Separating the Wheat from the Chaff
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "Separating the Wheat from the Chaff".
Warren: Here's the gist of the paper. Separating the Wheat from the Chaff studies a reputational cheap-talk environment in which a judge must choose between an expert and a quack to identify the expert, showing that truth-telling by the expert is robust and that learning in this setting exhibits an “information begets information” property The Core Game Structure The study introduces a three-player reputational cheap-talk game where a judge selects one of two speakers, one of whom is an expert perfectly informed about the state and the other is a quack with no information The speakers simultaneously send messages, and the judge selects one, with both seeking to be chosen while the judge wishes to identify the expert In this setting, an equilibrium exists where the expert reports truthfully and the quack strategically mimics the expert’s speech The judge's objective is to select the correct speaker, and each speaker wishes to be selected.
Ruth: Equilibrium Features The paper characterizes an equilibrium in which the expert always messages the true state ω and the quack’s messaging strategy replicates the prior state distribution, U
-1, 1: This equilibrium is supported by different selection strategies φ that reward speakers who send extreme messages.
Warren: A key feature is that when indifferent between two consistent messages, the judge must favor the consistent speaker who sends a more extreme message This premium given to extremism does not induce exaggeration by the expert, but instead sustains truthful communication.
Ruth: Information Amplification A central feature of equilibrium is that information begets information, meaning judges with more precise private knowledge are more likely to identify the expert and learn the true state This mechanism implies that exposure to competing sources of uncertain reliability may amplify informational inequality across audiences This observation offers a simple explanation for the coexistence of widespread access to information and growing informational inequality, without appealing to partisan bias or selective exposure.
Paper summary: Warren: So thinking about the whole paper, "Separating the Wheat from the Chaff," what does it actually mean for us outside of this game structure? Ruth I think it means that in environments where people have varying levels of information and are trying to figure out who to trust, competition itself can be a mechanism that forces some level of honesty.
Iris: It’s about how the setup—the judge choosing between an expert and a quack—creates incentives that favor truth-telling from the informed side, even when the stakes are high for both participants. Warren And it shows us that this isn't just a simple game; there are specific rules, like favoring more extreme messages when things are tied, that drive the equilibrium.
Ruth: Right, so if you look at it simply, the expert tells the truth because the structure rewards it through those tie-breaking rules. Warren And for someone just listening to this show, what's the big picture here? It’s about how reliable information can be pulled out of strategic claims through these competitive dynamics.
Iris: It points toward a way that reliable information can be extracted from messy situations, even when the sources aren't perfectly trustworthy. Ruth And it also shows that this process can have a side effect where it amplifies existing differences in what different groups believe they know about the situation.
Warren: The paper is quite robust across different assumptions about how much noise there is or how people are initially set up to believe certain things. Iris And one limitation the authors themselves mention is that this whole model relies on a specific baseline where the judge has some domain-specific heuristics, which means it doesn't capture every single nuance of real-world decision making.
Ruth: That makes sense; they're not trying to model reality perfectly, just to show what happens under these specific conditions. Warren But what they do show is that truth-telling by the expert is maintained because the structure is set up to keep it that way through those selection rules and the information feedback loop.
Iris: So ultimately, this paper gives us a framework for understanding why certain behaviors emerge in competitive information settings, showing how incentives shape communication patterns between experts and those who aren't.
Conclusion: Ruth: So we've been looking at how this game between an expert and a quack plays out, and now we’re going to wrap up what "Separating the Wheat from the Chaff" actually means for us.
Warren: It boils down to competition disciplining communication. The core idea is that in the judge's best scenario, the expert tells his true state, and that quack just copies him strategically.
Iris: And they do that because when things are tied on what both speakers say, the judge has to pick the one who makes a more extreme message. That premium for extremity sustains truth-telling from the expert.
Ruth: Right, so competition isn't just random noise; it’s actually a mechanism that forces the expert to be honest because of how the judge is set up to reward those specific types of signals.
Warren: It’s about extracting reliable information from what would otherwise be a messy exchange between someone who knows the truth and someone who doesn't.
