This competition entry has been selected for publication by the Forum team.
A challenge to Premise 1 of the unawareness argument
Feedback welcome in the comments during event week.
Anthony DiGiovanni states the normative premise this way: to justify c-preferring A over B, we need to argue that A has higher expected value broadly speaking, meaning that in some sense we "expect" that, if we were idealized agents who could aggregate all of A's and B's possible consequences into literal expected values, we would say A has higher expected value. We ourselves do not need literal expected values (DiGiovanni 2026).
So P1 is a claim about the verdict of a better-informed evaluator. It also lets us reach that verdict indirectly.
The counterargument table in the summary post labels one row as not directly addressed by the sequence: the objection that incomparability violates plausible deference and reflection principles, citing Hare (2010) and Tarsney, Lederman and Spears. DiGiovanni replies that those principles do not tell us about A's or B's "expected" consequences themselves, in our actual epistemic state, so they do not answer P1's question.
That reply succeeds against the deference arguments now in print. It does not reach a second kind of deference.
Hare (2010) considers an agent who lacks a preference between two options while preferring a mildly sweetened version of each. He argues that the resulting pattern of choice is defective. Tarsney, Lederman and Spears derive an impossibility from a weak negative dominance principle together with a weak form of ex ante Pareto. Both arguments concern the formal properties of a ranking. Call this structural deference. DiGiovanni's reply works against it: showing that incomparability conflicts with an attractive axiom does not tell us which option has better impartial consequences.
A second relation is different. An agent learns what a better-informed evaluator concluded about A and B after examining evidence the agent cannot inspect. The evaluator uses the same impartial welfare criterion. Her judgment is about the consequences themselves. Call this evidential deference.
I must choose between interventions A and B. A researcher has examined evidence I cannot inspect. I have independent grounds for treating her as better informed about this comparison. She reports that A has better expected consequences than B.
I cannot reproduce her analysis. Her report still changes my evidence about A and B, as a physician's advice changes a patient's evidence about two treatments.
Evidential Deference. When an agent has adequate grounds for treating another epistemic state as better informed about the impartial consequences of A and B, evidence that this state c-prefers A can supply a consequence-based reason to prefer A, absent a defeater specific to that deference relation.
Read strongly, DiGiovanni's sentence says that deference supplies no evidence about the comparative consequences. The sealed analysis gives reason to reject that.
Read weakly, it says that deference supplies evidence without supplying the first-order considerations behind the evaluator's judgment. That is correct. It does not yield the conclusion, because evidence need not contain the reasoning that produced it.
P1 is framed through an idealized evaluator. Let H be the proposition that an epistemically idealized version of me would assign A greater expected value than B. P1 asks for justified grounds for expecting H. Evidence about the verdict of a better-informed evaluator bears on H.
Excluding it needs a further requirement: that an impartial altruistic justification be reconstructible inside the agent's own model of the consequences. That is stronger than the wording of P1. It has not been defended. Asking again which option has higher expected value uses P1 to defend P1.
Molinari (2023) shows that deference does not require a precise prior. Good's theorem on the value of evidence fails for agents with imprecise credences. Molinari introduces two deference principles under which rational imprecise agents defer to their informed selves.
The case I need is weaker. Let F be a set of better-informed states the agent has adequate grounds to treat as superior for this comparison. Suppose every state in F c-prefers A to B. The agent need not assign a probability to any member of F. The agreement itself is evidence favoring A. When members of F disagree, the condition gives no ordering.
One tempting formalization does not help. If A is better than B conditional on every cell of an information partition, then A is better than B unconditionally under every measure in the credal set. That result presupposes a fixed state space. Unawareness is the claim that successor states involve hypotheses absent from the present space, so the theorem proves nothing here.
The unanimity condition instead mirrors the sequence's own criterion. Section 3.1.1 permits a comparison when one strategy is better under every plausible way of precisely evaluating coarse outcomes. Unanimity across better-informed states quantifies the same way over a different set. Precisifications of a present model are limited to hypotheses the agent has already conceived, which is why section 3.2.1 can argue that modeling the catch-all is subject to biased sampling. Successor states face no such limit.
