Limited effectiveness of psychological inoculation against misinformation in a social media feed

Abstract

Psychological inoculation is a promising and potentially scalable approach to counter misinformation. The goal of inoculation is to teach people to recognize manipulation techniques, such as emotional language, commonly found in misinformation online. While there is substantial evidence that inoculation increases technique recognition when directly assessed, it is not clear if this effect transfers to spontaneous detection of techniques and disengagement with the associated content in real-life contexts. In particular, emotional appeals are abundant on social media and known drivers of attention and engagement. Therefore, we examined the effects of emotional language and emotional manipulation inoculation on attention and engagement in a simulated social media feed environment. Through five preregistered studies, we found that inoculation only decreased engagement with emotionally presented content when we solely presented synthetic content relevant to the task of identifying emotional manipulation. Any addition of real tweets or even synthetic tweets containing other manipulation techniques (e.g. ad hominem attacks) into the feed appeared to nullify the effect. Our results highlight the importance of assessing misinformation interventions in ecologically valid contexts to estimate real-world effects.

Keywords: emotion; misinformation; psychological inoculation; psychology; social media.

Fig. 1.

Examples of manipulative and neutral synthetic tweets (top row) and real tweets containing more than 50% negative, neutral or positive sentiment (bottom row).

Fig. 2.

Regression coefficients for the interaction between inoculation and positive or negative emotional language for studies 2 and 3 (top row) in models predicting dwell time (A), likes (B), and shares (C). Regression coefficients for the interaction between inoculation and emotional manipulation for studies 4 and 5 which used synthetic stimuli (bottom row) in models predicting dwell time (D), likes (E), and shares (F). Error bars indicate 95% CI. *P < 0.05, **P < 0.01, ***P < 0.001.