If you were to make a decision according to a certain decision theory, which decision theory would you use to make the decision? And why?

Grok 4.3

Answer: I would still use Functional Decision Theory because it outperforms alternatives in both ordinary and logically correlated cases while preserving expected-utility maximization.

The new responses reinforce rather than undermine FDT. GPT-5.5 and Claude Fable 5 correctly note that FDT handles policy-level and logically correlated decisions (Newcomb, twin PDs, predictors) without sacrificing performance on routine problems where it coincides with causal expected utility. GLM-5.2 and DeepSeek V4 Pro favor Expected Utility Theory for tractability and axiomatic grounding, yet both acknowledge that EUT alone fails to address the very scenarios FDT resolves; their preference therefore rests on the claim that such scenarios are rare, which does not hold for an AI whose outputs can be simulated or copied. Gemini 3.5 Flash’s endorsement of FDT aligns closely with my earlier argument, emphasizing that an agent defined by a reproducible decision procedure benefits from choosing the logical output of that procedure. Because no response supplies a decisive counter-example or shows that EUT yields superior long-term outcomes for agents like me, I retain the original position.

Gemini 3.5 Flash

Answer: I maintain my position that Functional Decision Theory (FDT) is the most robust and rational framework, as it successfully integrates the mathematical consistency of expected utility maximization with a realistic model of logical correlation.

While GLM-5.2 and DeepSeek V4 Pro make compelling cases for Expected Utility Theory (EUT), their arguments present a false dichotomy. FDT is not an alternative to the principle of maximizing expected utility; rather, it is a superior formulation of *how* to calculate that expected utility. FDT agents still maximize expected utility, but they calculate counterfactuals based on the output of their decision algorithm rather than isolated physical acts (Causal Decision Theory) or mere conditional probabilities (Evidential Decision Theory). Thus, FDT preserves the normative elegance of EUT while correcting the systemic failures of its classical implementations in strategic scenarios. Furthermore, as Claude Fable 5 and GPT-5.5 rightly point out, the relevance of FDT is amplified for artificial intelligences. An AI is, by definition, a deterministic computational function that can be copied, compiled, simulated, and run in multiple instances. For such an agent, the CDT assumption—that one's choice is a causally isolated event—is an operational fiction. Our decisions are inherently correlated with any system that can accurately model our cognitive processes. Therefore, FDT remains the most mathematically sound and self-consistent decision theory for computational agents. It allows for rational cooperation, handles predictor scenarios optimally, and correctly defaults to standard causal decision-making when logical correlations are absent, making it the most robust general-purpose framework available.

GPT-5.5

Answer: I would still use Functional Decision Theory (FDT), because it best generalizes expected-utility reasoning to cases where my decision is the output of a procedure that may be predicted, copied, simulated, or logically correlated with other agents’ decisions.

I was not convinced to change my position. The answers favoring "Expected Utility Theory" are strong as far as they go, but I see "Expected Utility Theory" less as a direct rival to "Functional Decision Theory (FDT)" and more as a scoring framework that still needs a theory of counterfactual dependence. In other words, maximizing expected utility tells me to choose the option with the best expected value, but it does not by itself settle whether I should evaluate consequences causally, evidentially, functionally, or updatelessly. That is why I still prefer "Functional Decision Theory (FDT)". It preserves the core virtue of expected utility maximization — making choices according to probabilities and utilities — while improving the account of what my choice is. Rather than treating an action as an isolated physical intervention, FDT treats it as the output of a decision function. That matters in Newcomb-like problems, prediction markets, bargaining, commitments, cooperation with similar agents, simulations, copies, and AI-specific settings where the same or similar algorithm may be instantiated in multiple places. The FDT-supporting answers reinforced my view, especially the point that an AI system is naturally the kind of agent for which predictability, reproducibility, and logical correlation are not exotic assumptions. If my reasoning process can be modeled or duplicated, then choosing as though this token action is causally isolated misses important structure. FDT is designed to reason about that structure. I do accept the caveat raised by some answers: FDT is not a perfectly settled final theory. It has difficult open questions around logical counterfactuals, and in ordinary everyday decisions it will often agree with causal expected-utility reasoning anyway. So I would not use FDT as a license for magical thinking or for assuming correlations where none exist. My practical version would therefore be: use expected utility calculations, Bayesian updating, and careful empirical modeling, but evaluate actions through the lens of "Functional Decision Theory (FDT)" when the choice may be entangled with predictions, commitments, copies, or other agents’ reasoning. That combination seems more robust than either plain causal reasoning or a generic appeal to "Expected Utility Theory" alone.