If some successors use hypotheses I cannot presently conceive, how could I know that all of them rank A above B?
For one class of comparisons I do not need to know what they will believe. Policy P acquires information at a first stage, then chooses from the option set containing A and B. Policy Q commits to A now. Every successor under P chooses whichever option it then ranks highest. A remains available at that node, so every successor weakly c-prefers its own choice to A by its own lights. The unanimity comes from the structure of the option set. Preference is strict in any state whose ranking differs from mine.
Two restrictions apply. The information stage must not remove options. It must not have consequences of its own large enough to swamp the comparison.
Sections 4.1.5 and 4.1.6 discuss focusing on lock-in and building capacity. DiGiovanni concludes that we cannot control our impact on such events precisely enough, nor tease apart their relative value when we picture them coarsely. Those replies target the substantive consequences of particular punting interventions. They do not establish that option value supplies no c-reason at all.
I cannot know that every successor favors A unless I already know why. I can know the sign of a trusted expert's conclusion without possessing her evidence. The argument needs justified trust in the successor, not access to the successor's reasons.
Successors may be biased. Then the antecedent fails for those states and the condition gives no verdict. Admissibility is meant to be demanding.
Every successor is itself subject to the unawareness argument. DiGiovanni has granted what defeats this. Asked whether humans remain clueless in all situations, he replied that it is plausible that with much more evidence and more developed conceptual models of the cosmos we would be non-clueless about the comparisons of some actions, while adding that reaching such a state seems very hard. Once non-clueless states are possible, F can be non-empty. Which inquiries reach them is a question for P3.
An unknown defeater could undermine my trust. All evidence is defeasible that way. If the bare possibility of an unknown defeater erased the evidential force of testimony, ordinary epistemic deference would go with it.
Evidential deference restores little. It is silent whenever admissible states disagree. It does not show that familiar comparisons between interventions survive severe unawareness.
It is enough for the target. The conclusion of the unawareness argument is universal: we are not justified in c-preferring any action over any other. One surviving class of justified comparisons defeats it. Section 6 identifies such a class, whose unanimity does not depend on anticipating hypotheses we lack.
Whether warranted deference can ever ground a c-preference is a normative question about P1. Whether any present agent has adequate grounds to defer on a high-stakes altruistic choice is an empirical question. A negative answer to the second is compatible with a positive answer to the first.
The sequence is persuasive that invented precision is a poor response to unawareness. Nothing above reintroduces it. P1 therefore needs revision: a justified c-preference can rest either on a direct comparison of expected consequences or on warranted higher-order evidence about how better-informed evaluators compare them. With that second route admitted, impartial altruism remains partially action-guiding under unawareness without any return to precise Bayesianism.
DiGiovanni, Anthony. 2025. The challenge of unawareness for impartial altruist action guidance. EA Forum sequence, 2 June 2025. forum.effectivealtruism.org/s/rHqdsrieinyhM5KDv
DiGiovanni, Anthony. 2026. Cluelessness: Summary of the argument, why it matters, and counterarguments. EA Forum, 19 June 2026. forum.effectivealtruism.org/posts/NesgdEY6yPrE9wDap
DiGiovanni, Anthony. 2026. Comment in reply to K. Kirdan, under the post above. EA Forum. forum.effectivealtruism.org/posts/NesgdEY6yPrE9wDap?commentId=scZpLRcicQpM6jB9T
Hare, Caspar. 2010. "Take the sugar." Analysis 70 (2): 237-247. doi.org/10.1093/analys/anp174
Molinari, Giacomo. 2023. "Trust the evidence: two deference principles for imprecise probabilities." Proceedings of Machine Learning Research 215: 356-366. proceedings.mlr.press/v215/molinari23a.html
Tarsney, Christian, Harvey Lederman and Dean Spears. Share the Sugar. PhilArchive. philarchive.org/rec/TARSTS