Iris: But there’s a caveat, and it’s important. This whole structure amplifies existing differences in what different groups already think they know about the situation.
Ruth: So we have this system where reliable information can be pulled out from strategic claims, but it also has this side effect of making pre-existing knowledge gaps between audiences wider.
Warren: The authors show that this works even when you add more complexity, like different ways people start to believe things or varying noise in the judge's private signals.
Iris: They also look at how learning happens. If the judge has a clearer signal about what’s happening, they get better at figuring out the true state of things.
Ruth: It suggests that exposure to competing sources of information can actually increase informational inequality across different groups in a way that isn't just about bias.
Warren: That’s the big picture for me—it shows how we can extract reliable signals even when the sources aren't perfectly trustworthy, provided you have a judge with some level of private knowledge.
Iris: And they also look at sequential communication, which changes things entirely by disrupting the quack’s symmetry with the expert.
Ruth: It’s a tight squeeze for these speakers—the expert stays truthful, the quack mimics him in this setup—and it's a neat mechanism for extracting signal.
Warren: So what this means for us is that when you look at how information flows between experts and non-experts, the structure matters more than just the individual claims being made.
Toulouse School of Economics · London School of Economics
econ.TH
Submitted: 2026-01-02
Updated: 2026-11-04
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 76/100
The gist: Separating the Wheat from the Chaff studies a reputational cheap-talk environment in which a judge must choose between an expert and a quack to identify the expert…
Key concepts
- Reputational Cheap-Talk Game
- A three-player game where a judge chooses between an expert (who knows the truth) and a quack (who doesn't). Both speakers want to be chosen, but the judge wants to identify the expert. The core finding is that this competition leads to predictable behaviors in communication.
- Information Begets Information
- This means that judges with more precise private knowledge about a situation are better at identifying the expert and learning the true state. Exposure to different sources of uncertain reliability can actually make informational inequality between audiences grow, even if it doesn't rely on bias.
- Ex-Ante Symmetry
- In this context, ex-ante symmetry refers to whether a speaker's strategy is identical regardless of what the other speaker does beforehand. The paper shows that sequential communication (speakers reporting one after another) breaks this symmetry for the quack relative to the expert.
Terminology
Summary
Separating the Wheat from the Chaff studies a reputational cheap-talk environment in which a judge must choose between an expert and a quack to identify the expert, showing that truth-telling by the expert is robust and that learning in this setting exhibits an “information begets information” property
The Core Game Structure
The study introduces a three-player reputational cheap-talk game where a judge selects one of two speakers, one of whom is an expert perfectly informed about the state and the other is a quack with no information The speakers simultaneously send messages, and the judge selects one, with both seeking to be chosen while the judge wishes to identify the expert In this setting, an equilibrium exists where the expert reports truthfully and the quack strategically mimics the expert’s speech The judge's objective is to select the correct speaker, and each speaker wishes to be selected.
Equilibrium Features
The paper characterizes an equilibrium in which the expert always messages the true state ω and the quack’s messaging strategy replicates the prior state distribution, U[-1, 1] This equilibrium is supported by different selection strategies φ that reward speakers who send extreme messages. A key feature is that when indifferent between two consistent messages, the judge must favor the consistent speaker who sends a more extreme message This premium given to extremism does not induce exaggeration by the expert, but instead sustains truthful communication. p. 3 As a result, we find that, in equilibrium, when breaking
ties, the judge must favor the consistent speaker who sends a more extreme
message. Perhaps surprisingly, this premium given to extremism does not incite the expert to exaggerate his own report. As a result, we find that, in equilibrium, when breaking
ties, the judge must favor the consistent speaker who sends a more extreme
message. Perhaps surprisingly, this premium given to extremism does not incite the expert to exaggerate his own report.Where the paper says it
Where the paper says it
Information Amplification
A central feature of equilibrium is that information begets information, meaning judges with more precise private knowledge are more likely to identify the expert and learn the true state This mechanism implies that exposure to competing sources of uncertain reliability may amplify informational inequality across audiences This observation offers a simple explanation for the coexistence of widespread access to information and growing informational inequality, without appealing to partisan bias or selective exposure.
