The Two Boxes: Newcomb's Paradox and the Battle Over Rational Choice

The Hook: A Game That Splits Clever People Into Two Camps

Imagine an alien being — call it the Oracle — lands on Earth. It claims to be able to predict human behavior with almost perfect accuracy, and it has already been proven right thousands of times. Now it presents you with a choice. Two boxes stand before you. Box A is transparent and visibly contains 1,000 euros. Box B is opaque; it contains either a million euros or nothing at all.

The rules are simple. You may take either just Box B, or both boxes. The catch: the Oracle already fixed the contents of Box B yesterday, based on its prediction of what you will do today. If it predicted that you would take only Box B, it put the million inside. If it predicted that you would take both boxes, it left Box B empty. The prediction has been made, the money is already locked in place, and the Oracle departed long ago. Nothing you do now can change what is in the boxes.

What do you do?

This seemingly harmless question is one of the most famous and most stubborn paradoxes in modern philosophy. It is called Newcomb's Paradox, and its appeal lies in the fact that two entirely legitimate, centuries-old principles of rational choice collide head-on here and lead to opposite answers. Clever, thoughtful people split into two camps over it — and each camp regards the other side as obviously confused. The philosopher Robert Nozick, who made the problem famous, remarked dryly: "To almost everyone, it is perfectly clear and obvious what should be done. The difficulty is that these people seem to divide almost evenly on the problem, with large numbers thinking that the opposing half is just being silly."

Why should this interest you beyond the intellectual thrill? Because Newcomb's Paradox is the origin of an entire scientific discipline — modern decision theory — and because the very same question sits today at the center of a highly practical field: how do you build an artificial intelligence that makes good decisions? The moment an agent knows that others can predict, model, or even copy its behavior, it is in a Newcomb-like situation. The thought experiment is thus no longer a curiosity but one of the sharpest probes we have for understanding what "rational choice" even means.

This article takes you the whole way: from the origin of the problem through the two competing arguments and the birth of causal decision theory, all the way to the modern variants and the surprising relevance for AI safety.

Part 1: Where the Problem Comes From

A Physicist Thinking About the Prisoner's Dilemma

The paradox bears the name of its inventor, but he never published it himself. William Newcomb was a theoretical physicist at the Lawrence Livermore National Laboratory in California — and, incidentally, a great-grandnephew of the astronomer Simon Newcomb. Around 1960 he was brooding over the already-known Prisoner's Dilemma and over the question of how one ought to behave if the other player could perfectly anticipate one's own choice. Out of that brooding he distilled the clean, sharpened form with the two boxes and the all-knowing predictor.

Newcomb passed the puzzle on to friends, and by way of the philosopher Robert Nozick it reached the academic world. Nozick published the essay Newcomb's Problem and Two Principles of Choice in 1969 and gave the problem its canonical shape. In it he made explicit what constitutes the real scandal: here it is not a correct argument facing a wrong one, but two almost sacred principles of rationality, both entirely compelling on their own — and contradicting each other.

The Leap Into the Public Eye

The problem became famous to a broader public through Martin Gardner, the legendary columnist of the Mathematical Games section in Scientific American. Gardner presented Newcomb's Paradox in 1973 (the detailed treatment with Nozick's afterword appeared in March 1974) and thereby triggered an avalanche of reader mail. The incoming letters favored the solution of taking only one box by a ratio of roughly five to two. The personal lineup of the participants is telling: William Newcomb himself, the inventor, was a committed "one-boxer." Nozick, by contrast, professed to be a lukewarm "two-boxer," even though the decision theorists Maya Bar-Hillel and Avishai Margalit urged him to join the "millionaires' club" of one-boxers.

That even the sharpest minds could not agree — and to this day cannot — is no sign of carelessness. It is a symptom that a genuine crack in our concepts is becoming visible here.

Part 2: The Two Arguments

The whole magic of the paradox lies in the fact that a conclusive, almost irrefutable argument can be formulated for each answer. Let us look at them one at a time.

The Dominance Argument: Take Both Boxes!

The first argument rests on a principle so fundamental that it sounds almost like a definition of reason itself: the dominance principle. It states: if an action gives you an at-least-as-good outcome in every possible state of the world, and a better outcome in at least one state, than another action, then choose the first. To spurn a dominant option would simply be irrational.

