This article is part of a series titled “Machines and Votes: How artificial intelligence is reshaping elections abroad.”
Have you ever encountered a manipulated image? More importantly, can you distinguish between AI-generated photographs, audio clips, and videos and authentic ones?
As more than 2 billion individuals across 50 nations gear up for regional elections in 2024, the issue of deepfakes’ potential to sway voters has never been more critical.
Recent months have seen a surge in the identification of purported AI-generated deepfake content—images, audio, and videos—on X (formerly Twitter), as reported by the Brookings Institution’s review of the platform’s community notes, a collaborative fact-checking effort.
POLITICO took it upon itself to investigate.
They amassed a collection of genuine images alongside those created by artificial intelligence via Midjourney, an AI research lab renowned for generating lifelike images based on user input. According to tech company representatives interviewed by POLITICO, global studies have consistently shown that individuals can distinguish digital fabrications from authentic images about 60% of the time on average.
The potential to create highly realistic visuals with minimal effort using AI tools has raised concerns among politicians, policymakers, and propaganda analysts, despite the technology’s ongoing development.
With just a laptop, internet connection, and $50, anyone equipped with artificial intelligence (AI) capabilities could inundate social media platforms with deepfake political content in the upcoming months.
How adept are you at discerning these creations? Test your skills before delving further into the narrative. (Spoilers below!)
Seeking Copies?
Among the algorithmic threats looming this season, cybersecurity and propaganda experts are particularly wary of audio manipulation.
While most contentious AI-generated images are swiftly debunked within hours, thanks to the rapid error-spotting enabled by social media crowdsourcing, audio—especially the AI-generated distorted videos aimed at defaming British Labour Party Leader Keir Starmer—remains largely unexplored territory. The auditory deception created by algorithmic manipulation can easily deceive individuals due to the disparity between what is heard and what is seen.
POLITICO conducted experiments using readily available technologies costing less than $50 to explore this phenomenon. Initially planning to mimic real politicians, they opted instead to emulate the accents of POLITICO journalists, deeming falsehoods constitutionally dubious and a direct threat to the current election cycle.
You can assess for yourself the persuasiveness of these AI-generated videos.
THE REAL MARK SCOTT
Mark Scott serves as POLITICO’s Chief Technology Correspondent
THE REAL AOIFE WHITE
Aoife White holds the position of POLITICO’s Technology and Competition Editor
THE OTHER MARK SCOTT
Mark Scott is also POLITICO’s Chief Technology Correspondent
THE OTHER AOIFE WHITE
Aoife White also serves as POLITICO’s Technology and Competition Editor
AI Biden vs. AI Trump
The realm of video content, particularly interactive material, represents the next frontier in AI deepfakes.
In the domain of politically motivated AI-generated imagery, an old Soviet-era wall near the German-Polish border has emerged as a significant site showcasing the evolution of this technology.
Within this location, a group of activists known as the Singularity Group has orchestrated an ongoing, real-time online video debate between an AI-generated Joe Biden and an AI-crafted Donald Trump.
The project, operational for nearly nine months, leverages open-source technology, enabling individuals to input debate questions via the Amazon-owned streaming platform Twitch. Subsequently, the Biden/Trump bots engage, process the queries using Singularity’s AI systems, and respond, mimicking the politicians’ voices and appearances.
“Deepfakes pose a genuine threat,” remarked Reese Leysen, one of the project’s initiators, emphasizing its classification as parody on Twitch. The focus on politicians was intended to capture public attention.”
The fabricated Biden and Trump were confronted with various real-world debate questions by POLITICO. Given the AI system’s exposure to internet users’ inquiries over almost a year, many questions proved either too explicit or profane for publication.
Below are the two least explicit videos. Is the technology flawless? Certainly not. However, it provides a glimpse into the future trajectory.
Interacting with the Bots
We subjected fictitious Joe Biden and Donald Trump to real-world debate scenarios. Here’s how they responded.
Quick question No. 1: Which Disney character best embodies your political adversary, and why?
Quick question No. 2: If you were to emerge victorious in the November election, how would you address the Russia-Ukraine conflict?