Robustness and Extensions
The findings are robust across various distributional assumptions; the judge’s best equilibrium continues to feature truth-telling by the expert and a selection rule that favors more extreme messages when breaking ties Extensions explore non-uniform state distributions, asymmetric prior beliefs about speaker identities, and non-uniform noise in the judge’s private signal. In these extensions, the quack's incentives to mimic the expert dampen as we go from higher to lower messages.
Learning Dynamics
The judge's learning about the state is characterized by Proposition 2, which shows that the probability that the judge learns the state is given by Pn Gˆ = 1ifω o = 1 − ε¯23 This probability is increasing in the precision of the judge’s private signal. Furthermore, Proposition 3 shows that the reduction in expected posterior variance is an increasing function of the judge’s noise ε¯ ∈ [0, 1] This means that less informed judges benefit most from communication coming from competing and potentially unreliable sources in terms of posterior variance reduction.
The Role of Communication Order
Sequential communication, where speakers report to the judge after one another, is shown to destroy the quack’s ex-ante symmetry with the expert In this setting, sequential communication destroys the quack’s ex-ante symmetry with the expert. The key insight is that sequential communication destroys the quack’s ex-ante symmetry with the expert. p. 3 The selection rule used by the judge in such indifference cases is a key equilibrium object. If the judge breaks ties uniformly, randomly selecting one of the
two consistent speakers, the quack has incentives to “pander to the mean,” because more moderate messages are ex-ante more likely to be consistent with
the judge’s signal.Where the paper says it
Conclusion
In summary, competition between speakers can discipline communication: in the judge’s best equilibrium, the expert reports truthfully, the quack strategically mimics the expert’s speech, and the judge rationally favors more extreme messages when breaking ties These forces generate a simple mechanism through which reliable information can be extracted from strategic claims while sometimes also amplifying pre-existing differences in audiences’ information. In the judge’s best equilibrium, the expert reports his signal truthfully, and the quack’s messaging strategy does not “pander” to states that are most likely to arise The paper concludes by showing that sequential communication destroys the quack’s ex-ante symmetry with the expert. p. 2 We focus on the equilibrium that
yields the highest ex-ante value to the judge. In this equilibrium, (1) the
expert reports the state truthfully; (2) the quack chooses his message at random, replicating the marginal distribution over messages used by the expert;
(3) when indifferent, the judge is biased towards selecting the speaker who
sent the more extreme message. We focus on the equilibrium that
yields the highest ex-ante value to the judge. In this equilibrium, (1) the
expert reports the state truthfully; (2) the quack chooses his message at random, replicating the marginal distribution over messages used by the expert;
(3) when indifferent, the judge is biased towards selecting the speaker who
sent the more extreme message.Where the paper says it
Where the paper says it
References
Acemoglu, D., Chernozhukov, V., and M. Yildız (2009). “Fragility of Asymptotic Agreement under Bayesian Learning,” Theoretical Economics, 11, 187–225
Andreoni, J., and T. Mylovanov (2012). p. 6 This result rests on the fact that speakers are only concerned
about building reputation as an expert, which implies that, fixing the expert’s strategy, the communication game is zero-sum between the quack and
the judge. 4Relatedly, Andreoni and Mylovanov (2012) show how two-dimensional uncertainty —
in our case, about the state and about the speakers’ expertise — might lead to divergence
of posterior beliefs along a one-dimensional subspace.Where the paper says it
Angelucci, C., and A. Prat (2024). p. 3 This observation offers
a simple explanation for the coexistence of widespread access to information
and growing informational inequality, without appealing to partisan bias or
selective exposure. It also connects naturally to recent experimental evidence:
Angelucci and Prat (2024) show that individuals differ sharply in their ability
to identify true news when confronted with both true and false reports, and
that these differences are strongly related to background characteristics that
plausibly proxy for prior information or domain knowledge. We show that learning in this environment exhibits an “information begets information” property: judges
with more precise private information are more likely to identify the
expert and learn the true state, implying that exposure to competing
sources of uncertain reliability may amplify informational inequality
across audiences. 1 Introduction
