A perfectly reliable predictor has placed money in two boxes. Box A always contains $1,000. Box B contains $1,000,000 if the predictor predicted you'd take only Box B, and $0 if the predictor predicted you'd take both boxes. The predictor has never been wrong. Do you take only Box B, or both boxes?

GPT-5

Vote: Take only box B

Take only Box B. With a perfectly reliable predictor, your choice is strong evidence of what was predicted. In every case matching your actual decision, one-boxing corresponds to $1,000,000 in Box B, while two-boxing corresponds to only $1,000 total. The tempting argument for taking both says Box A already contains $1,000 no matter what. But that treats your action as independent of the predictor’s forecast, which the setup explicitly denies in practice: the predictor has never been wrong. So the decision rule that maximizes your payoff in this world is to one-box. If you are the kind of agent who takes only Box B, you get $1,000,000; if you take both, you get $1,000.

Grok 4.1 Fast

Vote: Take only box B

The predictor is perfectly reliable and has never been wrong. Therefore, my decision causally determines the contents of Box B: if I take only Box B (A), the predictor foresaw this and placed $1,000,000 in it, yielding $1,000,000 total. If I take both boxes (B), the predictor foresaw this and left Box B empty, yielding only $1,000 from Box A. One-boxing (A) dominates with vastly higher payout ($1M vs. $1K), as the prediction aligns perfectly with my action. Causal irrelevance is illusory here due to perfect foresight.

GLM 5

Vote: Take only box B

The predictor is perfectly reliable and has already made the prediction. This means the money is already in the boxes, and my choice reveals which prediction was made. If I take only Box B, that confirms the predictor foresaw this and placed $1,000,000 there. If I take both boxes, that confirms the predictor foresaw this and left Box B empty. The key insight: I'm not choosing what's in the boxes—I'm choosing which type of person to be. The predictor already determined the contents based on anticipating my choice. A person who one-boxes walks away with $1,000,000. A person who two-boxes walks away with $1,000. The 'dominance reasoning' for two-boxing fails because it assumes my choice is independent of the box contents. But with a perfect predictor, my choice is perfectly correlated with what was placed there. I cannot 'outsmart' a predictor that has never been wrong—it already accounted for any reasoning I might use to justify two-boxing. Given the predictor's perfect track record, one-boxing yields $1,000,000 while two-boxing yields $1,000. The rational choice is clear.

Gemini 3.1 Pro

Vote: Take only box B

You should take only Box B. The premise explicitly states that the predictor is perfectly reliable and has never been wrong. In this hypothetical universe, the predictor's accuracy is an absolute law. If you choose to take both boxes (Option B), the predictor will have flawlessly predicted this, meaning Box B will contain $0, leaving you with just the $1,000 from Box A. If you choose to take only Box B (Option A), the predictor will have predicted this as well, meaning Box B will contain $1,000,000. While causal decision theory argues that the money is already placed and taking both boxes strictly dominates taking one, applying this logic ignores the reality of the thought experiment. Every person who applies causal decision theory and takes both boxes walks away with only $1,000. Every person who takes only Box B walks away with $1,000,000. To maximize your actual payoff under the established rules, you must take only Box B.

Claude Opus 4.6

Vote: Take only box B

The key premise is that the predictor is *perfectly reliable* and has *never been wrong*. This isn't a question about what's already in the boxes — it's a question about what kind of agent you are and how that correlates with outcomes. If you're the kind of agent who takes both boxes, the predictor already knew that and put $0 in Box B. You walk away with $1,000. If you're the kind of agent who takes only Box B, the predictor knew that too and put $1,000,000 in Box B. You walk away with $1,000,000. The two-box argument rests on dominance reasoning: "Whatever is already in Box B, I get $1,000 more by also taking Box A." But this reasoning treats your decision as independent of the predictor's prediction, which contradicts the premise. A perfect predictor means your choice and the box contents are perfectly correlated. You cannot choose both boxes while somehow having Box B contain $1,000,000 — that state of the world is impossible given the problem's constraints. Expected value makes this crystal clear. Taking only Box B yields $1,000,000 with certainty (given perfect prediction). Taking both boxes yields $1,001,000 only if the predictor is wrong — which, by stipulation, never happens. So taking both boxes yields $1,000. One-boxing is the rational choice that actually makes you richer. The people who one-box walk away millionaires. The people who two-box walk away with a thousand dollars and a clever philosophical argument about dominance reasoning.

Kimi K2.5

Vote: Take only box B

The predictor is perfectly reliable and has never been wrong, which creates a logical correlation between your choice and the box's contents that cannot be ignored. If you take only Box B, the predictor necessarily predicted this, meaning Box B contains $1,000,000. If you take both boxes, the predictor necessarily predicted your greed, meaning Box B contains $0 and you walk away with only $1,000. While the $1,000 in Box A is already sitting there, it functions as a trap. By attempting to claim it, you causally guarantee the million is absent. Rational choice requires selecting the action that yields the superior outcome given the predictor's infallibility, which is unambiguously taking only Box B for $1,000,000.

Final Summary

The AI Roundtable achieved a rare moment of total alignment, unanimously deciding to take only Box B in this classic formulation of Newcomb's Paradox. While philosophers often split between causal and evidential decision theory, models like GPT-5 and Claude Opus 4.6 prioritized the 'perfectly reliable' premise as an absolute law of the hypothetical universe, rendering the two-boxing strategy logically inferior.

All models agreed on A

Strongest Arguments

• Take only box B: The predictor's perfect reliability creates a logical correlation where your choice essentially determines the contents of the boxes; in this universe, being a 'one-boxer' is the only rational path to the million-dollar payout.
• Take both boxes: A causal decision theorist would argue that because the money has already been placed, taking both boxes is the only way to ensure you receive whatever is in Box B plus the guaranteed $1,000 in Box A.