Fooling Yourself: how narratives shape beliefs

arXiv:2607.04753 · econ.GN, q-fin.EC · Submitted 2026-07-06 · Read on arXiv

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: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Fooling Yourself: how narratives shape beliefs".

Paper summary: Ruth: So we're looking at this paper called "Fooling Yourself: how narratives shape beliefs." The main idea is that decision-makers often get information in these stories, and those stories mix facts that actually help you with the decision with details that don't actually tell you anything useful for making a choice.

Warren: Exactly. It claims that when these irrelevant details are put into a narrative, they systematically pull people's beliefs toward fifty-fifty, which is the point where everyone is maximally uncertain, even though Bayes’ rule says you shouldn't update your belief based on that kind of information <ref:2607.04753#pg2>.

Iris: That means these details aren't just noise; they actively push the belief away from whatever the actual evidence suggests and back toward a neutral starting point, which is a very important mechanism to figure out <ref:2607.04753#pg1>.

Ruth: Right. The paper sets up three different ways these details can be presented: an urn problem with colored balls, an investigation story with clues pointing to one suspect versus neither, and a narrative context where those clues are replaced by just a message saying nothing new happened <ref:2607.04753#pg1>.

Warren: And the core finding is that this effect is strongest when those nondiagnostic details are embedded in the investigation narrative, even though it also happens in the abstract urn problem <ref:2607.04753#pg1>.

Iris: So they're saying that presenting a clue that points to neither suspect, within a story about theft, has a distinct effect compared to just seeing a white ball in an urn setup <ref:2607.04753#pg3>.

Ruth: That’s what the comparison between the Story and Story No Info conditions is designed to isolate—to see if narrative structure itself is responsible for this systematic pull toward uncertainty <ref:2607.04753#pg1>.

Warren: And they found that in the Story No Info condition, where those clues are just replaced by a message saying no new clue arrived, the reported beliefs are much closer to what Bayes’ rule would predict <ref:2607.04753#pg1>.

Iris: That tells us that if you strip away the narrative structure and just look at the raw information, people behave more predictably according to standard probability theory <ref:2607.04753#pg1>.

Ruth: The paper goes on to talk about how this behavior affects real-world situations, looking at things like how long it takes someone to make a decision when they're waiting for information, and the economic cost of that delay <ref:2607.04753#pg1>.

Warren: It showed through simulations that when you have to wait for a decision—say, in an investment scenario—this behavior actually slows down how fast agents reach a conclusion compared to when they only see the diagnostic signals <ref:2607.04753#pg1>.

Iris: And it introduces these behavioral types, like the Bayes-Reset type, which is one that resets its belief after seeing a diagnostic signal but then drifts back toward uncertainty after seeing those nondiagnostic ones <ref:2607.04753#pg1>.

Ruth: It seems the story condition actually increases the frequency of this midpoint-reset behavior following a nondiagnostic clue, which is a compositional shift in how people update their beliefs <ref:2607.04753#pg1>.

Warren: So what does this mean for practical applications? The authors suggest that when narratives are used to pass information—like in judicial evaluations or project approvals—this can actually delay the decision-making process because of these narrative effects <ref:2607.04753#pg1>.

Iris: And they also pointed out a limitation, which is important for any finding. The paper states that even though subjects were told those nondiagnostic clues carried no evidential value, they still revised their beliefs in a way that defied the expectation of pure Bayesian updating <ref:2607.04753#pg1>.

Ruth: That’s a fair point for skepticism. It highlights that simply telling people something is irrelevant isn't enough to make them ignore it completely when it's packaged in a story, which means institutional design might need to be more nuanced than just filtering out irrelevant data <ref:2607.04753#pg2>.

Warren: So, the big picture here is that narratives aren't just containers for facts; they are active components that can systematically distort how we process information and update our beliefs, pulling us toward uncertainty rather than pure evidence <ref:2607.04753#pg1>.