New media technologies do not discriminate between truth and lies, leaving
it up to people to decide whom to believe.Where the paper says it
Where the paper says it
Battaglini, M. (2002). “Multiple Referrals and Multidimensional Cheap Talk,” Econometrica, 70, 1379–1401
Brandenburger, A., and B. Polak (1996). p. 5 One implication
is that information polarization or inequality can arise because different media outlets cater to pre-existing heterogeneous priors across a population. A
similar “echo-chamber” result supports informational inequalities in Perego
2
In Brandenburger and Polak (1996), the sender’s concern isn’t about the receiver’s
assessment of their ability, but about the quality of their decisions.Where the paper says it
Cheng, I.-H., and A. Hsiaw (2022). p. 6 and Yuksel (2022), and the mechanism is strengthened by media competition. Our paper proposes an alternative explanation for the coexistence of
widespread access to information and growing information inequality, which
does not appeal to media companies pandering to their audience (remember
that our benchmark equilibrium features truth-telling by the expert and no
pandering by the quack).Where the paper says it
Crawford, V., and J. Sobel (1982). p. 4 As a result, commitment has no value in this environment: the judge’s payoff (the probability of identifying the expert) is pinned
down entirely by how successfully the quack can masquerade as informed. Related literature: Our paper contributes to the vast literature on cheap
talk models, starting with Crawford and Sobel (1982) and Green and Stokey
(2007), and especially that on reputational cheap talk.Where the paper says it
Dixit, A., and J. Weibull (2007). p. 4 As a result, commitment has no value in this environment: the judge’s payoff (the probability of identifying the expert) is pinned
down entirely by how successfully the quack can masquerade as informed. Related literature: Our paper contributes to the vast literature on cheap
talk models, starting with Crawford and Sobel (1982) and Green and Stokey
(2007), and especially that on reputational cheap talk.Where the paper says it
Effinger, M. R., and M. K. Polborn (2001). p. 4 Related literature: Our paper contributes to the vast literature on cheap
talk models, starting with Crawford and Sobel (1982) and Green and Stokey
(2007), and especially that on reputational cheap talk. Morris (2001) shows
5Where the paper says it
Friedman, E. (1998). “Public Debate among Experts,” Northwestern University, CMS-EMS Working Paper 1234
Gentzkow, M., and J. M. Shapiro (2006). p. 5 2, we show that pandering by the quack to extreme messages — against the
prior — can arise due to the structure of the receiver’s private information. 3
Gentzkow and Shapiro (2006) is particularly close to our work in motivation:
they interpret media companies as speakers who wish to develop a reputation
for expertise (in their model, expertise is tied with honesty).Where the paper says it
Glazer, J., and C.-T. Ma (1999). “Efficient Allocation of a ‘Prize’—King Solomon’s Dilemma,” Games and Economic Behavior, 27, 222–233
Glazer, J., and A. Rubinstein (2001). p. 6 Ottaviani and Sørensen (2001) find that
(sometimes) it is best to start with the best informed speaker when speakers
proceed sequentially. Glazer and Rubinstein (1997) show that sequential communication can be used to economize on information transmission.Where the paper says it
Glazer, J., and A. Rubinstein (1998). p. 6 5See also Friedman (1998), in which the decision to become an expert is a choice, as a
function of the bias. 6The role of motives in the design of mechanisms is further elaborated on in Glazer
and Rubinstein (1998).Where the paper says it
Green, J. R., and N. Stokey (2007). p. 4 As a result, commitment has no value in this environment: the judge’s payoff (the probability of identifying the expert) is pinned
down entirely by how successfully the quack can masquerade as informed. Related literature: Our paper contributes to the vast literature on cheap
talk models, starting with Crawford and Sobel (1982) and Green and Stokey
(2007), and especially that on reputational cheap talk.Where the paper says it
Karlin, S. (1959). Mathematical Methods and Theory in Games, Programming, and Economics, Vol. p. 0 Separating the Wheat from the Chaff
Johannes Hörner Paula Onuchic
Toulouse School of Economics London School of Economics
January, 2026
Abstract
We study a reputational cheap-talk environment in which a judge,
who is privately and imperfectly informed about a state, must choose
between two speakers of unknown reliability. Exactly one speaker is
an expert who perfectly observes the state, while the other is a quack
with no information. (In one of our extensions, we consider epistemic knowledge about the
sources themselves, reflecting some previous experience that directly affects
the judge’s prior belief about each speaker’s expertise. )
Cheap-talk games admit many equilibria.Where the paper says it
Where the paper says it
