Manipulating Exposure to Emotional Deepfakes

Deepfakes are synthetic media manipulated or generated using artificial intelligence (AI) to appear authentic and human-like. A particular class of deepfakes involves imagery or videos of human faces with manipulated appearance and or behaviour. While numerous policy and opinion-based papers have raised concerns about the societal impact of synthetic media, particularly the risk of increased scepticism and public distrust (Twomey et al., 2023; Weikmann & Lecheler, 2023), studies directly and experimentally examining the effects of deepfake exposure remain limited and inconsistent in their findings. Some experimental work suggests that increased exposure may induce scepticism, which can result in a shift in response bias, where individuals become more prone to labelling real content as fake (a ‘false positive bias’) (Groh et al., 2022; Janssen et al., 2025; Lewis et al., 2023), a concept that has been referred to as the “liar's dividend”(Chesney, Bobby; Citron, Danielle, 2019). In contrast, others have reported a ‘false negative bias’, with prior exposure resulting in participants’ being more likely to incorrectly judge fake media as accurate (Ahmed et al., 2024; Batailler et al., 2022; Pennycook et al., 2018), a finding consistent with the “illusory truth effect”, wherein repeated statements are more likely to be judged as truthful simply because they feel familiar. By isolating the types of errors people make when detecting deepfakes i.e. false positives and negatives, and how this relates to the amount of deepfake exposure, the present study seeks to clarify the direction of exposure effects and contribute insights for countering the negative impact of misinformation. Whether exposure to deepfakes leads to general confusion, hyper-scepticism, or misplaced trust has different implications for media literacy interventions and policymaking. Our previous study highlighted the importance of emotive cues in deepfakes, revealing that our participants were impaired and biased in the deepfake detection task by deepfaked faces which had emotional, compared to neutral expressions. Given this, the study will also investigate how increased exposure and the types of detection errors participants make are impacted by emotional expressions in the stimuli. This distinction is of interest as deepfakes are often used to spread emotionally charged or divisive material, such as political misinformation. In the research literature the impact of emotion on deepfake detection is varied, with results pointing towards emotion causing an increase in false positives and false negatives. There is evidence suggesting the perceived ‘uncanniness’ of virtual characters is increased in cases of emotional expression (especially negative emotions) (Tinwell et al., 2011). Koschate et al., 2016 however found that adding emotional expressions to humanlike robots reduced how “uncanny” participant’s rated them. Valenti et al., 2023 showed that emotional expressions are not processed equally. Faces depicting happy expressions are remembered better, even under high cognitive demand/when participants are distracted, than angry expressions. This supposed difference in memory and attention for happy expressions may influence detection accuracy and in turn false negative and false positive rates. This is also suggested from Ng, 2023, who demonstrate an overly cautious/false positive effect in people more likely to report an image as ‘fake’ if they perceive the individual in the image as being more dangerous. This effect may be linked to negative emotional expressions e.g. anger may be perceived as dangerous. The current study will manipulate the amount of exposure to deepfakes experimentally, by assigning participants to either a no, low, or high exposure group, and use a deepfake detection task where the stimuli used will have both emotional and neutral expressions. We aim to investigate not only overall detection accuracy but also response bias, via the type of detection errors participants make.