Testing interventions for deepfake discernment: fact-checks and digital literacy tips

Studies show that individuals struggle to discern between authentic and fake images (Kasra et al. 2017; Köbis et al. 2021; Korshunov and Marcel 2020; Nightingale et al. 2017; Shin and Lee 2022; Vogler et al. 2025). Content generated with generative Artificial Intelligence (AI) tools are increasingly more realistic and easier to produce, and automated detection models lag behind in their accuracy and the pace at which they can adapt to such realistic deepfakes (e.g. Goldstein et al. 2023; Mirsky 2021; Shiohara and Yamasaki 2022). In this context, interventions targeting individuals - such as digital literacy tips about, and debunks of, deepfakes - can also help identify such realistic AI-generated videos. At the same time, these interventions can also make people more cautious and skeptical about news more generally (Hoes et al. 2024), including information and videos from reputable news organisations. In this study, we test the effect of exposure to digital literacy tips and fact-checks on individuals’ ability to distinguish between real videos and deepfakes, as well as on their trust in AI. We hypothesise that these two interventions should improve people’s ability to spot deepfakes, at the cost of making them more skeptical about other things they see online - including real content. Our assumption is that individuals might default to assuming other things are fake when prompted to think about the prevalence of gen-AI. We embed these interventions in a 2-wave panel survey in Spain (N = 1,700). Following wave 1, participants are randomly assigned to one of three treatment groups: the first exposing them to digital literacy tips, the second to fact-checks, and the third to news from reputable outlets (note that in this study, we do not formulate any hypotheses in relation to the third group, but rather register this as an exploratory analysis). We then compare treated participants’ ability to distinguish between deepfakes and real videos, as well as their trust in AI, to that of a control group. We partner with the main fact-checking (Maldita), digital literacy (LearnToCheck), and news organisations (El Mundo and El Pais) in Spain to expose participants to realistic interventions. Participants will be delivered the treatment via separate websites managed by researchers and updated three times a week. Throughout the intervention period, each participant will be exposed to up to 22 articles. Note that this study is part of a larger project in which we measure the effects of our interventions on several other outcomes: individuals’ misperceptions, skepticism, belief accuracy, trust in democratic actors, policy positions, and polarisation. You can find other pre-registrations related to this intervention in this same repository.