While generative artificial intelligence (AI) has opened a wealth of creative possibilities, its ability to produce convincing misinformation has made it harder to discern artificial content from real content (Park, 2024). In recent times, misinformation has taken the form of deepfakes – artificially generated videos that manipulate individuals to appear to say or do things they never did (Somers, 2020). Deepfakes have gained notoriety in the public sphere primarily for their use in generating pornographic content, and more recently, to erode trust by creating high-quality, fabricated videos that falsely depict influential figures making controversial statements.

Existing studies (Heidari et al., 2024; Jung et al., 2020; Pan et al., 2020) on deepfake detection are primarily focused on those involving deep learning. Few studies investigate human deepfake detection strategies. Among those that do, they are often limited in terms of sample size and generalisability due to the use of qualitative data collection methods (Goh et al., 2022; Zeng et al., 2023). Videos used in these studies also lack diversity and consist largely of entertainment or political videos (Goh, 2024), which may not sufficiently encompass the genres of videos that participants encounter day-to-day.

This study addresses these gaps by determining human ability to identify deepfake videos through an online quantitative survey. The first objective ascertains deepfake and real video identification performance; and the second examines the strategies that people employ to identify deepfake and real videos.

Deepfakes pose significant threats to society, particularly through their potential to create fabricated videos for spreading misinformation. The misuse of deepfakes risks undermining public trust and has the potential to deepen social divisions (Westerlund, 2019). The impact of deepfakes is further amplified when considering that individuals generally perceive video content as more credible than textual information (Sundar et al., 2021), making the spread of falsehoods through deepfakes particularly damaging.

Given the rise in deepfakes (Sumsub, 2023), there is a wealth of research on deepfake detection using deep learning models (El-Gayar et al., 2024; Lee et al., 2023). These approaches enable the rapid processing of large volumes of video data while providing objective assessments (Heidari et al., 2024), which may explain the heightened research interest over human detection capabilities.

While there is a growing body of work on human deepfake detection strategies, data collection methods are often limited to exploratory and qualitative modes such as interviews (Goh, 2024) or diary studies (Zeng et al., 2023). Despite yielding detailed insights, they are constrained by small sample sizes and limited generalisability. Findings are also often limited to the most frequently used strategies regardless of video type. For instance, in Goh (2024), strategies for identifying video authenticity are described without differentiation between methods used for detecting deepfakes or real videos. This approach limits the ability to understand which specific strategies are most effective for each type of video.

Our study utilized an online survey. Participants were shown four videos — two real and two deepfake — randomly selected from a pool of ten authentic and ten deepfake videos. These videos were publicly available from the Web and covered a range of topics, including entertainment, politics, education, and sports. Video descriptions can be found in Table 1.

After viewing each video, participants were asked to: (1) identify whether the video was real or deepfake, (2) report their confidence level, and (3) select the strategies they used to arrive at their decision. The strategies were categorized into three types: visual (e.g., facial features, background, and environmental issues), auditory (e.g., vocal features, sound quality), and knowledge-based (e.g., online tools, knowledge of video subject). These strategies were adapted from prior research (Goh et al., 2022; Zeng et al., 2023; Goh, 2024). Participants could select multiple strategies.

Young adults aged 21 to 35 were recruited as they represent a group that is active online and would likely have experience with deepfake videos (Petrosyan, 2024). A total of 195 participants were recruited via convenience and snowball sampling.

Overall, participants demonstrated an ability to differentiate between deepfake and authentic videos. When examining performance based on authenticity type, participants exhibited stronger proficiency in detecting real videos over deepfakes.

Across all videos watched, participants used a variety of strategies. This highlights the difficulty in human detection as there is no singular strategy that individuals can rely on to accurately distinguish deepfakes from real videos (Goh, 2024; Groh, 2020). This finding thus highlights the importance of media literacy and the use of multiple methods to ascertain the authenticity of videos.

Participants largely focused on the subjects in the video, as seen by facial features, and physical and behavioural characteristics emerging as frequently used strategies. In contrast, strategies that considered general video and audio attributes such as production quality were less frequently utilised. This may stem from the understanding that the subject is often the main focus of deepfake manipulation (Sundar et al., 2021), prompting participants to ignore other details (Groh et al., 2021). The danger however is that as the quality of deepfake manipulation increases, people relying primarily on such strategies may fail to spot falsified content.

The strategies used for identifying deepfake videos differed greatly from those used for image and text identification. A study on deepfake images found that participants often relied on peripheral features, such as accessories or clothing texture, to identify manipulated images (Bray et al., 2023). Similarly, strategies for evaluating the credibility of blog posts emphasised the importance of arrangement and alignment features (Jo et al., 2019). In contrast, deepfake video detection requires the analysis of dynamic components, such as facial expressions and body movement, which adds a layer of complexity compared to static media types. This distinction illustrates the need for a more nuanced approach to video detection, focusing on dynamic strategies that are not as relevant in text or image analysis.

Vocal features were the most used strategy in the identification of real videos. This suggests that participants perceive the human voice as a convincing feature associated with real videos. The complexity of the human voice and the difficulty of deepfakes to replicate these complexities makes vocal features a critical differentiator (Kulangareth et al., 2024), aiding participants in distinguishing between real and fabricated content. However as with the concern about visual features, deepfake technology is increasingly able to accurately clone voices (Mai et al., 2023). Reliance on vocal features only would again pose misidentification dangers.

Notably, regardless of video type and the accuracy of identification, the two least used strategies were employing multiple sources for verification and communication with others. Preference for individual decision-making may be explained by confidence in individual ability to identify deepfakes (Köbis et al, 2021). However, as pointed out by Goh (2024), the use of more cognitively demanding strategies such as referencing multiple sources increases the likelihood of correct identifications. This is especially important due to the rapid advances in deepfake generation technology.

This study reveals the strategies used in deepfake video identification, addressing a gap in the current literature. The findings underscore the complexity of detecting deepfakes, as evidenced by the array of identification methods utilized.

Our study offers theoretical contributions to the field of human deepfake detection and more generally, information credibility assessment, by revealing the strategies young adults use. In contrast to existing research which often reported on most used strategies (Goh et al., 2022; Zeng et al., 2023; Goh, 2024), our larger participant sample allowed for a detailed analysis of both most and least used ones. The exploration of lesser-preferred strategies offers a more comprehensive understanding of identification approaches and their potential shortcomings. Our findings provide a nuanced, multi-dimensional perspective that enhances existing literature on deepfake detection.

In terms of practical implications, our findings can inform educators in the development of media literacy curricula aimed at enhancing digital wellness and safety. Additionally, authorities and online platforms can use these insights to craft targeted strategies for combating deepfakes, including public service announcements, improving resilience against misinformation.

Despite the insights uncovered, several limitations should be acknowledged. First, the strategies recorded in the survey may not reflect the full suite of those employed by individuals in real world scenarios. Participants were informed that they had the option to rewatch videos multiple times, potentially leading to heightened scrutiny than typical online encounters, inflating the accuracy rates. Future research should aim to capture more naturalistic viewing behaviours. Second, as deepfakes become increasingly sophisticated, new detection strategies may emerge while existing ones may become less effective. Ongoing research is thus essential in providing insights into these evolving methods, ensuring that detection strategies remain relevant and effective in countering ever advancing deepfake technologies.