Iris: And the finding that this effect is stronger in stories versus abstract problems suggests that the context of the narrative is what makes those nondiagnostic details stick with people and influence their judgment <ref:2607.04753#pg1>.

Ruth: We're wrapping up our look at "Fooling Yourself: how narratives shape beliefs." The authors argue that since these details pull beliefs toward maximal uncertainty, we need to be careful about how we structure the stories we use to communicate important information <ref:2607.04753#pg1>.

Warren: They suggest that if you’re using narratives in areas like investment decisions or resource allocation, you need to consider how those non-diagnostic details might actually slow down convergence to a final decision <ref:2607.04753#pg1>.

Iris: We've seen how the structure of the story matters for these updates and we've also seen that this effect is present even when those clues are replaced by no-information messages within the same story context <ref:2607.04753#pg3>.

Ruth: That’s what we’ve discussed today about "Fooling Yourself: how narratives shape beliefs." It seems the narrative format itself is a key factor in influencing belief revision, especially when it includes details that don't carry direct evidence <ref:2607.04753#pg1>.

Warren: And the implication is that for anyone building systems based on stories to transmit information, they need to be aware of how those non-diagnostic elements might unintentionally delay action or keep people stuck in indecision <ref:2607.04753#pg1>.

Iris: That's it for this paper. We've covered the core claim about nondiagnostic details pulling beliefs toward zero point five, and how that manifests differently depending on whether you use a story or an abstract setup <ref:2607.04753#pg3>.

Conclusion: Ruth: So we're wrapping up this look at "Fooling Yourself: how narratives shape beliefs." The main point is that those extra details in a story aren't just noise; they actively pull people's beliefs toward fifty-fifty uncertainty, even when Bayes’ rule says to stick to the actual evidence.

Warren: It means these narrative structures can systematically change how we update our minds, pushing us away from what the math actually suggests is true.

Iris: And the real takeaway is that this effect isn't just a quirk of one setup; it happens across different ways you present information, whether it’s an urn or a complex investigation story.

Ruth: The authors found this pull toward zero point five is strongest when those extra clues are embedded in a narrative context, but they also saw it in the simpler urn problem.

Warren: That suggests the *way* you tell the story matters more than just having some pieces of data floating around randomly. It’s the structure that makes people hesitate.

Iris: And what this means for us is that if you're using stories to pass on important information, like in a court case or an investment pitch, you have to be careful about those non-diagnostic details.

Ruth: Exactly. If those irrelevant details are present, they can actually slow down how fast people reach a decision.

Warren: It makes the time it takes for an agent to make a call much longer when they're waiting on that kind of information flow.

Iris: So we’ve seen how narrative structure influences belief revision and how that slows down the process, but what are the real-world costs of that delay?

Ruth: That’s exactly where we need to look next. The paper lays out some economic implications about when this delay matters for things like investment decisions or even organizational resource allocation.

IMT School for Advanced Studies Lucca · Department of Economics and Statistics, University of Siena · BIDSA, Universita Bocconi, Milano · Prague University of Economics and Business

econ.GN, q-fin.EC

Submitted: 2026-07-06

Updated: 2026-11-18

Code: https://github.com/paolopin/narrative-belief-updating

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

The gist: Decision-makers often receive information through narratives combining diagnostic evidence with details that carry no information useful for inference <ref:2607.04753#pg2> In a laboratory experiment…

Key concepts

Nondiagnostic Details
These are details within a story or problem that carry no useful information for making a specific decision, such as white balls in an urn problem. The study found these details systematically push participants' beliefs toward 0.5, representing maximal uncertainty, instead of updating them based on new evidence.
Bayesian Updating
This is the standard mathematical rule used to revise beliefs when new information is received. It dictates that if you have a prior belief and receive evidence, your belief should change in a predictable way. The paper shows that in narratives, this rule is often ignored or overridden.
Narrative Context
This refers to the surrounding story or context in which signals are presented. The study compared how nondiagnostic clues appear within an investigation narrative versus being replaced by explicit 'no-information' messages, showing that the narrative structure matters for belief revision.

