Epistemic Updates in Transformative Experiences — Epoche C1
The bedrock of contemporary epistemology often rests on the premise that rational agents update their degrees of belief, or 'credences', through a process of conditionalisation — a fixed arithmetical rule for revising those degrees in the light of new evidence — even when confronting novel experiences. This Bayesian framework, with its elegant mathematical precision, posits that new evidence simply refines prior probabilities, leading to a more accurate posterior belief. However, the model faces profound challenges when confronted with phenomena such as L. A. Paul's concept of 'transformative experience'. Paul (2014) argues that certain experiences — becoming a parent, undergoing a significant career change, or even tasting a durian for the very first time — are not merely additive; they fundamentally alter one's preferences, values, and even one's very identity. The epistemological implications of such experiences compel us to question the universality of standard Bayesian updates. To feel the force of the challenge, one must first see clearly what the Bayesian model claims and why it is held to be the standard of rationality; only then does it become visible exactly which of its load-bearing assumptions transformative experience knocks away. The Bayesian Baseline: What Conditionalisation Says, and Why Bayesian epistemology represents an agent's state of belief not as a list of things believed outright but as a set of credences : numerical degrees of confidence between $0$ and $1$ assigned to propositions, where a proposition is simply a claim that could be true or false ('it will rain tomorrow', 'this durian tastes of caramel and onions'). Rationality is then held to impose two kinds of constraint. The first is synchronic — at any one time, credences should obey the axioms of probability (for instance, credences in a claim and its negation should sum to $1$). The second is diachronic , governing change over time, and its canonical rule is conditionalisation : when you learn a piece of evidence $E$ with certainty and nothing more, your new credence in any hypothesis $H$ should equal your old conditional credence in $H$ given $E$ — the confidence you already had in $H$ on the supposition that $E$ was true: $$ C_{\text{new}}(H) = C_{\text{old}}(H \mid E) = \frac{C_{\text{old}}(H \wedge E)}{C_{\text{old}}(E)}, $$ defined whenever $C_{\text{old}}(E)$ is positive. The rule has a simple mechanical meaning: learning $E$ deletes the possibilities incompatible with $E$ and rescales the credence in those that remain, so that the surviving possibilities keep their old proportions. In the standard jargon, the credences held before the evidence are the prior , and those held after are the posterior . This rule is not an aesthetic preference; it has celebrated justifications, which is precisely why a genuine counterexample matters. The classic 'Dutch book' theorems show that an agent whose updating policy departs from conditionalisation can be offered a series of bets, each fair by her own lights at the time she accepts it, that jointly guarantee her a loss — a pragmatic incoherence. A more recent and purely epistemic justification measures the inaccuracy of a credence as its distance from the truth (a natural such measure, the Brier score, charges $(1 - C(X))^2$ when $X$ is true and $C(X)^2$ when $X$ is false, so confident error is punished heavily). Greaves and Wallace (2006) proved that, judged by expected accuracy, conditionalisation is the optimal updating plan: any agent who plans to respond to evidence in some other way expects, by her own prior lights, to end up less accurate. Pettigrew (2016) develops this accuracy-first programme systematically, deriving the probability axioms and the conditionalisation rule from the single ideal of getting as close to the truth as possible. On the practical side, credences feed decision theory: the orthodox rule says to choose the act $a$ that maximises expected utility, $\mathrm{EU}(a) = \sum_{o} C(o \mid a)\, U(o)$, the sum over possible outcomes $o$ of their probability given the act, weighted by their utility $U(o)$ — a number representing how much the agent values that outcome. Note the quiet assumptions doing the work: a fixed stock of propositions over which credences are spread, and utilities that are available in advance and stable across the decision. Both will fail. Paul's Challenge: Epistemic Opacity Paul distinguishes two properties an experience can have. It is epistemically transformative if undergoing it gives you knowledge you could not have acquired any other way — before it, you cannot know what it will be like. It is personally transformative if it changes your core preferences and values — the very utilities the decision rule needs. The experiences that interest her, and this essay, have both properties at once; her book opens with a deliberately fantastical case, the offer to become a vampire, precisely because it makes both features vivid: you cannot know in advance what vampiric life is like from the inside, and becoming one would change what you care about. The core of the argument is that before a genuinely transformative experience, one cannot assign meaningful subjective probabilities or utilities to the outcomes. Consider the decision to become a parent. From a Southeast Asian perspective, where family and community ties are paramount, the social pressure and cultural expectations surrounding parenthood are immense. Yet the profound shift in one's perception of love, responsibility and self that accompanies becoming a parent cannot be adequately predicted by someone who has not lived it. It is not merely a matter of accumulating more information; it is a fundamental re-calibration of one's entire epistemic and axiological framework — 'axiological' meaning the framework of values itself, what one takes to be worth pursuing. The process by which such experiences break the Bayesian machinery can be delineated in three steps, each