Misinformation interventions and online sharing behaviour: lessons learned from two pre-registered field studies.

The spread of misinformation on social media continues to pose challenges. While prior research has shown some success in reducing susceptibility to misinformation at scale, how individual-level interventions impact the quality of content shared on social networks remains understudied. Across two pre-registered longitudinal studies, we ran two Twitter/X ad campaigns, targeting a total of 967 640 Twitter/X users with either a previously validated 'inoculation' video about emotional manipulation or a control video. We hypothesized that Twitter/X users who saw the inoculation video would engage less with negative-emotional content and share less content from unreliable sources. We do not find evidence for our hypotheses, observing no meaningful changes in posting or retweeting post-intervention. Our findings are most likely compromised by Twitter/X's 'fuzzy matching' policy, which introduced substantial noise in our data (approx. 7.5% of targeted individuals were actually exposed to the intervention). Our findings are thus probably the result of treatment non-compliance rather than 'true' null effects. Importantly, we also demonstrate that different statistical analyses and time windows (looking at the intervention's effects over 1 h versus 6 h or 24 h, etc.) can yield different and even opposite significant effects, highlighting the risk of interpreting noise from field studies as signal.