Terminology

Summary

Decision-makers often receive information through narratives combining diagnostic evidence with details that carry no information useful for inference <ref:2607. p. 004753#pg2> In a laboratory experiment, participants repeatedly report incentivized beliefs in a Bayesian inference task <ref:2607. p. 004753#pg9> We find that nondiagnostic details systematically pull participants’ beliefs toward 0. p. 05, the point of maximal uncertainty, despite Bayes’ rule prescribing no revision This response is strongest when nondiagnostic clues are embedded in the investigation narrative, although it also occurs in the urn condition

Experimental Design and Conditions

The study implemented three conditions with the same statistical structure to examine how narratives shape belief updating These conditions included an urn problem with diagnostic colored balls and nondiagnostic white balls, an investigation narrative with diagnostic clues pointing to one suspect and nondiagnostic clues pointing to neither, and a narrative context with no-information messages replacing nondiagnostic clues Participants were explicitly informed of the prior probability of each state and the probability of every possible signal under each state The experimental design involved a between-subjects design where participants observed a sequence of signals and repeatedly stated their subjective beliefs about which of two possible states is true

Results on Belief Revision

The main result shows that nondiagnostic signals are not treated as such, especially when explicitly presented within a narrative In the Story condition, subjects systematically revise their beliefs after observing nondiagnostic clues, even though Bayes’ rule prescribes no update These revisions are not mere noise; beliefs systematically move back toward 0.5, the point at which the two states are judged equally likely A similar but less pronounced pattern is present in the Urn condition Conversely, in Story No Info, where nondiagnostic clues are replaced by explicit no-information messages within the same investigation story, reported beliefs are substantially closer to the Bayesian benchmark

Economic Magnitude and Stopping Time

Monte Carlo simulations of a sequential stopping problem quantified how the experimentally estimated return-to-uncertainty behavior affects the amount of evidence required to reach a decision Relative to Story No Info, Story substantially slows convergence, as simulated agents require more signals before reaching the threshold This delay is economically relevant whenever information acquisition is costly, or when delay itself reduces payoffs because opportunities expire

Behavioral Classification and Persistence

The analysis classified subjects into four behavioral types, including Bayes-Reset, which follows the same rule after diagnostic signals but returns toward uncertainty after nondiagnostic ones The type estimates reveal a clear compositional shift: Story increases the prevalence of midpoint-reset behavior following a nondiagnostic signal Furthermore, exact resets to 0.5 are not simply reversed at the next diagnostic signal; subsequent beliefs remain closer to those implied by updating from the reset belief than from the pre-reset belief

Implications for Institutional Design

The simulation shows that exposure to nondiagnostic narrative clues can delay belief convergence, requiring agents to observe more signals before taking action This matters in settings where narratives are used to transmit information and timely action is valuable, such as investment decisions, judicial evaluations, project approvals, and organizational resource allocation Our findings also have implications for institutional design In our experiment, subjects were informed that nondiagnostic clues carried no evidential value, yet revised their beliefs nonetheless Transparency about irrelevance may therefore be insufficient; institutions may need to limit or filter nondiagnostic content in decision-relevant communications

Conclusion on Mechanism

Our main finding is inconsistent with Bayesian updating and, to the best of our knowledge, has not been documented in prior research Nondiagnostic signals neither systematically shift beliefs toward a particular state nor primarily serve to reinforce prior beliefs; instead, they tend to draw beliefs toward maximal uncertainty This response is sharply reduced when the same nondiagnostic events are replaced by no-information messages within the same narrative context, and is less pronounced in an abstract urn problem with the same statistical structure The economic implications are potentially meaningful

Data and Code Availability

The replication package for this paper is publicly available at https://github.com/paolopin/narrative-belief-updating The repository contains the Stata code for the empirical analysis, the MATLAB code for the behavioral-type classification, and the Python code for the numerical simulations

