18 Months In: Reflections and Early Insights from our PhDs with the TRUE Project

Ever since they first appeared in a Reddit sub-thread in 2017, ‘deepfakes’ - synthetic media created or manipulated using artificial intelligence (AI) – have become much more realistic-looking, more prevalent, and harder to detect. Some authors have warned of the impending “infocalypse”, bringing about an era in which seeing is no longer believing. At the same time, “user-generated evidence” (i.e., information captured by ordinary users through their personal digital devices and used in legal adjudication) is increasingly being used as evidence to pursue accountability for atrocity crimes. As we have noted, one of the biggest dangers surrounding deepfakes is not that they could end up as evidence in court, but rather that real footage might be dismissed as possibly fake. But how aware are people of deepfakes and how good are we at detecting what is authentic and what is real? How do judges, lawyers, and investigators deal with the increasing prevalence of deepfakes? Can we still trust digital evidence in an era of deepfakes?

These are the questions Ruben, Rebecca, and Anne have been exploring for the past 18 months as PhD students working on the TRUE project. We are using our linguistics, psychology, and legal backgrounds to explore the impact of deepfakes on trust in user-generated evidence. In this post, we reflect on our experience working in a multi-disciplinary team, the challenges and benefits of conducting research on a current and emerging topic, and the ethical considerations of researching deepfakes. We further present some initial findings from our first year as doctoral researchers and outline some of the key considerations with regards to user-generated evidence in an era of deepfakes.  

Doing a PhD in a multi-disciplinary research project 

Being part of a multi-disciplinary research project that combines law, psychology, and linguistics has provided us with many opportunities. In our weekly lab meetings, we get together with our TRUE team members to discuss what we are currently working on, get feedback and toy with ideas, and discuss upcoming events (such as the Advanced Advocacy Training the TRUE project will co-host in July 2025). 

Working in a team with individuals with very different research backgrounds has prompted us to learn to formulate our research in laypeople’s terms. This has been very helpful when presenting at conferences. We also started to include aspects from our respective disciplinaries in our research studies and outputs, thus opening our research up to a wider audience. Ruben’s studies on the gap between our ability to spot deepfakes and our confidence in our accuracy in spotting deepfakes have fed into Anne’s article on Deepfakes and the Law.  

The TRUE project has organised a series of events over the past 18 months, such as the TRUE Workshop on Digital Misinformation we hosted in March 2025. The connections and insights we have gained from those events, as well as the experience of organising events, have been invaluable. 

Deepfake research 

When we started our PhDs, we were apprehensive about the fact that we were researching an area that was quickly evolving, fearing that our research might be redundant by the end of our 3-year PhD program. The term ‘deepfake’ had only been termed as recent as 2017, and there was thus relatively little academic literature relevant to our research. We thus had to revisit the extant literature and draw parallels to earlier trends. For Anne, this meant drafting a chapter on the history of audio-visual manipulation, in which she mapped how the evolution of X-Rays, photography, DNA fingerprinting as evidence was always accompanied by hesitance, uncertainty, and oftentimes quickly followed by new forms of manipulation. Ruben had to draw on previous research on other forms of misinformation, also investigating the interplay between confidence and accuracy in how we form decisions when engaging with content we encounter online or in more traditional forms of media.

Working on a newly emerging phenomenon, which has been faced with much fearmongering, has also made us aware of the ethical implications of our research. Despite an increasing proportion of the public (90.4%) indicating significant concern for the dangers of deepfakes, a study released by the Turing Institute indicated a much smaller percentage reported to have been exposed to harmful deepfakes. Ideally, therefore, our research outputs will lead to a higher level of skepticism, rather than cynicism. User-generated content, whether it is used as evidence in court, for entertainment, or for information purposes, can be of great value, and should not be automatically dismissed simply because it could be manipulated. Instead, user-generated content should be approached with a healthy dose of skepticism, and we should always seek to identify the source, motives for why the content was created and shared online, and look for alternative sources to corroborate the statements made in a piece of user-generated content.  

Building on this foundation, we turn to our early research findings from our first 1,5 year as doctoral researchers and examine key considerations surrounding user-generated evidence in an era of deepfakes. 

How good are we at spotting deepfakes? 

Psychological studies examining the human ability to detect deepfakes increasingly indicate two main issues: competence and over-confidence. Not only are we bad at differentiating deepfakes from real content, but we are also falsely that we can detect the difference. As deepfake generation capabilities improve, we will become increasingly reliant on other sources to differentiate real from fake content, such as AI-detection models, which may already surpass human detection abilities. Unfortunately, we are less likely to take advice when we are overconfident in our abilities. Reducing our overconfidence may be imperative in making us less susceptible to deepfake-based misinformation.  

