, time, energy, the capacity for care.

We have built entire institutions around satisficing. Standard operating procedures. Rules of thumb. Bureaucratic protocols. These are not failures of intelligence. They are adaptations to scarcity, the scarcity of cognitive bandwidth in a world of infinite complexity.

Cheap prediction changes that equation.

When a machine can rapidly and accurately fill in the gaps that our mental shortcuts were designed to paper over, we no longer need to settle for good enough. We can afford to ask: what if we considered every possible scenario? What if we didn’t rely on the rule, but on the actual prediction for this specific case, in this specific context, at this specific moment?

Instead of relying on rigid rules or mental shortcuts to manage uncertainty, our thought processes will become more nuanced and more contingent.

This is genuinely liberating. But it also carries a weight we may not be ready for.

Rules are not just cognitive shortcuts. They are also moral insulation. When we apply a rule uniformly, we are protected from accusations of arbitrariness, of favouritism, of discrimination. The rule is the reason. But when every decision is bespoke, when the machine produces a tailored prediction for every individual case, the question of the reward function becomes inescapable. Why are we optimising for this outcome and not that one? Whose good is the machine predicting toward?

The end of satisficing does not just liberate us from cognitive laziness. It exposes us to the full weight of our own values. And that is a much harder thing to carry than a rulebook.

* * *

IV. Reframing — the cognitive skill of the age

Perhaps the most quietly radical idea in Prediction Machines is this: the future will belong not to those who have the answers, but to those who know how to ask the right questions. Specifically, how to reframe a problem as a prediction problem.

The authors use autonomous driving as their central example. For decades, the engineering approach was rule-based: code every possible scenario, write an if-then response for every hazard, build a decision tree comprehensive enough to cover the road. This approach failed, not because the engineers were insufficiently clever, but because the world is irreducibly complex. You cannot enumerate all possible situations in advance.

The breakthrough came from reframing: instead of asking what is the correct response to this road condition, engineers asked what would a human driver do? The problem shifted from rule-generation to prediction. And once framed that way, it yielded to machine learning in a way it had never yielded to explicit programming.

This is a cognitive move. It is a specific way of looking at a problem, stepping back from the content of the problem to ask whether it has a different shape than the one we first assumed. The authors call the capacity to do this AI insight, and they suggest it will become one of the defining intellectual skills of our era.

Human creativity will shift toward problem decomposition and reengineering, toward asking which problems are, at their core, problems of missing information.

But I want to sit with something the book does not fully resolve: the act of reframing is not neutral. When you decide to treat a medical diagnosis as a prediction problem, you are making a choice about what medicine is. When you reframe education as a prediction problem, predicting which student will succeed, or which intervention will help, you are making a choice about what learning means. The reframe is not just a technical move. It is a philosophical one.

And this is where the title comes back, with its full weight. We are the generation that will decide which human experiences to reframe as prediction problems, and which to guard from that framing. That is a form of power. It is also a form of responsibility. And it is ours alone.

* * *

V. The purity problem — bias as the price of meaning

There is an inherent contradiction lurking at the heart of all of this, and I find it the most generative tension in the book.

Prediction, in its ideal statistical form, aspires to purity. It wants to be unbiased; correct on average, free of prejudice, a clean mirror held up to reality. And yet, as we have seen, a prediction without a reward function is useless. It is a probability with no direction, a number with no instruction attached. To be meaningful, a prediction must be coupled to a judgment about what we want. And judgment, by its nature, is partial. It reflects values. It reflects history. It reflects the particular, contingent, interested self that is doing the judging.

When we train a machine to predict human judgment, as in the autonomous driving case, we are not just asking it to mirror reality. We are asking it to mirror us. And we are not “pure”. We carry centuries of structured inequity in the data we generate. When the machine learns from that data, it learns our biases too. Not as exceptions. As features.

We are not so much corrupting the purity of prediction as recognising that prediction alone is useless without a value system.

So here is the contradiction laid bare: we need bias, in the sense of a value system, a reward function, a point of view, for prediction to be useful. But bias, in the sense of prejudice and historical injustice, is exactly what we hope machines will help us escape. These two meanings of the same word are in permanent, productive tension.

The answer the book gestures toward is reward function engineering; the deliberate, reflective articulation of what we actually want, as opposed to what we have historically done. This is harder than it sounds. It requires a generation willing to be honest about its values, precise about its priorities, and humble about the ways its judgments have already been shaped by forces it did not choose.

Which is to say: it requires exactly the kind of self-knowledge that an age of outsourced prediction might, if we are not careful, allow us to avoid.

* * *

Coda: what we owe the machines we are building

This is what it means to be a selfish generation, and I mean that not as an accusation but as a precise description of our situation. We are the generation that will take from AI the burden of prediction while retaining for ourselves the privilege of judgment. We will hand over the guesswork and keep the values. We will outsource the statistics and claim authorship of the outcomes.

That is, in one sense, an extraordinary gift. It frees us to be more fully human, to focus on what we want, on what we believe, on who we are. The machines will fill in the missing information. We will decide what the information is for.

But the gift comes with an obligation that we are perhaps only beginning to understand. If the machines will predict toward our reward functions, then we must be exquisitely careful about what we reward. We must be willing to examine not just what we want, but why, whether our wants are truly ours, or merely inherited biases dressed in the language of preference.

The authors of Prediction Machines remain, by their own admission, agnostic on the deepest questions: whether prediction is the foundation of intelligence, whether the gap between machine prediction and human thought will one day close, whether what we are building is a tool or a mirror or something else entirely.

I find that agnosticism honest and right. We do not yet know what we are building.

What we do know is this: we are the first generation to have the choice. The prediction machines are here. They are cheap, they are fast, and they are waiting. What they are waiting for; what they will always be waiting for, until something fundamental changes is us to tell them what we want.

That question what do we actually want? may be the most important thing our generation is asked to answer. And unlike every question we have handed to the machines, this one, we cannot hand away.

This essay is a philosophical reading developed through a conversation with NotebookLM, using Prediction Machines: The Simple Economics of Artificial Intelligence by Ajay Agrawal, Joshua Gans and Avi Goldfarb (Harvard Business Review Press) as the primary source. The ideas in italics and quotation are drawn from or paraphrased from that text. The interpretations, contradictions and conclusions are the author's own.