Novel media technology is changing the nature of political communication. The emergence and rapid improvement of generative AI has ushered in an era when scammers, revenge porn tormentors, the official White House Twitter account, and other cruel/stupid actors can whip up sophisticated fake images and videos.
Thanks to the vagaries of academic publishing timelines, two of my projects on this topic were published in the past few weeks — though one of the experiments was conducted in 2020, and the other in 2024. They treat video manipulation quite a bit differently, in a way that can help us understand the changing nature of digital communication.
“Political Deepfakes Are as Credible as Other Fake Media and (Sometimes) Real Media” (Journal of Politics) with my excellent co-authors Soubhik Barari and Chris Lucas. (ungated version here)
“Thirst Traps and Quick Cuts: The Effects of TikTok Edits on Evaluations of Politicians” (Social Media + Society) with my excellent co-author Valerie Li
In the former, we found that deepfakes are approximately as credible as misinformation delivered through text or audio formats, rather than being uniquely deceptive. However, in discernment tasks, subjects often mistook authentic videos for deepfakes when the videos showed politicians from their own party involved in scandals. We also found that informational interventions and accuracy primes only somewhat reduced the impact of deepfakes. Overall, while deepfakes may not be uniquely deceptive compared to other forms of misinformation, they may still contribute to eroding trust in media and increasing partisan polarization.
I think this paper does an excellent job laying out the potential pathways by which real and false videos might affect voter perception. In terms of research design, we pioneered new attention checks for audiovisual content, and used an experimental double-debrief to ensure that subjects didn’t persist in believing in the fake videos we showed them. There’s a ton of cool results in the paper, but my favorite compares the importance of partisanship in evaluating the veracity of true vs fake video clips:
There is a massive performance gap between Democrats and Republicans, but only for real videos. The top panel demonstrates that some of the fake videos were more compelling than others, but that partisanship plays only a very minor role in predicting whether people identified them correctly or not. In contrast, there are discrepancies by partisanship for real clips like the Obama hot mic and the Trump Covid announcement. Overall performance suggests that partisans fare much worse in correctly identifying real clips, but not deepfakes, portraying their own party’s elites in a scandal.
This is the first article in a “top” political science journal to study the effect of AI-manipulated video. Designing and executing this kind of study is a significant endeauvor, and passing peer review (and now, code replication checks) at an outlet like the JOP requires a lot of rigorous back and forth. And I mean a lot. We first submitted the manuscript to the JOP in September 2021.
I’ve published over a dozen papers evaluating the effects of political media on citizens’ preferences and beliefs. This one is by far the most “rigorous.” The sample size was much larger than usual, made necessary by the fact that our experimental design was especially complex in order to understand not just the effects of deepfakes but different mechanisms through which those effects might operate. The construction of the deepfakes themselves was a serious undertaking in 2020, and we collaborated with a professional video editing firm. The Table of Contents for the Appendix barely fits on a single page.
But all this rigor comes at a cost. This experiment was conducted during the 2020 Democratic primary — but was only just published. When we study contemporary phenomena, there is a tradeoff between speed and rigor, between internal validity and temporal validity. This is clearly the case in the Meta2020 collaboration. The race-to-the-most rigorous occasioned by the need to publish in top journals puts too much weight on the former form of validity. But both forms of validity are important.
There’s no easy way out of this problem. Science is hard to do both well and quickly. But I firmly believe that we can do better — in fact, that it is scientifically necessary that we do better. Citizens are now dramatically more likely to have been exposed to digitally manipulated videos than they were in 2020 — and those videos have gotten much better and much more varied.
If there had been some flaw in the randomization procedure in our experiment, our paper would have been rejected, and rightly so: the scientific value would be zero. But if we accept the importance of temporal validity, the scientific value of this paper would have been higher if it had been published in 2023 than now in 2025. The scientific community recognizes this — the working paper was cited 64 times already. I take this as further evidence that we should treat Metascience as Political Methodology: that we should insist that the macro processes of science be selected with the same rigor as we treat the micro processes.
The second paper is about TikTok thirst traps, so you can guess it’s a bit more recent. Indeed, it was conducted during the 2024 US Presidential Election, when the Democratic Party was still pretending Biden could run for re-election. So, a lifetime ago in US politics but less distinct from today in terms of media technology.
Social Media + Society is less prestigious than the Journal of Politics. It’s also, institutionally, better suited to studying digital phenomena. The journal is committed to a 30-day turnaround time between the paper is accepted and when it is published online — the second-fastest publication time I’ve ever seen, after the JQD.
The phenomenon of “TikTok Edits” was brought to our attention by media commenter Jules Terpak, in a long-form YouTube video titled “The Power of TikTok Edits”. One of Terpak’s most important observations is that the creators of these “edits” are participating in a kind of political fandom. Rather than engaging in traditional rhetoric with the goal of persuading their audience that a given politician will implement beneficial policies, their goal is immediately aesthetic: they want their favorite politician to look good. Creators of “edits” are, crucially, constrained by the video they have available—in this way, “edits” amplify the importance of existing media coverage, giving an advantage to already famous politicians.
The key argument in the paper is that for these TikTok “edits,” facticity is beside the point. Edits do not pretend to show politicians as they “are”—the proliferation of technical images of these politicians means an explosion in the points of view from which we might see them. Beyond the necessary step of attracting viewer attention, the goal of these edits is to present a specific point of view: that the politician is attractive, or intimidating, or funny. The importance of candidate attractiveness for voter evaluation grows as the capacity to communicate their attractiveness audiovisually becomes democratized and contested.
A highly edited video, splicing together video from across decades and with a sentimental or bombastic backing track obviously did not “take place.” The point of these “technical images,” following the framework of the media theory GOAT Vilém Flusser, is to produce a compelling video that demonstrates the point of view of the creator—to allow the creator to share their point of view with others. We demonstrate that these videos are effective at changing evaluations of politicians even though they are clearly neither true nor false.
We test the effects of two types of edits: thirst trap edits and badass edits. In the “thirst trap” genre, the edits make the target politician appear attractive, using soft video filters and portraying them in stereotypically attractive activities—playing sports, being surrounded by stereotypically attractive women, being rich, and winking/smiling/making other flirtatious or charming facial expressions. And since these were all older men, the thirst trap edits were used on some video clips when they were younger.
The “badass” genre, on the contrary, made the politician look intimidating or impressive, using dark video filters and dramatic audio tracks, intense cuts and visual effects, and portraying them dressed powerfully (in suits) in some official capacity. This often involved giving bombastic speeches and the inclusion of patriotic (e.g. US flags) or military symbols.
The most fun part of the paper, in my opinion, is an unusual dependent variable: politician attractiveness. First, we demonstrate significant variation in baseline levels of attractiveness rating, with Trump by far the least attractive and RFK the most attractive.
We demonstrate that the “thirst trap” edits are able to significantly increase evaluations of attractiveness across the board.
Given that these candidates were older White men who had been covered extensively in the media already, we believe that these results represent a lower bound compared to a diversifying population of political candidates. For non-White, female, and less prominent politicians, appearance should play more of a role—and the ability to effectively manipulate appearance will therefore play a larger role in the practice of political campaigning.
We expect this to be a fertile area for future research. The scope of the “political” that can be expressed with video edits goes beyond active political campaigns. There are a variety of social and cultural topics displayed in this kind of video, which we expect to play a growing role in many areas of 21st-century communication.