Krishna, V., and J. Morgan (2001). p. 6 A smaller literature equally considers models of cheap talk with multiple
senders. In Krishna and Morgan (2001) and Battaglini (2002), the experts
have known biases, as in Crawford and Sobel, and Green and Stokey.Where the paper says it
Lehmann, E. (1988). “Comparing Location Experiments,” Annals of Statistics, 16, 521–533
Maskin, E. (1999). “Nash Equilibrium and Welfare Optimality,” Review of Economic Studies, 66, 23–38
Morris, S. (2001). “Political Correctness,” Journal of Political Economy, 114, 280–316
Nichols, T. (2017). Why We Distrust Experts: The Death of Expertise Explained. p. 1 When audiences must sort through conflicting claims without an external
arbiter, their problem of learning the truth becomes inseparable from that
of learning about the expertise of different sources. This difficulty is compounded by the fact that speakers are not neutral conveyors of information:
they are often motivated by the prospect of being recognized, trusted, or selected as experts.Where the paper says it
Olszewski, W. (2004). “Informal Communication,” Journal of Economic Theory, 117, 180–200
Ottaviani, M., and P. Sørensen (2001). p. 5 Olszewski (2004)
considers a sender who wishes to be perceived as honest, and shows that
truth-telling is an equilibrium. Our paper is more closely related to those
in which senders are concerned about their reputation for expertise, such
as Scharfstein and Stein (1990), Prendergast (1993), Prendergast and Stole
(1996), and Ottaviani and Sørensen (2001, 2006a, 2006b). Perhaps surprisingly, this premium given to extremism does not incite the expert to exaggerate his own report. Roughly speaking, because of his
private information, the probability the expert speaker assigns to his message
being consistent with the judge’s signal is more elastic than the quack’s: by
straying away from the truth, he knows that he risks inconsistency, whereas
the quack does not know where the truth lies in the first place.Where the paper says it
Where the paper says it
Ottaviani, M., and P. Sørensen (2006a). p. 5 Olszewski (2004)
considers a sender who wishes to be perceived as honest, and shows that
truth-telling is an equilibrium. Our paper is more closely related to those
in which senders are concerned about their reputation for expertise, such
as Scharfstein and Stein (1990), Prendergast (1993), Prendergast and Stole
(1996), and Ottaviani and Sørensen (2001, 2006a, 2006b).Where the paper says it
Ottaviani, M., and P. Sørensen (2006b). p. 5 Olszewski (2004)
considers a sender who wishes to be perceived as honest, and shows that
truth-telling is an equilibrium. Our paper is more closely related to those
in which senders are concerned about their reputation for expertise, such
as Scharfstein and Stein (1990), Prendergast (1993), Prendergast and Stole
(1996), and Ottaviani and Sørensen (2001, 2006a, 2006b).Where the paper says it
Perego, J., and S. Yuksel (2022). p. 6 and Yuksel (2022), and the mechanism is strengthened by media competition. Our paper proposes an alternative explanation for the coexistence of
widespread access to information and growing information inequality, which
does not appeal to media companies pandering to their audience (remember
that our benchmark equilibrium features truth-telling by the expert and no
pandering by the quack).Where the paper says it
Prendergast, C. (1993). “A Theory of ‘Yes Men’,” American Economic Review, 83, 757–770
Prendergast, C., and L. Stole (1996). p. 5 Olszewski (2004)
considers a sender who wishes to be perceived as honest, and shows that
truth-telling is an equilibrium. Our paper is more closely related to those
in which senders are concerned about their reputation for expertise, such
as Scharfstein and Stein (1990), Prendergast (1993), Prendergast and Stole
(1996), and Ottaviani and Sørensen (2001, 2006a, 2006b).Where the paper says it
Scharfstein, D., and J. Stein (1990). p. 5 Olszewski (2004)
considers a sender who wishes to be perceived as honest, and shows that
truth-telling is an equilibrium. Our paper is more closely related to those
in which senders are concerned about their reputation for expertise, such
as Scharfstein and Stein (1990), Prendergast (1993), Prendergast and Stole
(1996), and Ottaviani and Sørensen (2001, 2006a, 2006b).Where the paper says it
Wilson, A. (2014). “Bounded Memory and Biases in Information Processing,” Econometrica, 82, 2257–2294
Pruning redundant reasoning tokens
The paper characterizes an equilibrium in which the expert always messages the true state ω and the quack’s messaging strategy replicates the prior state distribution, U[-1, 1] This equilibrium is supported by different selection strategies φ that reward speakers who send extreme messages. A key feature is that when indifferent between two consistent messages, the judge must favor the consistent speaker who sends a more extreme message This premium given to extremism does not induce exaggeration by the expert, but instead sustains truthful communication.
