Despite the increasing relevance of multiple prior beliefs in various domains of economics and beyond and the significant theoretical work on them, little progress has been made on developing choice-based techniques for eliciting them. This paper proposes a new choice-based, incentive-compatible elicitation method for multiple prior beliefs, and implements it in two experiments on continuous sources of uncertainty to elicit the multiple prior equivalent of subjects’ CDFs. The method is theoretically robust, insofar as it applies under a wide range of multiple prior decision models and with few assumptions about the nature of beliefs. In its implementation, we find a significant majority of subjects have non-degenerate sets of priors, with larger sets for more unpredictable events. Finally, we use our method to provide the first elicitation of the mixture parameter in the Hurwicz alpha-maxmin EU model that fully controls for beliefs.

Scientific models structure our perception of reality. This paper studies how we choose among them under expert advice. Scientific models are formalised as probability distributions over possible scenarios. An expert is assumed to know the most likely model and seeks to communicate it to a decision maker, but cannot prove it. As a result, communication about models is a cheap talk game. The decision maker is in a situation of model-uncertainty and is ambiguity sensitive. I show that information transmission depends on the strategic misalignment of players and, unlike similar models in the literature, a form of consensus among scientific models. When science is divided, there is an asymmetry in information transmission when the receiver has maxmin expected utility preferences. No information can be conveyed about models above a certain threshold. All equilibria of the game are outcome equivalent to a partitional equilibrium and the most informative one is interim Pareto dominant. 

This paper examines whether teachers’ gender biases stem from discrimination, and focuses on two of its potential drivers: gender preferences and beliefs. In an online experiment, 1,840 teachers evaluated fictitious transcripts with randomized gender information. Preferences were measured via dictator games, and beliefs through an Implicit Association Test. While teachers showed no gender preferences, they did hold gender beliefs. Analyzing 19,000 transcript evaluations, we find no evidence of gender-based discrimination. Our findings suggest that simply disclosing a student's gender does not trigger biased evaluations, implying that discrimination is more likely to emerge during direct student-teacher interactions.

This paper studies communication failure regarding risks in the commons. It argues that strategic issues might explain why years of scientific reporting regarding future risks have failed to trigger sufficient climate awareness. We consider a game of contribution to a public bad, where there is uncertainty regarding the damage generated by externalities. Prior to the game, agents receive non-certifiable information regarding the damage from an informed utilitarian expert. We show that in large-scale public good problems information transmission usually fails. We compare this result with the cases of Rawlsian and anti-Rawlsian experts and discuss the implication for climate expert panels. We also investigate the influence of the value of information in our result, which provides us with some insight on the part played by deontology on scientific communication.

Despite years of scientific reporting regarding climate change, public acceptance of economic regulations is still limited. Why scientific forecasts regarding climate change fail to trigger public willingness for the appropriate mitigation is still a highly debated question. This paper surveys the economic literature on environmental information transmission in search of an explanatory mechanism for this paradox. Combining empirical results with existing theoretical mechanisms, I argue that failure in information transmission naturally arises from the strategic setting in which scientific authorities and citizens find themselves in.

We consider a two tiered wholesale electricity market made of a day-ahead and a real-time market. In the retail market, consumers can subscribe a contract with a conventional retailer or cooperate through an aggregator who directly participates in the wholesale electricity market by taking forward positions. These latter depend on the consumer aggregated demand as estimated in the day ahead. Consumers are then penalized in real time on the basis of their prediction errors. To plan the aggregator’s pricing strategy, we model the consumers in a behavioral economics framework and take into account their possibility to churn. We characterize analytically the core of the game and give conditions for its non emptiness. Then we propose an algorithm based on Machine Learning methods (SVR, Neural Network, Regret) to optimize the aggregator’s pricing strategy in a competitive framework. Our results are finally evaluated on a case study based on time series of the power consumptions of 370 Portuguese consumers.

Can a hearer be rationally justified to have beliefs based on testimony alone when the source of his information is known to have conflicting epistemic goals? Building on a game-theoretical approach I suggest that, contrary to the existing views, he can. But this justification relies on an equilibrium concept, which is only reached on the long run. In addition, the hearer's justified beliefs will always be more imprecise than the one held by the original source. These results highlight the importance of scientific norms which, in practice, are the embodiment of these equilibrium mechanisms and thus of scientific credibility. 

Whether it is permissible to form credences about individuals on the basis of accurate statistics is a central question in epistemology. Existing approaches share a common structure: they treat the statistical generalisation as an accurate input and ask whether projecting it onto a particular individual is permissible. We argue that this shared structure obscures something crucial: statistical regularities are not model-independent; they are generated against the background of causal assumptions that agents collectively adopt and act upon. When those assumptions sustain a self-fulfilling prophecy, the dominant model produces the very data that appear to confirm it while suppressing the evidence that would reveal it as false. We use Bayesian Networks to make this mechanism explicit, showing that self-fulfilling prophecies correspond to a breakdown of the modularity and invariance conditions that underwrite causal inference. We advance two claims. Moral Exploration: when the dominant model is morally consequential, the risk of self-fulfilling prophecy gives agents reason to explore alternative causal models, even when they are less supported by existing evidence. Zetetic Improvement: exploration generates data that the dominant model had suppressed, restoring the conditions under which competing models can be discriminated. The inference problem and the model selection problem cannot be separated.

