Arguably, the most concerning downside of encrypted messaging applications is an inability to screen, analyze, and trace messages that may proliferate fake news. This is evidenced by WhatsApp becoming a breeding site for misinformation in its biggest user base in India. The inability of such platforms to take down incorrect information and identify or penalize its perpetrators makes it nearly impossible to control and fact-check the information being shared.
A point of interest in such a conundrum is the implication that people are highly susceptible to misinformation. This project aims to understand if incentivizing the sharing of factual information and disincentivizing the sharing of misinformation, either through micropayments or through social feedback, reduces the sharing, and potentially the spread, of misinformation. This article will summarize the study design and mainly focus on the behind-the-scenes of how the content for this study was developed.
In association with Tattle, Monk Prayogshala built a web service platform named Meshi to host the experimental study. Here, participants were asked to respond to 25 messages over a period of three days with 5 messages on the first day, and 10 messages each on the following two days. The options available for each message were sharing it, and/or reacting to it with emojis that expressed happiness, anger, and disgust. There was also a choice for participants to select a ‘read more’ option to gain more information about the message before they react to it. Last, each participant was assigned to one of two conditions in this information-sharing experiment: one where they received real money if they shared true information (and lost real money if they shared fake news) and the other where they gained or lost followers based on the veracity of information shared.
Similar to other studies of misinformation, content was formed using current or recent news. However, while other studies used real headlines as they were printed in the media, we made sure that we removed most, if not all, identifying elements across 250 headlines. For example, instead of saying TikTok is banned in India, we rephrased it to “the most popular non-gaming app is banned in India.” Identifying information was also removed from the ‘read more’ content. The motivation for doing this was to minimize any potential internet searches by participants during the experiment and maintain data accuracy.
Furthermore, in an effort to provide structure and congruence to the messages (i.e. headlines) being presented, we organized them into different types. Inspired and in sync with seminal work in the field, these were divided into the broad categories of plausible, implausible, true, false, and wholesome. The headlines chosen for the plausible category were ones that participants were more likely to be certain about (e.g. “There is a low chance of mobile phone use leading to cancer”). On the other hand, implausible category headlines were those that participants would not certainly know to be false (e.g. “There is an 87% chance that Indian women report a better quality of life than men”).
Headlines in the true and false categories respectively contained the same information as headlines in the plausible and implausible categories but without the probability or likelihood phrasing. As a result, the content for the true and false categories reads more like a traditional news headline. In the aforementioned examples, the true headline was “Using mobile phones is not associated with an increased risk of cancer” and the false headline was “Indian women report a better quality of life than men”. The rationale behind using the same basic source news but phrasing it differently was to assess if changing the certainty phrasing affects the spread of misinformation. The wholesome category was used as a control and it was not incentivized in either condition. Its headlines were mainly employed to test whether participants in the social condition were more likely to share wholesome content compared to those in the financial condition. All the messages were randomized and counterbalanced, with each participant receiving all five types of messages across the three days of the experiment.
Beyond categories, the messages were further organized by themes using Tattle’s Khoj database that displays words in IFCN-certified fact-checking articles in a particular week. After careful consideration of the overall word frequency and the number of clusters a specific word is in, some keywords were chosen to help build an overarching theme (e.g., the words “COVID-19”, “myths”, and “vaccines” led to the creation of the theme “health”). The Khoj database was also supplemented by manually tracking fact-checking websites like WebQoof, Alt News, The Logical Indian, and Boom, and keeping up-to-date with various online news media outlets for over a month. The final themes generated were health, science, technology, religion, politics, and miscellaneous (with items related to sports, entertainment, wildlife, etc.).
The content for this project was carefully chosen and all messages were related to India and were culturally relevant. To confirm that the chosen messages truly measured misinformation and not mere unawareness, we asked a few research personnel about their perception of the fake messages prior to the main study. After receiving a hit rate of 36-50%, we concluded that the content was effective enough to be perceived as either real or fake and was not lopsided in either direction. We soon started a pilot study on the Meshi platform and the data collection concluded towards the end of July 2022.
Swanaya Gurjar
This blog was originally published on the Tattle website.