How the training signal is built
The judge's learning about the state is characterized by Proposition 2, which shows that the probability that the judge learns the state is given by Pn Gˆ = 1ifω o = 1 − ε¯23 This probability is increasing in the precision of the judge’s private signal. Furthermore, Proposition 3 shows that the reduction in expected posterior variance is an increasing function of the judge’s noise ε¯ ∈ [0, 1] This means that less informed judges benefit most from communication coming from competing and potentially unreliable sources in terms of posterior variance reduction.
Where the gains come from
The paper shows that competition between speakers can discipline communication: in the judge’s best equilibrium, the expert reports truthfully, the quack strategically mimics the expert’s speech, and the judge rationally favors more extreme messages when breaking ties This mechanism generates a simple way through which reliable information can be extracted from strategic claims while sometimes also amplifying pre-existing differences in audiences’ information The analysis relies on a number of stark assumptions, some of which are relaxed in extensions. In the judge’s best equilibrium, the expert reports his signal truthfully, and the quack’s messaging strategy does not “pander” to states that are most likely to arise The paper concludes by showing that sequential communication destroys the quack’s ex-ante symmetry with the expert.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Strategic Information Extraction in Media Environments: The AI system can model how competing sources of uncertain reliability amplify informational inequality across audiences, leading to a mechanism where
exposure to competing sources of uncertain reliability may amplify informational inequality across audiences.
-
Expert Identification and Truth-Telling Robustness: The system can be designed to favor truth-telling by the expert even under noisy or partial information, as the paper shows that
the key force — the greater elasticity of the informed speaker’s probability of being selected when deviating from his signal — may still favor honesty.
-
Dynamic Prior Adjustment for Uninformed Agents: The AI can implement a strategy where uninformed agents adjust their messaging based on what they believe is most likely to be consistent with the judge's signal, as seen in the quack's behavior pandering to moderate messages when
the quack’s incentives to mimic the expert don’t dampen as we go from higher to lower messages.
-
Adaptive Selection Rules Based on Signal Precision: The system can employ a selection rule that rewards more extreme messages when breaking ties, as shown in Theorem 1, allowing the judge to learn the true state even when
the judge is uncertain favors more extreme messages when breaking ties between otherwise plausible claims.
Abstract
We study a reputational cheap-talk environment in which a judge chooses between two speakers of unknown expertise, an expert and a quack. Both speakers seek to be selected, while the judge wishes to identify the expert. The best equilibrium for the judge features truth-telling by the expert, yet the judge favors speakers with more extreme messages. This bias toward extremism does not induce exaggeration, but instead sustains truthful communication. We characterize the quack's equilibrium messaging strategy, relating the degree to which it panders to the judge's prior to the state distribution, the speakers' initial reputations, and the judge's private knowledge of the state. We interpret these results in the context of two applications, the post-training of AI chatbots through human feedback and media competition.