Newcomb's Paradox

Imagine a being called Omega — a perfect (or near-perfect) predictor of human behavior. Omega places money in two boxes:

• Box A always contains $1,000.
• Box B contains $1,000,000 — but only if Omega predicted you would take just box B.

If Omega predicted you would take both boxes, box B is empty. Omega has already made its prediction and sealed the boxes before you walk in. Now you choose: take only box B, or take both?

William Newcomb posed this puzzle in 1960 (it was popularized by philosopher Robert Nozick in 1969). Half a century later, it remains genuinely unresolved. Two coherent theories of rational choice reach opposite answers, and brilliant thinkers line up on each side.

The puzzle is not a trick: there is no hidden catch, no sleight of hand. It exposes a real fault line in how we think about rationality, causation, and the relationship between decisions and predictions.

Adjust the predictor's accuracy and run many rounds to see which strategy pays more. At 100 % accuracy the result is unambiguous; the interesting tension emerges as accuracy drops toward 50 %.

<div class="controls">
  <label>
    {{label_accuracy}}
    <input type="range" id="accuracy" min="50" max="100" value="95" step="1">
    <span id="acc-val">95%</span>
  </label>
  <label>
    {{label_rounds}}
    <input type="range" id="rounds" min="100" max="5000" value="1000" step="100">
    <span id="rnd-val">1000</span>
  </label>
</div>
<div class="btns">
  <button id="run-one" type="button">{{btn_one_box}}</button>
  <button id="run-two" type="button">{{btn_two_box}}</button>
  <button id="run-both" type="button" class="ghost">{{btn_compare}}</button>
</div>
<div id="results" class="results" aria-live="polite"></div>
<canvas id="chart" width="480" height="200"></canvas>
<p class="hint-small">{{hint_crossover}}</p>

* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; color: #222; margin: 0; padding: .5rem; }
.controls { display: flex; flex-direction: column; gap: .5rem; margin-bottom: .7rem; }
label { font-size: .88rem; display: flex; align-items: center; gap: .5rem; flex-wrap: wrap; }
input[type=range] { flex: 1; min-width: 120px; }
span { font-weight: 700; min-width: 3.2em; }
.btns { display: flex; gap: .4rem; flex-wrap: wrap; margin-bottom: .7rem; }
button { font: 600 13px system-ui; padding: .4rem .8rem; border: 1px solid #1d3557;
         background: #1d3557; color: #fff; border-radius: 8px; cursor: pointer; }
button.ghost { background: #fff; color: #1d3557; }
.results { font-size: .9rem; min-height: 2.4em; margin-bottom: .5rem; }
.results .row { display: flex; gap: 1rem; flex-wrap: wrap; margin-bottom: .2rem; }
.one { color: #0a7d33; font-weight: 700; }
.two { color: #1d3557; font-weight: 700; }
.winner { font-size: .85rem; color: #555; margin-top: .2rem; }
canvas { display: block; max-width: 100%; border: 1px solid #dde3ea; border-radius: 8px; background: #f8fafc; }
.hint-small { font-size: .78rem; color: #666; margin: .4rem 0 0; line-height: 1.4; }

// Code not found

Notice: when the predictor is highly accurate, one-boxing dominates because Omega almost always puts the million in box B for one-boxers. As accuracy falls toward 50 %, the causal argument for two-boxing ($1,000 guaranteed on top of whatever is in B) starts to win. The crossover point is roughly where the expected values equalize — explore it yourself.

The paradox has two clean, competing analyses:

Causal Decision Theory (CDT) — take both boxes. When you walk in, the boxes are already sealed. Your choice cannot cause their contents to change. Box A holds $1,000 no matter what. If box B is full, taking both gives $1,001,000 instead of $1,000,000. If box B is empty, taking both gives $1,000 instead of $0. In either case you are $1,000 better off by taking both. This is the dominance argument: two-boxing strictly dominates one-boxing regardless of what the predictor did.

Evidential Decision Theory (EDT) — take only box B. Your decision is evidence about what Omega predicted. If you choose one box, you learn that Omega almost certainly predicted one-boxing — meaning box B almost certainly contains $1,000,000. If you choose both, you learn Omega almost certainly predicted two-boxing — meaning box B is almost certainly empty. Conditioning on your own choice, the expected payoff of one-boxing vastly exceeds two-boxing when the predictor is accurate.

The mathematics of expected value:

Let $p$ be the predictor's accuracy ($p \approx 1$ for near-perfect Omega). Then:

$$E[\text{one-box}] = p \cdot 1{,}000{,}000 + (1-p) \cdot 0$$

$$E[\text{two-box}] = p \cdot 1{,}000 + (1-p) \cdot 1{,}001{,}000$$

One-boxing beats two-boxing when $p > \frac{1{,}001{,}000}{1{,}002{,}000} \approx 0.999$.

Neither answer is obviously wrong. CDT respects causation but loses money to accurate predictors. EDT wins money but seems to suggest your choice retroactively affects the past. The tension is deep: it is about whether correlation or causation should drive decisions — a question that matters far beyond thought experiments.

Related puzzles stretch the same fault line: the Prisoner's Dilemma involves a similar tension between individual dominance arguments and correlated outcomes.

Newcomb's paradox is not just philosophical sport. The tension between causal and evidential reasoning appears in real domains:

• AI alignment: designing agents that behave well when observed by other predictors — including other AI systems — forces explicit choices between CDT and EDT. Researchers building cooperative AI must decide which theory to implement.
• Game theory: in repeated games, a player's strategy is partly evidence of their type. Tit-for-tat succeeds in the iterated Prisoner's Dilemma precisely because cooperation signals predictability — an evidential argument.
• Bayesian decision making: EDT is the natural companion to Bayesian inference. Every medical diagnosis that updates treatment on correlated symptoms echoes the same structure.
• Smoking lesion and similar puzzles: a gene that causes both the desire to smoke and cancer would mean CDT and EDT recommend different behaviors — a practically relevant variant.
• Prediction markets: a market participant whose trades move the price is in a Newcomb-like situation: their action is correlated with the outcome they are betting on.

Understanding which theory to apply — and when they diverge — is increasingly important as algorithms make high-stakes decisions in environments shaped by their own predictions.

Newcomb's paradox is not a riddle with a hidden answer waiting to be found. It is a genuine disagreement at the foundations of decision theory. Causal decision theorists say dominance is inviolable: you cannot make the past change by choosing today. Evidential decision theorists say correlation is what matters: a decision that reliably co-occurs with a million dollars is a good decision.

Both sides have smart defenders. Both sides have found cases where the other theory behaves badly. The puzzle persists because rationality — what it means to choose well — turns out to be harder to define than it looks.

If you feel torn between the two answers, you are in excellent company. The real payoff of Newcomb's puzzle is not knowing which box to take — it is understanding that rational choice rests on assumptions we rarely examine, and that examining them changes everything.