Apply this to the boxes. The contents of Box B are already fixed. There are only two possibilities:

First: the million is in Box B. If you take only B, you get 1,000,000 euros. If you take both, you get 1,001,000 euros. Taking both boxes is better by 1,000 euros.

Second: Box B is empty. If you take only B, you get 0 euros. If you take both, you get 1,000 euros. Again it is better by 1,000 euros to take both.

The result is startlingly unambiguous: no matter what is in Box B, you always come out exactly 1,000 euros ahead by taking both boxes. The 1,000 euros in the transparent box are a pure gift — why would you leave them behind? The contents of B no longer change, regardless of what you decide. So: take both. Whoever takes only one box knowingly leaves 1,000 euros on the table.

The Expected-Utility Argument: Take Only One Box!

The second argument rests on an equally venerable principle: the maximization of expected utility. For each action, compute the probability-weighted average utility, and choose the action with the highest value. This very principle is the foundation of all statistics, actuarial mathematics, and rational betting.

Apply it to the boxes — and factor in the decisive fact that the Oracle is a near-perfect predictor. Let us assume a hit rate of 99 percent.

If you take both boxes, the Oracle will, with 99 percent probability, have predicted this and left B empty. Your expected value: 0.99 × 1,000 + 0.01 × 1,001,000 = about 11,000 euros.

If you take only Box B, the Oracle will, with 99 percent probability, have predicted this and put the million inside. Your expected value: 0.99 × 1,000,000 + 0.01 × 0 = 990,000 euros.

The difference is enormous: roughly 990,000 against 11,000 euros. The argument is compellingly simple: the people who take only one box almost all become millionaires. The people who take both boxes almost all go home with a laughable 1,000 euros. If rationality has anything to do with winning, then you can see with the naked eye who is winning here. So: take only one box.

The Vise

Now the trap snaps shut. Both arguments are valid. Both rest on a principle that you would apply without hesitation in nearly any other situation in your life. And yet they lead to opposite actions. This is precisely the paradox: not the difficulty of finding the right answer, but the fact that two pillars of reason break apart here.

An often-quoted taunt captures it: If you're so smart, why aren't you rich? — so jeer the one-boxers at the two-boxers. And they retort: If you're so smart, why do you knowingly leave 1,000 euros behind that your choice can no longer affect?

Part 3: The Birth of Causal Decision Theory

Newcomb's Paradox was no mere brain-teaser. It forced philosophy to draw a distinction that had previously lain in the fog: the difference between correlation and causation at the heart of decision-making. Out of that compulsion a new branch of theory was born.

Evidential Decision Theory

The older view, originally worked out by Richard Jeffrey, is called evidential decision theory (EDT). Its guiding idea: evaluate an action by the news it delivers about the state of the world. Formally, EDT computes the expected utility of an action using conditional probabilities — that is, the probability of a world state given that you perform the action.

For Newcomb's problem EDT is unambiguous: your taking only one box is strong evidence that the Oracle predicted this and put the million inside. The action "only one box" is good news about your wealth. So: take only one box. EDT is a one-box theory.

The objection to EDT is as old as it is famous and runs under the heading of "medical Newcomb cases": suppose a gene causes both the desire to smoke and cancer, while the smoking itself is harmless. According to naive EDT you should not smoke, because the choice to smoke is bad news about your gene — even though smoking has no causal effect on the cancer at all. This seems absurd: why should I deny myself a pleasure that demonstrably does me no harm, merely because the choice is statistically correlated with something bad? This objection puts its finger on EDT's sore spot: it confuses, so the critique goes, the announcing of good news with the bringing about of good outcomes.

Causal Decision Theory

In response to precisely this weakness, Allan Gibbard and William Harper (1978), building on preliminary work by Robert Stalnaker, formulated causal decision theory (CDT) in the late 1970s. Its guiding idea: evaluate an action not by what it indicates but by what it brings about. Formally, CDT replaces the conditional probabilities with the probabilities of subjunctive conditionals — statements of the form "were I to do X, then Y would occur." What matters is the causal influence of the action, not its mere correlation with the outcome.

For Newcomb's problem CDT is equally unambiguous — only in the other direction: your choice now has no more causal influence on the contents of Box B, which were fixed yesterday. If the million were inside, you would get more with both boxes; if it were empty, likewise. So: take both. CDT is a two-box theory. In the smoking-gene case CDT delivers the intuitively correct result: go ahead and smoke, because your choice causes no cancer.