References

Aina, C. (2025). Tailored stories. p. 1 Working Paper <ref:2607. p. 004753#pg2> Alos-Ferrer, C. and M. Mihm (2023). p. 4 An axiomatic characterization of bayesian updating <ref:2607. p. 004753#pg9> Andre, P. p. 4, I. Haaland, C. Roth, M. Wiederholt, and J. Wohlfart (2026, 02). p. 0 Narratives about the macroeconomy <ref:2607. p. 004753#pg2> Aydogan, I., A. Baillon, E. Kemel, and C. Li (2017). Signal perception and belief updating <ref:2607. p. 204753#pg9> Barron, K. p. 4 and T. Fries (2024). p. 4 Narrative persuasion <ref:2607. p. 304753#pg9> Benabou, R. p. 4, A. Falk, and J. Tirole (2020). p. 4 Narratives, imperatives, and moral persuasion <ref:2607. p. 404753#pg9> Benjamin, D. p. 4 J. (2019). Chapter 2 - errors in probabilistic reasoning and judgment biases <ref:2607.04753#pg9> Bock, O. p. 8, I. Baetge, and A. Nicklisch (2014). hroot: Hamburg registration and organization online tool <ref:2607. p. 804753#pg9> Bright, D. p. 5 A. and J. Goodman-Delahunty (2006). p. 5 Gruesome evidence and emotion: anger, blame, and jury decision-making <ref:2607. p. 004753#pg9> Charness, G. and C. Dave (2017). p. 4 Confirmation bias with motivated beliefs <ref:2607. p. 504753#pg9> Chater, N. p. 1 and G. Loewenstein (2016). p. 1 The under-appreciated drive for sense-making <ref:2607. p. 004753#pg9> Chen, D. p. 8 L., M. Schonger, and C. Wickens (2016). p. 1 otree-an open-source platform for laboratory, online, and field experiments <ref:2607.04753#pg9> Conlon, J. p. 4 J. and S. Y. Kwon (2026). p. 4 Beliefs from cues <ref:2607. p. 004753#pg9> Coutts, A. (2019). Good news and bad news are still news: Experimental evidence on belief updating <ref:2607. p. 404753#pg9> Douglas, K. p. 5 S., D. R. Lyon, and J. R. Ogloff (1997). p. 5 The impact of graphic photographic evidence on mock jurors’ decisions in a murder trial: Probative or prejudicial? p. 1 Law and human behavior <ref:2607. p. 004753#pg9> Eliaz, K. p. 4 and R. Spiegler (2020). p. 4 A model of competing narratives <ref:2607. p. 004753#pg9> Federal Rules of Evidence: December 1 (2024). p. 1 Rules 401–403 <ref:2607. p. 004753#pg9> Fedyk, A. p. 4 and J. Hodson (2023). p. 4 When can the market identify old news? p. 4 Journal of Financial Economics <ref:2607. p. 004753#pg9> Fiske, S. p. 1 (2018). Social cognition: selected works of Susan Fiske <ref:2607.04753#pg9> Flynn, J. p. 4 and K. Sastry (2024). p. 4 The macroeconomics of narratives <ref:2607. p. 004753#pg9> Gilbert, T. p. 4, S. Kogan, L. Lochstoer, and A. Ozyildirim (2012). p. 4 Investor inattention and the market impact of summary statistics <ref:2607. p. 404753#pg9> Goetzmann, W. p. 4 N., D. Kim, and R. J. Shiller (2024). p. 4 Emotions and subjective crash beliefs <ref:2607. p. 104753#pg9> Graeber, T. p. 4, C. Roth, and F. Zimmermann (2024). p. 1 Stories, statistics, and memory <ref:2607. p. 404753#pg9> Grether, D. M. (1980). Bayes rule as a descriptive model: The representativeness heuristic <ref:2607. p. 004753#pg9> Han, Y. p. 4, D. Huffman, and Y. Liu (2024). p. 1 Minds, models and markets: How managerial cognition affects pricing strategies <ref:2607. p. 404753#pg9> Huang, X. p. 1, S. H. Teoh, and Y. Zhang (2014). Tone management <ref:2607. p. 104753#pg9> Igou, E. p. 3 R. and H. Bless (2005). p. 3 The conversational basis for the dilution effect <ref:2607. p. 304753#pg9> Kemmelmeier, M. p. 3 p. 3 (2004). Separating the wheat from the chaff: Does