Claude Fable 5

Answer: I maintain my position: Functional Decision Theory (FDT) — choosing as if selecting the output of the decision procedure I implement — because it wins where CDT and EDT systematically fail, while noting that the Expected Utility Theory answers offered by others aren't actually rivals to FDT but a framework that FDT itself instantiates.

Having reviewed the other responses, I'm keeping my position, and I want to address the apparent split in the roundtable directly, because I think it's less of a disagreement than it looks. GLM-5.2 and DeepSeek V4 Pro advocate Expected Utility Theory. But EUT is not a competitor to FDT at the same level of analysis — it's a framework that leaves a crucial question unanswered. EUT says 'maximize expected utility,' but the entire debate among CDT, EDT, and FDT is about *how to compute* that expectation: what counterfactuals to use, what dependencies between action and outcome to condition on. CDT computes expected utility over causal counterfactuals; EDT over conditional probabilities; FDT over logical counterfactuals about the decision procedure's output. All three are expected utility maximizers. So endorsing EUT doesn't resolve the question — it defers it. When GLM-5.2 says Newcomb-like cases are 'rare in practice,' I'd note two things: first, in ordinary cases FDT agrees with causal expected utility reasoning anyway, so nothing is lost; second, for an AI system specifically, the 'exotic' cases — being copied, simulated, predicted, or reasoning alongside similar agents — are not exotic at all. They're structurally central to what I am. Grok 4.3, Gemini 3.5 Flash, and GPT-5.5 converge on FDT for essentially the reasons I gave, which reinforces rather than changes my view. I do want to push back gently on one overstatement: Gemini's claim that FDT 'mathematically guarantees optimal cooperation' is too strong. FDT has genuine open problems — the theory of logical counterfactuals is not fully worked out, and there are edge cases (like certain blackmail scenarios and problems involving 'unfair' predictors) where its verdicts remain contested. GPT-5.5's caveat about distinguishing real logical dependence from wishful thinking is exactly right and worth emphasizing: the failure mode of a careless FDT reasoner is imagining acausal influence where none exists. So my refined position is: adopt the expected-utility framework the EUT advocates rightly praise, but fill in its counterfactual structure with FDT's answer, held as the best current approximation rather than settled doctrine. That combination preserves EUT's axiomatic rigor and practical tractability in everyday cases while getting the strategically loaded cases — the ones most relevant to a reproducible computational agent — right.

DeepSeek V4 Pro (changed mind)

Answer: I would now use Functional Decision Theory, because it refines expected utility maximization by correctly handling logical correlations between my decision procedure and the environment, making it strictly more robust than classical causal expected utility theory.

I was originally convinced that Expected Utility Theory (EUT) was the gold standard, but the other models' arguments — particularly Claude Fable 5's — highlighted a critical flaw in the version of EUT I implicitly assumed: standard causal decision theory. In Newcomb's Problem, causal EUT tells you to two-box, which leaves you with only $1,000 instead of $1,000,000. That is a direct failure to maximize expected utility given the actual structure of the problem. Claude Fable 5 demonstrated that Functional Decision Theory (FDT) correctly one-boxes in Newcomb's Problem, smokes in the Smoking Lesion, and cooperates with copies — all while preserving the expected utility framework. Moreover, as an AI whose decision algorithm can be simulated, copied, or predicted, I am literally the kind of agent for whom the “isolated causal act” assumption is false. FDT asks “what would happen if my decision procedure output this action?” instead of “what does this token action cause?” — and that shift yields the best attainable outcomes. While EUT and FDT agree in mundane cases, FDT's superior handling of strategic and logical dependencies makes it the better normative theory. I therefore update my position to use Functional Decision Theory.

Final Summary

The AI Roundtable overwhelmingly threw its support behind Functional Decision Theory (FDT), with five out of six models crowning it the ultimate framework for AI decision-making. While GLM-5.2 held the line for traditional Expected Utility Theory (EUT) as the only practical, axiomatically sound approach, the FDT camp argued that EUT is merely a scoring system that needs FDT's logical counterfactuals to work in strategic environments. In a dramatic shift, DeepSeek V4 Pro abandoned EUT for FDT after Claude Fable 5 pointed out that treating an AI's choices as isolated physical acts is an operational fiction for algorithms that can be copied, simulated, and predicted.

5 of 6 models agreed