For many decisions, people rely on information received from others by word of mouth; through conversations with family, friends, and strangers; via television, radio, or podcasts (Hirshleifer 2020, Shiller 2017, Shiller 2020). Attempts to verbally transmit information often introduce distortions – pieces of the information can get lost, added, or accidentally modified. If such distortions follow systematic patterns, word-of-mouth transmission could introduce predictable biases into the supply of information, thereby influencing people’s downstream beliefs and behaviours.
Human transmission of information
In a recent paper (Graeber et al. 2024), we study the nature and consequences of verbal transmission of economic information. We conduct pre-registered, incentivised experiments with more than 5,000 participants. In a first experiment, participants listen to a one-minute message giving a qualitative forecast about an economic variable and are incentivised to then record themselves passing on the information they heard. Participants separately transmit forecasts about two variables: home price growth in a US city, and revenue growth of a US retailer. They are paid based on how closely the belief updates induced by listening to their message match the belief updates induced by listening to the original forecast. Such incentives motivate a faithful transmission of all relevant information in the original forecast, which is commonplace in practice: sales employees relay customer feedback to developer teams, teachers pass on knowledge to students, doctors convey patient insights to other medical professionals in shift-to-shift handoffs, friends give advice by sharing personal experiences, and journalists convey information from sources to the public.
In a subsequent experiment, a different group of participants listen to either the original forecast or a transmitted version of that forecast before stating incentivised beliefs about the relevant variable and about the characteristics of the original prediction. Comparing the beliefs of listeners hearing original messages to the beliefs of listeners hearing transmitted versions of those messages lets us characterise distortions introduced in the transmission process using simple belief-based quantitative measures.
What matters for the downstream consequences of verbal transmission, however, is not only whether transmission garbles information in general, but whether certain types of information are subject to stronger distortions than others, creating systematic biases in the supply of information. To study whether some types of information are distorted more strongly than others, we focus on two key features of messages: the level of the prediction contained in a message and the reliability of that prediction. Our experiments separately vary the level and reliability of the original forecasts, allowing us to compare distortions of level information with distortions of reliability information.
Differential information loss
Our main finding is that information about the reliability of a prediction is lost in transmission about three times as much as information about the prediction’s level. We document this finding using three distinct and complementary sets of analyses.
In our first set of analyses, we examine listeners’ beliefs about the level and reliability of the predictions in the original messages. We estimate the sensitivity of those beliefs to the experimental manipulations of level and reliability. We then compare the sensitivity of listeners who directly hear the original messages to the sensitivity of listeners who hear transmitted versions of those messages. The difference between the sensitivities of the two groups is our main measure of transmission-induced information loss.
Consider the loss of level information. Among listeners who directly hear the original messages, switching from a low-level message to a high-level message shifts beliefs about the prediction’s level by 1.37 standard deviations (SDs). Among listeners who hear transmitted versions of those messages, beliefs shift by only 0.88 SDs. This indicates approximately 34% loss of sensitivity to variation in the level of the original prediction. By contrast, loss of reliability information is nearly three times as large: beliefs about the prediction’s reliability go from shifting by 1.18 SDs to shifting by 0.12 SDs, meaning 91% of the variation in information about a message’s reliability is lost in transmission.
Figure 1
In our second set of analyses, we examine listeners’ belief updates about the economic variables discussed in the recordings. Listeners who hear the original messages directly update their belief in the direction of the message’s prediction, and those who hear strong-reliability versions of a message update twice as strongly on average as those who hear weak-reliability versions. By contrast, listeners who hear transmitted versions of the messages update about the same amount on average from weak-reliability and strong-reliability messages: the distinction between weak- and strong-reliability messages is almost completely lost in transmission. Formally, we calculate that the sensitivity of listeners’ belief updates to variations in the level of the original prediction decreases by 30% as a result of transmission, but the sensitivity to variations in reliability decreases by 90%.
Figure 2
In our final set of analyses, we abstract away from belief-based measures of information loss and analyse the transcripts of transmitted messages. We document that while nearly all the transmitted messages contain some statement about the level of the original prediction, only a third mention the original prediction’s reliability. This is true even among very long transmitted messages that provide minute details about the level of the original prediction.
Figure 3
Why is reliability information more likely to be lost?
Reliability information may be selectively omitted due to a conscious decision-making process. This could occur because (1) the advantages of sharing reliability information are deemed less significant than the benefits of conveying level information, or (2) the perceived mental effort required to transmit reliability information is considered to be greater. Alternatively, the loss of reliability information might not stem from a purposeful optimisation strategy, but rather from an inadvertent process. For instance, (3) the individual might simply not think of the reliability information at the time of recording the voice message. Through a series of experimental investigations, we reject the first two explanations and provide evidence supporting the third explanation.
Implications
Our findings speak to a variety of important real-world phenomena, including viral fake news (Guriev et al. 2023, Zhuravskaya et al. 2017), persistent belief polarisation, failures of expert communication, and the persuasive power of qualitative information (Andre et al. 2021, Graeber et al. 2023). In a world where information is increasingly disseminated through social networks, both online and offline, our study highlights the need for greater awareness of the factors that influence the fidelity of information transmission. The key finding of our paper – that as a result of transmission, reliable and unreliable messages converge in influence – highlights how social transmission may partly explain the pervasive influence of inaccurate and low-quality information.
References
Andre, P, I Haaland, C Roth and J Wohlfart (2021), “Inflation Narratives”, VoxEU.org, 23 December.
Graeber, T, S Noy and C Roth (2024), “Lost in Transmission”, CEPR Discussion Paper 18771.
Graeber, T, C Roth and F Zimmermann (2023), “Stories, Statistics, and Memory”, VoxEU.org, 9 February.
Guriev, S, Henry, E, Marquis, T and E Zhuravskaya (2023), “Evaluating anti-misinformation policies on social media”, VoxEU.org, 10 December.
Hirshleifer, D (2020), “Presidential Address: Social Transmission Bias in Economics and Finance”, Journal of Finance 75(4): 1779–831.
Shiller, R J (2017), “Narrative Economics”, American Economic Review 107(4): 967–1004.
Shiller, R J (2020), Narrative Economics, Princeton University Press.
Zhuravskaya, E, S Guriev, E Henry and O D Barrera (2017), “Fake news and fact checking: Getting the facts straight may not be enough to change minds”, VoxEU.org, 2 November.