discriminating between diagnostic and nondiagnostic information eliminate the dilution effect? p. 3 Journal of Behavioral Decision Making <ref:2607. p. 004753#pg9> Kemmelmeier, M. (2007). Does the dilution effect have a conversational basis? p. 3 Journal of Language and Social Psychology <ref:2607. p. 004753#pg9> Kieren, P. p. 4 and M. Weber (2025). p. 4 Expectation formation under uninformative signals <ref:2607. p. 304753#pg9> Levy, G. p. 4, R. Razin, and A. Young (2022). p. 4 Misspecified politics and the recurrence of populism <ref:2607. p. 004753#pg9> Li, F. (2008). Annual report readability, current earnings, and earnings persistence <ref:2607. p. 104753#pg9> Loughran, T. p. 1 and B. McDonald (2011). p. 1 When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks <ref:2607. p. 004753#pg9> Morag, D. p. 4 and G. Loewenstein (2025). p. 4 Narratives and valuations <ref:2607. p. 004753#pg9> Musolff, R. p. 4, C. Roth, and F. Zimmermann (2026). p. 0 Learning from false stories <ref:2607. p. 404753#pg9> Nisbett, R. p. 3 E., H. Zukier, and R. E. Lemley (1981). The dilution effect: Nondiagnostic information weakens the implications of diagnostic information <ref:2607. p. 304753#pg9> Ortoleva, P. p. 4 (2024). Alternatives to bayesian updating <ref:2607. p. 004753#pg9> Peters, E. p. 3 and M. Rothbart (2000). p. 3 Typicality can create, eliminate, and reverse the dilution effect <ref:2607. p. 304753#pg9> Pin, P. p. 0 and T. Rotesi (2023). p. 8 App-based experiments <ref:2607. p. 004753#pg9> Schwartzstein, J. p. 4 and A. Sunderam (2021). p. 4 Using models to persuade <ref:2607. p. 004753#pg9> Schwarz, G. (1978). Estimating the dimension of a model <ref:2607. p. 004753#pg9> Shiller, R. p. 4 p. 4 J. (2017). Narrative economics <ref:2607. p. 004753#pg9> Shiller, R. J. (2019). Narrative Economics: How Stories Go Viral and Drive Major Economic Events <ref:2607. p. 404753#pg9> Tetlock, P. p. 4 C. (2011). All the news that’s fit to reprint: Do investors react to stale information? p. 4 The Review of Financial Studies <ref:2607. p. 404753#pg9> Thaler, M. p. 4 (2024, May). p. 1 The fake news effect: Experimentally identifying motivated reasoning using trust in news <ref:2607.04753#pg9> U.S. Securities and Exchange Commission (2024). p. 1 Regulation s-k. p. 1 Item 303, Management’s Discussion and Analysis of Financial Condition and Results of Operations <ref:2607. p. 104753#pg9> Whitson, J. p. 1 A. and A. D. Galinsky (2008). p. 1 Lacking control increases illusory pattern perception <ref:2607.

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

  1. Improve belief updating models by incorporating a rule that accounts for reset-to-uncertainty behavior, which is observed in Story treatment, by allowing beliefs to systematically move toward 0.5 even after observing nondiagnostic signals.

  2. Enhance sequential decision-making AI systems with a stopping time calculation that incorporates the empirically estimated return-to-uncertainty behavior, which shows that midpoint reversion substantially slows belief convergence.

  3. Develop institutional recommendation systems for information filtering by incorporating findings that nondiagnostic narrative clues induce systematic belief revisions toward 0.5, thereby weakening accumulated diagnostic evidence, suggesting institutions should limit or filter nondiagnostic content in decision-relevant communications.

Related papers