Our study takes the first step by investigating whether performance feedback can lead people to recalibrate their confidence level to be more in line with their detection ability. Our study corroborated previous findings indicating poor detection performance (58% accuracy) and overconfidence (75% confidence level). Providing performance feedback was ineffective in improving participants’ accuracy or reducing their overconfidence. However, feedback did lead participants to become more skeptical of future videos, where they predicted that just over half (51%) of the videos were deepfakes compared to the control, who predicted that only 40% of the videos were deepfakes. Additional findings indicated that active open-mindedness was associated with better detection and lower overconfidence, whilst the consensus formed by participants as a group outperformed their individual detection ability (82.5% vs 58%). Subsequent studies investigating whether providing crowd-sourced feedback can improve detection and how it compares to feedback from other sources (such as AI detection models or expert fact checkers) has recently been conducted and is currently in the process of data-analysis. 

The impact of deepfakes on criminal trials 

Robert Chesney and Danielle Citron have warned that, as deepfakes become more prevalent, it becomes easy to dismiss every piece of (authentic) evidence as fake. The authors call this the “liar’s dividend”. So far, there have only been a few cases in which the liar’s dividend was – albeit unsuccessfully – used. Examples include the criminal trial of Guy Reffitt, who was one of the insurrectionists during the January 6th storming of the US Capitol Hill in 2021. An FBI specialist testified for the prosecution about her analysis of the defendant’s electronic devices and was asked by the defence in cross-examination whether she had considered that the pieces of evidence could be deepfakes. She stressed that she had authenticated the evidence thoroughly and had no reason to suspect that any of the pieces of evidence were deepfakes. Further examples include the case of ‘cheerleader mom’ Raffaella Spone, falsely accused of creating deepfakes of her daughter's teammates, and a case against Tesla for the failure of a semiautomatic vehicle, where it was claimed that an interview with founder Elon Musk might have been a deepfake. 

Anne is currently working on a database of domestic and international cases featuring user-generated evidence, which will be launched on the TRUE website in June. So far, with regards to accountability procedures for mass atrocities, we have not seen the liar's dividend being employed - yet. But as we see more and more deepfakes being circulated in the context of the Ukraine/Russia, and the Gaza/Israel conflicts, the liar's dividend may well become a popular strategy in upcoming trials. 

Trust in user-generated evidence 

As part of our Mock Jury Study, a fictional criminal trial involving a (real) video of an airstrike in Yemen was recorded by a team led by award-winning documentary maker James Newton in September 2023. The video of an airstrike was posted to Twitter and verified using open-source analysis by the investigative organisation, Bellingcat, and served as the central piece of evidence to the prosecution’s case.  

In 2024, we began hosting research participants (140 in total) at Swansea University to act as mock jurors, who were invited to watch the trial proceedings and then asked to deliberate in groups of 12. The deliberations of these juries were transcribed and linguistically analysed to uncover laypeople’s perceptions of user-generated evidence. Our juries generally perceived the recorded piece of user-generated evidence to be authentic, owing to the trustworthiness and credibility of the prosecution’s expert witness’ testimony. The notion of ‘trust in experts and science’ has received increased attention within the professional contexts of jury trials and policymaking. Thus, our analyses suggest that concerns around deepfakes and manipulation were not central to juries’ evaluation of the evidence. The main point of discussion was the sufficiency and strength of this central piece of evidence to prove the defendant’s guilt beyond reasonable doubt. Despite generally accepting the expert analysis of the video, our mock jurors expressed doubts about relying on it due to the person who created the video being unavailable to testify. Therefore, the biggest concern we are seeing from this study’s data is not of deepfakes, but of the trustworthiness of the source of user-generated content. 

Can we still trust digital evidence in an era of deepfakes?  

Audio-visual manipulation is not a new phenomenon, and judges, lawyers and juries have been challenging the adage of “seeing is believing” for close to two centuries. Abraham Lincoln’s head was famously “photo-shopped” on the slightly more imposing body of Southern politician John Calhoun in the 1860s, and the first reported trial on photographic manipulation took place in 1869. Every advance in photography, film, and audio recording was quickly followed by the emergence of means and technologies to manipulate content. Yet, deepfake technology has significantly enhanced the speed and volume at which fake content can be created, and human and machine capabilities in spotting deepfakes are lagging behind. A similarly troubling phenomenon is that of misattribution – such as the video purporting to show the Israeli Embassy in Bahrain on fire or this picture claiming to show a demonstration in Canberra, Australia. 

Legal practitioners should thus start preparing for the increasing prevalence of deepfakes and misattributed content and ensure that authentication measures are in place to prevent fake evidence from being tendered. Initiatives such as the Coalition for Content Provenance and Authenticity (C2PA) are invaluable in ensuring the provenance of user-generated content by providing a way to check the provenance of audio-visual materials. Media literacy and awareness of deepfake technology and detection tools are key. In addition, civil society actors and investigators should employ robust and transparent methodologies when collecting user-generated evidence and thoroughly document the chain of custody (i.e. the documentation of how a piece of evidence was obtained and has been dealt with). Experts on deepfake technology will be in high demand, and steps should now be taken to clarify what makes an expert on user-generated evidence and deepfakes, and to train the next generation of experts. With those measures in place, user-generated evidence can continue to play an important role in criminal trials. 

The authors are Rebecca Jenkins (PhD Candidate in Applied Linguistics, Swansea University), Ruben Lamers James (PhD Candidate in Psychology, Swansea University), and Anne Hausknecht (PhD Candidate in Law, Swansea University).