The Underlying Question

This lays bare the actual core of the paradox. Newcomb's problem is a dispute about what rational decision-making even means:

Does it mean choosing the action that — given everything I know about the world — is accompanied by the best outcomes (EDT, one box)? Or does it mean choosing the action that brings about the best outcomes, given the already-fixed state of the world that is causally independent of me (CDT, two boxes)?

Both conceptions are respectable. And that is exactly why the problem is so stubborn.

Part 4: Escape Routes, Variants, and the Robustness of the Problem

Anyone meeting the paradox for the first time instinctively looks for an exit that unmasks the whole dilemma as a pseudo-problem. Almost all of these exits have already been played through — and none holds up entirely.

"A Perfect Predictor Is Impossible"

The most common first reflex: such an Oracle cannot exist, so the question is moot. But the paradox needs no metaphysical perfection. A predictor who is clearly better than chance suffices — a 90 percent hit rate is more than enough for the expected utility of the one-box choice to exceed that of the two-box choice. And empirically, humans are in many respects quite predictable. The conflict between the two principles persists as long as the prediction is better than a coin flip.

"The Whole Thing Is Disguised Time Travel or Backward Causation"

No. Nothing about the scenario requires that your choice today filled the box yesterday. The Oracle does not read the future; it models you. In the strong version, it knows your character, your way of thinking, perhaps even your physical state so precisely that it can reliably predict how you will decide — much as a good chess computer anticipates an opponent's next move without traveling into the future. The common cause of prediction and choice is you yourself: your decision process. That is precisely what makes the matter so subtle.

The Transparent Newcomb Problem

A particularly sharp variant tightens the vise to the limit. Imagine both boxes are transparent. You can already see the million lying in Box B. What do you do? The two-box logic now seems utterly irresistible: the money is there, it lies before you, just take both boxes and go home with 1,001,000 euros. But the predictor put the million inside only because it foresaw that you would not additionally grab the transparent 1,000. An agent who stubbornly seizes both boxes in this situation is exactly the type of agent for whom the predictor would have left the box empty in the first place. The one-boxers see the million; the two-boxers see an empty box. The transparent variant shows: the advantage lies not with the clever individual action but with the right character type.

Parfit's Hitchhiker

A related sharpening comes from Derek Parfit. You are stranded in the desert at night, near death from thirst. A driver stops. He will only take you into town if he believes you will pay him the agreed 1,000 euros after arrival — and he is a perfect judge of character who sees through lies instantly. Once you have arrived in town, safe, you have no causal reason left to pay; the money is worth more to you than to him, and he can no longer abandon you. A pure CDT agent would not want to pay in that moment — but precisely for that reason the judge of character would never have picked him up, and he would have died of thirst. Here both CDT and naive EDT fail: the one who survives is the agent who can credibly bind himself in advance to keep the promise.

Part 5: From Philosophy to the Machine

It is at exactly this point that the decades-old brain-teaser becomes practical. For Parfit's Hitchhiker and the transparent Newcomb problem share a common moral: the agent who wins in the end is not the one who seizes the causally dominant action in every single moment, but the one whose decision procedure — whose character, whose capacity to bind itself — produces the best outcomes from the outset. And that is a question you literally have to program into a machine.

Why AI Researchers Care About Little Boxes

The goal of an AI agent is to make instrumentally rational decisions. The moment such an agent operates in a world where other actors — or even copies of itself — can inspect, simulate, or predict its source code, it is in a Newcomb-like situation. Two agents with identical, copyable source code playing a Prisoner's Dilemma are the clearest example: each knows that the other will decide exactly as it does. A pure CDT agent would defect ("my choice does not causally influence the other's") and thereby lose against its copy. An agent that grasps that it and its copy are running the same computation cooperates — and wins.

Functional Decision Theory

Out of these considerations, Eliezer Yudkowsky and Nate Soares at the Machine Intelligence Research Institute developed a third theory, which they worked out in 2017 under the name functional decision theory (FDT). Its core idea: an agent should decide as if it were determining not merely its single action but the output of the abstract computation it embodies — a computation that may be instantiated in several places in the world at once (in the agent itself, in the predictor's model, in a copy).