Three challenges pervade the philosophy of scientific modelling: the limits of external validity, the problem of model pluralism under deep uncertainty, and the management of non-epistemic values at the science--policy interface. This paper argues that the social tipping point literature (STPL) provides a particularly instructive case study in which all three arise simultaneously and interact in ways that have not been sufficiently recognised. We show, first, that widely used agent-based models are prone to overfitting and therefore have limited external validity beyond specific cases such as electric vehicle adoption in Norway. Drawing on the notion of nomological machines, we argue this is a structural rather than contingent limitation: social systems do not generally constitute the stable causal arrangements that would allow inferences to travel reliably across contexts. Second, current practice under-represents model uncertainty by collapsing a genuine plurality of competing models into single best-guess specifications, giving an illusion of precision in a domain characterised by deep structural uncertainty. Third, non-epistemic values in model design and interpretation are underestimated; drawing on the argument from inductive risk, we show these values routinely play a direct rather than appropriately managed indirect role in model selection. These three challenges are not independent: the choice to idealise in a particular direction is itself value-laden, and resolving model uncertainty reflects prior commitments about whose interests the model should serve. What unifies them is a persistent mismatch of scale between locally calibrated models and the global policy ambitions they are asked to support. A philosophically more mature approach to socio-ecological modelling requires addressing all three simultaneously.

Social science models exhibit remarkable diversity in the combination of their epistemic virtue. While different epistemic aims partially explain these variation, in practice competing traditions often pursue the same aims using different combinations. We propose a formal framework analysing models along three dimensions: accuracy, tractability, and normative plurality in the assumed practical and epistemic values of agents. By characterizing how these dimensions relate, we show that normative plurality plays an instrumental role mediating the accuracy-tractability trade-off. The feasible space of models forms a geometric structure with no single optimal point: different epistemic aims correspond to different optimal regions, and multiple models can legitimately occupy the same region. This provides decidability criteria for model pluralism while explaining both disciplinary specialization and legitimate competition between modelling traditions. Our framework clarifies when normative plurality is epistemically justified and moderates the claim that robustness across normative assumptions is always valuable.

Individual responsibility seems to dissolve in collective action problems: when one's causal contribution is negligible and others will not comply, demanding that one act rightly looks both futile and unfair. This paper argues that the inefficacy objection which generates that verdict has its limits. It is an argument about the causal contribution of actions, but is silent about an essential component of an agent's implication in collective harm: the beliefs about how the other agents reason and decide. I call these behavioural beliefs. The Nash counterfactual independence assumption, which underwrites the ``why bother'' reasoning endemic to collective action problems, is one such belief. I argue that where the evidence cannot settle the question of how others reason, what to believe becomes partly a moral question. I defend the Cooperative Belief Principle}: agents bear a pro tanto obligation to adopt and propagate behavioural beliefs conducive to cooperative outcomes, where Pareto-superior equilibria exist and non-cooperative beliefs are not better supported by the evidence. The principle requires no individual causal efficacy, it identifies an obligation the inefficacy objection cannot reach, and its conditions are satisfied well beyond the climate case.

 We study how teachers value school behaviour using a large-scale online experiment with 1,670 French secondary-school teachers who evaluate realistic fictitious transcripts. Academic performance and behavioural records---unjustified absences and late arrivals---are independently randomized, allowing us to isolate the causal effect of behaviour while holding academic performance constant. Teachers reward behaviour through higher grades and distinctions, but the good-behaviour premium is modest and equally distributed across student gender or academic profiles. This suggests that the behavioural component of  assessment generates differences in evaluations primarily through differences in student conduct, rather than through unequal treatment of otherwise comparable students. The content of written feedback, however, is unaffected by behavioural signals. 

We judge how dangerous a warming climate is largely from the damage we can see. But adapting to a visible hazard erases the evidence of how bad it is, and so prevents mitigation. I model this mechanism as a self-confirming equilibrium with an endogenously misspecified causal model. I show that society converges to a systematically over-optimistic belief whose bias grows with the degree of normalisation neglect. Because the neglect is itself endogenous to how salient the cost of adaptation is, classical environmental-policy instruments act on it very differently, and not as standard theory predicts. A disclosure requirement that makes the cost of adaptation salient corrects beliefs while leaving comfort intact; an adaptation tax corrects beliefs only by inducing people to adapt less; and a subsidy to adaptation deepens the complacency by censoring the damage record further. 

Storylines are emerging in climate science as a new way to communicate under deep uncertainty. This paper compares them to probabilistic communication. An expert who can only bound a physically plausible range of states can either compress that range into a probabilistic statement and communicate it through cheap talk or convey the range directly as a storyline: a physically coherent account of what could unfold. I model an ambiguity-averse listener and ask which channel serves her better. The answer depends on how sharp the uncertainty is. When the plausible range is well resolved, storylines convey strictly more and lead to better decisions; as uncertainty deepens, that advantage fades until the two approaches become equivalent. In addition, the deeper the uncertainty, the higher the evidential bar for a listener to trust the most alarming storyline.