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.
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
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
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?
Why is accepting contextual, in a way that judging is not?
Merely saying that acceptances, "being tied to action"
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
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
The two-step theory sketched above accounts nicely for the cases of acceptances discussed in the literature. Judging P true flat-out is an accepting under a stringent norm of accurate truth, while "judging P likely" is an accepting under a norm of plausibility, conducted on the background of previous probabilistic beliefs regarding P. Adopting P as a matter of policy divides into accepting a set of premises to be used in collective reasoning under a norm of consensus, and accepting it under a norm of coherence, (as in judgments by contradiction, legal reasoning, etc.). Assuming, imagining, supposing do not automatically qualify as acceptances. Only their controlled epistemic forms do, in which case they can be identified as forms of premising.
This theory predicts that errors in acceptances can be either strategic or epistemic. Strategic errors occur when selecting an epistemic norm inappropriate to a context, (for example, trying to reconstruct a shopping list accurately, when comprehensiveness is sufficient), or when incorrectly setting the decision criterion given the stakes (taking an epistemic decision to be non-important when it objectively is, and reciprocally). Epistemic errors, in contrast, can occur both in applying a given norm to its selected material, (for example, seeming to remember that P when one does not) or in forming an incorrect judgment of confidence about one’s epistemic performance (for example, being confident in having correctly remembered that P when one actually failed to do so). Appropriate confidence judgments have an extremely important role as they help filter out a large proportion of first-order epistemic mistakes (illusory remembering, poor perceivings, incoherent or irrelevant reasonings etc.).
Is our two-tiered theory relevant to AI research? Although the author has no personal competence in this domain, it appears to be clearly the case. The two-tiered theory of acceptance is inspired by empirical research on epistemic self-evaluation, and by requirements on epistemic reliability. For these reasons, it should help epistemic agency to be modeled in a more realistic way, and conferred to artificial systems with the suitable cognitive complexity. Indeed an artificial agent, for example a social robot, should be able to monitor the epistemic responses of others as well as its own not only for their accuracy, but also for their comprehensiveness, their coherence, and the consensus in a group. Monitoring relevance in speech may, at present, be a trickier issue. Even artificial agents with no linguistic ability, as far as they need to evaluate whether they can solve a given problem, need to have various forms of acceptance available (e.g.: do they have the resources to create a path to a solution? Should they list, rather, all the solutions already used in similar contexts? How plausible is it that one of these solutions is applicable here and now? Epistemic planning depends on such acceptances being alternative options). Furthermore, it is crucial for artificial system designers to clearly distinguish an epistemic evaluation, which only depends on internal assessment of what is known or accessible, and a utility evaluation, which varies with the importance of the task.
Our two-tiered theory of acceptance raises potential objections that will be briefly examined.
It might be objected that, if acceptance can be governed by
epistemic norms as disparate as intelligibility, coherence,
consensus and accuracy, it should not be treated as a
natural kind. In other words, there is no feature common to the
various forms of acceptance, and for that very reason, the
concept of acceptance should be relinquished. To address
this objection, one needs to emphasize that normative
diversity in acceptances is observed in metacognitive
studies: agents, according to circumstances, opt for accuracy or
comprehensiveness, or use fluency as a quick, although
loose way, of assessing truthfulness
A second objection might find it implausible that ordinary agents have the required sophistication to manage acceptances as described, by selecting the kind of epistemic acceptance that is most profitable given a context of planning, by keeping track of the implicit or explicit payoffs for a particular option, and by setting on their basis their response criterion.
It must be acknowledged that agents do not have in
general the conceptual resources that would allow them to
identify the epistemic norm relevant to a context.
Acceptances, however, can be performed under a given norm
without this norm being represented explicitly. Agents
learn to associate implicitly a given norm with a given
cognitive task and context: their know-how is revealed in their
practical ability to monitor their acceptances along the
chosen normative dimension
Concerning decision making, robust evidence indicates
that the ability to re-experience an emotion from the recall
of an appropriate emotional event is crucial in integrating
the various values involved in an option
Finally, epistemologists might observe that such a variety
of epistemic standards pave the way for epistemic value
pluralism, i.e., the denial that truth is the only valuable goal
to pursue. Our variety of epistemic acceptings should
indeed be welcome by epistemic value pluralists, who claim
that coherence, or comprehensiveness, are epistemic goods
for their own sake
The proposed two-tiered theory of acceptance does not invite a relativistic view of epistemic norms, but rather combats it: Even though agents can accept propositions under various norms, there are rational constraints on norm selection. For example, it may be rational to look for accurate retrieval when making a medical decision, while looking for comprehensive retrieval when shopping. Once a norm has been identified as instrumentally justified, acceptance can only be successful if it is conducted under the epistemic requirements prescribed by its norm. Thus agents, whether natural or artificial, can build, through acceptances, a stable epistemic representation of the facts relevant to their action that is not contaminated by the strategic importance of the decisions to be taken on its basis.
Given the limited cognitive resources available to agents at any given time, it is rational for them to focus on the epistemic goals that will maximize their epistemic potential both in terms of correctness and utility. Our two-tiered theory of acceptance accounts for this consequence of bounded rationality. Our future work aims to clarify the informational basis on which the various epistemic norms operate. We also aim to study norm sensitivity in agents from different cultures, measured by their confidence judgments about particular tasks and performances. Finally, we will study how potential conflicts for a given acceptance between epistemic norms can be generated, and how they are overcome.
Thanks to Richard Carter for his linguistic revision and critical discussion of a prior version, and Anne Coubray, Martin Fortier and Mouhamadou el Hady Ba for useful feedback. This research has ben made possible by an ERC advanced grant, "Dividnorm" project # 269616, conducted at the Fondation Pierre-Gilles de Gennes pour la Recherche, Ecole Normale Supérieure, Paris.