Audio-based misinformation and radicalization are topics of increasing concern in digital environments (Muhammed T & Mathew, 2022; Wirtschafter 2023), where lies spread faster than truth (Vosoughi et al., 2018) and can cause serious harm. While there is a canon of new research on media misinformation interventions, audio misinformation and polarization interventions have not been sufficiently studied. Pew Research has found many Americans listen to and get information from audio sources, including podcasts which have gained popularity over the previous decade, and terrestrial radio which has continued to reach large audiences (Jurkowitz, 2023). Research on podcasts in particular has found high trust in hosts and advertisements (Brinson & Lemon, 2023), and significant spread of misinformation. The affordances of audio-centric content are well-suited for manipulation: voice-based misinformation and radicalization in Lebanon were found to exploit the medium’s conveying of emotion and speaker source credibility (El Masri et al 2022). Of the scant research on audio misinformation interventions, one salient study successfully deployed audio dramas designed by researchers to debunk common health misconceptions among listeners in Sierra Leone (Winters et al 2021).

One strategy of increasing interest among misinformation interventions is prebunking, or the act of trying to build mental defences against manipulation techniques before problematic content is seen or consumed. Prebunking typically involves presenting consumers with a pre-emptive warning about the tactics used in misinformation, coupled with a refutation before they encounter the false information itself (Roozenbeek & van der Linden 2019). In some studies, prebunking has been found to be significantly more effective than debunking (Banas & Miller 2013; Lewandowsky & van der Linden, 2021). A foundation of prebunking is McGuire’s (1961) Inoculation Theory, positing that exposing individuals to a weakened form of an argument along with refutations can build cognitive resistance against misinformation (McGuire 1964).

Our current work focuses on prebunking because this approach has shown promise in the current media landscape, where rapid misinformation dissemination often outpaces the ability to correct falsehoods effectively. To create prebunking messages, members of our research team drew on our backgrounds in audio production and theatre including sketch and improvisation (improv) to script, perform, record, and edit short audio sketches designed to be played as ads during podcasts. In acknowledgment of the growth of video podcasting and of limitations in our capacity to create video ads, we also used AI to add visualization to the ads we created. Here we report on our ad creation processes and our experimental pilot study conducted to gauge one ad’s influence on information literacy.

The five ads we created all humorously dramatized situations in which people interact when misinformation or manipulation is involved, resolving when the problem is managed civilly. All ads end with a reference to our project website to learn more. Our first ad took three months to create, after which we began creating ads in one month or less in overlapping phases. Our ad creation process included steps described in this section.

We designed media to be 60 to 90 seconds in length, to optimize audience attention and allow for eventual podcast ad airing at reasonable cost. Production began with script drafting, followed by recording and editing.

Ads created in this project are designed to be aired during podcasts, particularly those known to spread misinformation. Our work focused on audio prebunking is modelled in part on other campaigns designed to intervene in misinformation spread. Previously researched prebunking video campaigns have been scaled up and distributed broadly to mitigate the harms of misinformation in locations including Europe (Lewandowski and Van der linden, 2021) and India (Parihar et al, 2024). While our campaigns on audio misinformation interventions could be similarly scaled up with corporate partnership, we intend at the time of this writing to scale up more modestly, through modeling collaborative and user-based ad creation in workshops and educational settings, and through production accelerated using AI assistance.

The primary goal of our study was to evaluate the effectiveness of our prebunking ads in reducing susceptibility to misinformation, increasing the ability to identify false news headlines, and influencing attitudes towards civil discourse. We formulated specific research questions to guide our inquiry thus far:

Will exposure to a specific prebunking ad enhance an individual’s ability to identify false news headlines?

These questions framed our study design, which involved a controlled experimental setup with undergraduate students at a large university in the southwestern US as participants.

Participants who followed the invitation link provided through the recruitment materials first completed a pre-screening survey to assess their eligibility. Eligible participants were then invited to participate in Part 1 of the study, which involved providing informed consent, demographic information, and details about their social media practices.