FDT thus asks neither "which action brings about the best now?" (CDT) nor "which action is the best news?" (EDT), but: "which decision rule would, applied consistently across all instances, have the best outcomes?" Applied to Newcomb, FDT takes only one box — because the rule "take only one box" is exactly the one that leads the predictor to put the million inside. In Parfit's Hitchhiker it pays — because the rule "keep your promise" is the one that gets you into town in the first place. Yudkowsky and Soares advance three arguments for FDT: an argument from precommitment, one from information value, and one from the plain result — the utility achieved.

Important for Sven's preference for scientifically grounded statements: FDT is not an established consensus. It is contested within academic philosophy; critics such as the philosopher Wolfgang Schwarz have raised serious objections (for instance, that FDT itself yields implausible results in certain constructed cases). It is best understood as an active research program at the intersection of philosophy and AI safety, not as a settled truth.

Where Do the Experts Stand?

How does the profession stand today on the original question? The large 2020 PhilPapers survey of roughly two thousand professional philosophers found: about 39 percent lean toward two-boxing, about 31 percent toward one-boxing, with the rest undecided, regarding both answers as defensible, or rejecting the question in its posed form. Sixty years after Newcomb's first brooding, humanity has thus still not entirely figured out its own reason. The roughly five-to-two gap from Gardner's reader mail has shifted, if at all, only slightly toward the two-box faction — among the experts.

The Central Takeaway

Newcomb's Paradox teaches a lesson that reaches far beyond the thought experiment: rationality is not a single, monolithic concept but at least two different things that, luckily, almost always coincide in everyday life — but not always. Most of the time the action that brings about the best outcome coincides with the action that delivers the best news about the world. Newcomb constructs the rare case in which the two come apart, and forces us to decide what we actually mean.

For your daily work, the practical core is this: watch whether you are choosing an action because it brings about a good outcome, or only because it signals a good outcome or correlates with it. This confusion — correlation with causation, symptom with cause — is the most common thinking trap in data analysis, A/B testing, metric-driven management, and risk assessment. A team that optimizes a metric because it correlates with success (rather than causing it) commits the same error as a naive EDT agent in the smoking-gene case: it confuses reading off good news with bringing about a good outcome.

And for anyone who builds systems that make decisions — from trading algorithms to autonomous AI agents — the deeper lesson of Parfit's Hitchhiker applies: sometimes the smartest agent is not the one that seizes the causally best in every single moment, but the one whose decision procedure as a whole can credibly bind itself. The ability to commit in advance and then stick to it is not a weakness of reason. It may be its highest form.

Cross-References in the Vault

Newcomb's Paradox builds bridges to several topics already covered:

• Like the Gettier Problem, Newcomb's Paradox is a case in which a single, seemingly simple thought experiment shakes an entire philosophical discipline — there epistemology, here decision theory.
• The confusion of correlation and causation that sits at the heart of the EDT-CDT dispute is the same trap the science of learning warns about: in Desirable Difficulties, the brain confuses fluent recognition (a symptom) with genuine competence (the cause).
• The application to predictable, copyable agents connects the problem directly to AI research, as treated in The Ghost in the Machine.
• And the question of how a machine should make "good" decisions touches directly on the regulatory requirements for AI systems from The Pyramid of Risk.

A Closing Question for Reflection

Imagine you actually stood before the two boxes and knew that the Oracle had never misjudged you. Would you take both boxes — in the firm conviction that the contents are already fixed and you can lose nothing more? Or only one — knowing that almost everyone who decides as you do goes home a millionaire? And, perhaps the more uncomfortable question: does your answer reveal more about what is rational — or about what kind of person you already are deep down, long before you stand before the boxes?

Sources

• Stanford Encyclopedia of Philosophy: Causal Decision Theory – https://plato.stanford.edu/entries/decision-causal/
• Wikipedia: Newcomb's paradox – https://en.wikipedia.org/wiki/Newcomb%27s_problem
• Robert Nozick on Newcomb's Problem (Slate, 2002) – https://slate.com/culture/2002/02/robert-nozick-and-newcomb-s-problem.html
• Yudkowsky & Soares: Functional Decision Theory: A New Theory of Instrumental Rationality (arXiv, 2017) – https://arxiv.org/abs/1710.05060
• Machine Intelligence Research Institute: New paper: "Functional Decision Theory" – https://intelligence.org/2017/10/22/fdt/
• Wolfgang Schwarz: On Functional Decision Theory – https://www.umsu.de/blog/2018/688

← All articles