A Two-tiered View on Acceptance

Experimental studies in metacognition indicate that a variety of norms are used by humans and some non-human agents to control and monitor their cognitive performances, such as accuracy, comprehensiveness, intelligibility, coherence, relevance, or consensus. This diversity of epistemic norms motivates a revision of the concept of acceptance. First, there are different forms of acceptance, corresponding to the specific epistemic norm(s) that constitute(s) them. Furthermore, acceptances need to include a strategic component, from which the epistemic component is insulated, whose function is to adjust the epistemic output to expected utility. Experimental evidence suggests that this twotiered analysis of acceptance is empirically adequate. Relevance to AI is briefly discussed. Acceptance and its Norms Intelligent agency requires an ability to control and monitor one's cognitive states, e.g. retrieve memories, check one's perceptions or one's utterances. The aim of cognitive control is to acquire cognitively reliable properties, such as retrieving a correct answer. Intelligent agents in realistic settings, however, whether natural or artificial, need to track other epistemic norms beyond accuracy, such as the comprehensiveness of a list, the intelligibility of a text, the coherence of a story, the relevance of a remark, or the consensuality of a claim. Experimental studies in metacognition suggest that such norms are indeed used by human and some non-human agents to control and monitor their own cognitive performance (Goldsmith and Koriat, 2008, Couchman et al. 2010). Furthermore, the particular cognitive task in which performance is being monitored has been shown to dictate which epistemic norm is appropriate to a given context. The goal of this article is to sketch a theory of acceptance that takes advantage of these studies. Acceptances, in contrast with beliefs, are generally recognized as voluntary (Jeffrey 1956, Stalnaker 1987, Bratman 1999, Lehrer 2000, Velleman 2000, Frankish, 2004). Accepting is an epistemic action, involving deliberation, i.e. various forms of cognitive control and their associated norms. There is no consensus, however, about the norm(s) of acceptances. While for Velleman (2000) accepting is regarding a proposition P as true, even though it may not be "really true", Cohen takes acceptance to be "a policy for reasoning, (..) the policy of taking it as a premise that P" (1992, 5, 7). For Stalnaker, "sometimes it is reasonable to accept something that one knows or believes to be false". Circumstances where this is reasonable include cases where P "may greatly simplify an inquiry", where P is "close to the truth", or "as close as one needs to get for the purposes at hand". Granting that accepted propositions are subject to contextual variation in their sensitivity to evidence and truth, they cannot be freely agglomerated in a coherence-preserving way, in contrast with beliefs (Stalnaker 1987). Finally, Bratman (1999) claims that acceptances conjoin epistemic and practical goals. These features of acceptance, however, fail to offer an intelligible and coherent picture of the epistemic action of accepting, and of its role in practical reasoning and decision-making. First, it is left unclear how a context of acceptance is to be construed in a way that justifies applying fluctuating epistemic standards. Second, how can one possibly conjoin an epistemic requirement, which is essentially passively recognized and applied, and utility considerations, which require an active decision from the agent as to what ought to be accepted in the circumstances? The Context Relevant to Accepting P Why is accepting contextual, in a way that judging is not? Merely saying that acceptances, "being tied to action" (Bratman, 1999), are sensitive to practical reasoning, is not a viable explanation: other mental actions, such as judgments, also tied to action, do not adjust their contents to considerations of practical reasoning. Saying that they are context-dependent because coherence, consistency, and relevance apply within the confines of the existing plan, rather to a theoretical domain, does not explain how epistemic correctness survives instrumental adequacy. Our first proposal consists in the following claim: Utility may dictate the norm of acceptance relevant to a given context of action, without dictating the output of the corresponding acceptance. As said above, accepting P can be driven, among other things, by a goal of comprehensiveness, accuracy, intelligibility, consensus or coherence. For example, you may accept that the shopping list you just reconstructed from memory is comprehensive (all the items previously listed are included), but not that it is accurate (new items are also mistakenly listed). On the proposed view, an acceptance is always indexed to its specific norm: a proposition is never merely accepted, it is rather acceptedat or acceptedct etc. (where at is short for: accurate truth, and ct for comprehensive truth). Although the selection of a particular epistemic goal responds to the practical features of one's plan, there is no compromise between epistemic and instrumental norms concerning the content of acceptances. Agents' epistemic confidence in acceptingn P (accepting P under norm n) is not influenced by the cost or benefit associated with being wrong or right. Thus we don't need to endorse the view that an epistemic acceptance of P is yielding to utility considerations, as Bratman suggests. This proposal offers a natural way out of a puzzle, called the preface paradox, which is raised by traditional views about acceptance: A writer may rationally accept that each statement in his book is true, while at the same time rationally accepting that his book contains at least one error (Makinson 1965). This puzzle is dissolved once it is realized that the author's epistemic goal is one of offering an ideally comprehensive presentation of his subject matter: it will thus not be contradictory for him to acceptct all the sentences in her book, while acceptingpl (accepting as plausible or likely) that one of them is false. Hence, a mental act of acceptancect does not allow aggregation of truth, because its aim is exhaustive (include all the relevant truths) rather than accurate truth (include only truths). Similarly, in the lottery puzzle, an agent acceptsat that there is one winning ticket in the one thousand tickets actually sold. It is rational for her, however, not to acceptpl that the single ticket she is disposed to buy is the winning one. From Epistemic to Strategic Acceptance The output of an epistemic acceptance so construed needs, however, to be adjusted to the final ends of the agent’s plan. The decision to act on one's epistemic acceptance, i.e., strategic acceptance, constitutes a second, distinct step in accepting P. On our view, utility does not just influence the selection of certain epistemic norms of acceptance. It also influences decision in a way that may depart greatly from the cognitive output of epistemic acceptance. The first argument in favor of this two-step view of acceptance is conceptual. The existence of an autonomous level of epistemic acceptance enables agents to have a stable epistemic map that is independent from local and unstable instrumental considerations. Thus, it is functionally adaptive to prevent the contents of epistemic evaluation from being affected by utility and risk. Second, empirical evidence shows that agents are indeed able to adjust their cognitive control both as a function of their confidence in accepting P, and of the strategic importance of the decision to be finally taken. In situations where agents are forced to conduct a cognitive task, strategic acceptance is ruled out: agents merely express their epistemic acceptance. In contrast, when agents can freely consider how to plan their action, given its stakes, they can refrain from acting on the unique basis of their epistemic acceptance. A decision mechanism is used to compare the probability for their acceptance being correct and a preset response criterion probability, based on the implicit or explicit payoffs for this particular decision. Here agents are allowed to strategically withhold or volunteer an answer according to their personal control policy (risk-aversive or risk-seeking), associated with the cost or benefit of being respectively wrong or right (Koriat and Goldsmith, 1996). A third reason in favor of our two-tiered view is that strategic acceptance can be impaired in patients with schizophrenia, while epistemic acceptance is not (Koren et al. 2006): this suggests, again, that epistemic and strategic acceptances are cognitively distinct steps.