In addition to answering questions about themselves and their social media use, participants completed qualitative questionnaires and three scales: the generic conspiracist belief scale (Brotherton et al., 2013), the misinformation susceptibility test or MIST-20 (Maertens et al., 2024). We also developed and applied our own misinformation susceptibility test based around the model of the MIST-20, called ITMIST, using different real and fake headlines.

Participants were then exposed to one of the prebunking ads or a control ad, in a format of either audio-only or an audiogram with AI-generated visualizations. Following ad exposure, participants completed a survey response questionnaire, marking the end of Part 1. The next day, participants received a link to Part 2 of the study, where they were asked to take the GCBS, MIST-20, and ITMIST scales and completed another qualitative questionnaire within one week of their completion of the first survey.

143 respondents filled out Part 1 of the survey; of these, 99 went on to complete Part 2 of the survey within one week as was required. Participants were predominantly young adults, with ages ranging between 18 and 48. 59 identified as female (59.6%), 38 as male (38.4%), 2 preferred not to disclose their gender (2.1%), and none chose other gender options. 54 identified as White/Caucasian (54.5%), 9 as Hispanic (9.1%), 8 as Asian/Pacific Islander (8.1%), 5 as Black/African American (5.1%), 2 as American Indian/Alaskan Native (2.0%), 19 as multiracial or from other backgrounds (19.2%), and 2 preferred not to disclose their ethnicity (2.0%). Politically, 34 participants identified as Democrats (34.4%), 32 as Republicans (32.3%), 18 as Independents (18.2%), and 15 affiliated with other parties or chose not to disclose political preference (15.6%). In terms of religious affiliation, 37 identified as Christian (37.4%), 16 as non-Religious but Spiritual (16.2%), 11 as Roman Catholic (11.1%), 6 as Jewish, 6 as Atheist (6.1% each), 5 as Muslim (5.1%), 3 as Agnostic, 3 as Protestant (3% each), 5 as another affiliation (5.1%), and 6 preferred not to say (6.1%).

100% of respondents stated that they used a social media platform at least once a week. Of the major platforms, 89 respondents use Instagram (89.9%), 67 use Tik Tok (67.6%), 66 use Snapchat (66.7%), 34 use Twitter/X (34.3%), 21 use Facebook (21.2%), and 8 use other platforms (8.1%). In terms of average daily time spent on social media, 43 respondents reported usage over 2 hours a day (43.4%), 26 between 90 minutes and 2 hours (26.3%), 12 between 60-90 minutes (12.1%), 14 between 30-60 minutes (14.1%), and 4 less than 30 minutes (4.0%). When asked how much of their information on current events/politics/news respondents get from social media, 28 said Most (28.3%), 67 said Some (67.7%), and 4 said None (4.0%).

For podcast engagement, 23 respondents said they frequently tune in to podcasts (23.2%), 50 said sometimes (50.5%), and 26 said never (26.3%). Of respondents who listen to podcasts, 52 reported they always skip podcast ads (71.2%), 19 said they sometimes skip them (26.0%), and 1 said they never skip them (1.4%). There was a wide range of podcasts that respondents said they listened to for entertainment and education, and clustered most strongly by political affiliation. Of respondents who identified as Democrats, the most popular podcasts for entertainment were ‘Call Her Daddy’ (CHD) and ‘Cancelled’, while the most popular answer for educational podcasts was ‘None’. Of respondents who identified as Republicans, ‘The Joe Rogan Experience’ (JRE) was the most popular podcast both for entertainment and for education. Of Independents and those who preferred not to specify their political affiliation, the most popular entertainment podcasts were JRE and CHD, and for education the most common response was ‘None’ followed by JRE.

Between the Visual Control group and the Visual Experimental group, the Experimental group showed a significantly larger average improvement on the MIST-20 (Visual Control x̄ = 0.33, Visual Experimental x = 0.93), and a moderate improvement on the ITMIST (Visual Control x̄ = 0.81, Visual Experimental x̄ = 0.98).

Between the Audio Control group and the Audio Experimental group, there were more mixed results. The Control group showed higher improvement than the Experimental group on the MIST- 20 (Audio Control x̄ = 1.41, Audio Experimental x̄ = 0.85), but lower improvement than the Experimental group on the ITMIST (Audio Control x̄ = 0.45, Audio Experimental x̄ = 0.78).