– A ‘complete breakdown in trust in all media’ is at stake, as people increasingly doubt the authenticity of content whether real or not, over the risk it could be generated, according to deepfake expert Henry Ajder
– ‘Even with the current level of AI usage, we are not well-prepared for an escalation of both deceptive use of generative AI, or an escalation of claims that compromising real footage is made with AI,’ Sam Gregory, AI expert, tells Anadolu
Artificial intelligence (AI) has been all the talk since the advent of AI language model ChatGPT, which brought the technology into the mainstream.
The possibility of AI replacing human jobs one day has led to ever-rising anxiety in the workforce all around the world. Nevertheless, it is not just the average Joe’s job that might be at stake — AI may well be instrumental dethroning US President Joe Biden, or keep him in the Oval Office, considering the unprecedented speed at which the technology is advancing.
AI-based manipulation methods and deepfakes, which are footage or content that looks or sounds convincing enough to fool at least some voters, pose a substantial threat, experts speaking to Anadolu said.
Sam Gregory, an AI expert who is also the executive director of technology advocacy website witness.org, mentioned last year’s parliamentary polls in Slovakia, where he said “significant” election manipulation has been attributed to the AI-generated content.
“In some notable cases, there is attribution of significant effect, for example with a falsified audio impugning a key candidate in the Slovakian elections last year, and implying vote-rigging,” he said.
“This audio appeared during the final days before the polls, when media and politicians were not supposed to comment, and it also took time to verify that it was manipulated,” added Gregory, who suggested that the upcoming Nov. 5 US elections could also be preceded by such an “October surprise” of compromising content published on social media.
While signs of AI being used with such malicious intent certainly raises concern, deepfake audio of politicians are particularly worrying, says deepfake expert Henry Ajder.
“It’s still not entirely clear whether these deepfakes are realistic or persuasive enough to actually change voters’ opinions and change the way that they vote at the polling booth,” he told Anadolu, adding that the “changes that we’ve seen in the landscape are significant.”
‘Complete breakdown of trust in media’
Ajder raised another damning prospect of the increasingly common use of AI-based manipulation methods, referring to what he called “a complete breakdown of trust in media.”
While it remains unclear whether people are really being influenced by deepfakes, Ajder said there is “certainly a strong chance that they will be.”
“So the danger for me really is actually about a complete breakdown in trust in all media, whether it’s real or not,” he said, explaining that people are “increasingly doubting the authenticity of real media because they’re saying it could be a deepfake, it could be generated.”
“So the jury really is still out on how effective deepfakes are going to be, but we are starting to see signs of the undermining of the information ecosystem because people now know that deepfakes are possible.”
Absence of ‘strong regulations’
According to Gregory, the AI expert, the problem is compounded by a lack of proper regulations on how to deal with manipulative AI-based content.
“One notable dimension is the absence of either strong regulation to determine what is a legitimate use of synthetic content, and safeguards against malicious usage,” he said, warning that “there are no strong technical counter-measures.”
“Methods to detect deepfakes are not widely available or very reliable, and are not paired with skills and resources in the hands of most journalists, while new approaches to show more transparently how AI was used in content creation are not fully implemented across platforms and production tools,” he said.
“So, even with the current level of AI usage, we are not well-prepared for an escalation of both deceptive use of generative AI, or an escalation of claims that compromising real footage is made with AI.”
Gregory also pointed to what he called “levels of polarization, the absence of robust legislation in some states, and the inadequate technical mitigations, particularly as they are available to minority communities who are targeted with disinformation and voter suppression,” adding that these prospects “highlight the need to worry.”
Risk of being misled
At least some voters are at risk of being manipulated or misled by synthetic content made through the use of AI, with two main reasons to worry, according to the expert.
“First, most voters are seeking out content that confirms their positions, so deceptive AI content that reinforces their views or humanizing AI content that elevates their candidates will be welcome,” he said.
Gregory also emphasized that AI technology is improving exponentially over time and that it is getting equally more difficult to detect synthetic content.
“Voters are encouraged to scrutinize images, or listen closely to audio — but this is not a valid strategy as AI improves, to actually spot it. However, publicly available detection tools available via a Google search do not function well and give frequent false positives and negatives, furthering confusion and doubt,” he said.
Many others also share this fear, with Gregory noting that the World Economic Forum (WEF) had determined “misinformation and disinformation powered by AI” as the number one threat going forward.
“In reality, AI so far has been a complement to traditional campaigning and influence operations, not a replacement or had a significant multiplying effect,” he said.
For his part, Ajder underlined that “the average Internet user does really have that ability to distinguish deepfakes from authentic content,” further adding that “the technology has rapidly improved the quality” of synthetically generated voice audio.
“I can speak from my own perspective as someone who’s worked in this space for a long time, it is increasingly difficult to listen to a clip of someone allegedly speaking and to be able to say that’s definitely real or definitely fake. So, for the average person, it’s even more difficult,” he added.
Data harvesting
Another result of AI’s breakneck progress is the amount of data that this prompts tech companies to harvest, which Gregory said has reached “unprecedented” levels.
“There is an unprecedented degree of data harvesting going on right now as AI companies try to secure training data for their models,” he said.
“Similarly, many individuals are using chatbots and other AI services without thinking whether they are sharing private information that will be incorporated into the datasets,” Gregory warned.
The expert also pointed to India’s recent general elections as an example of the potential manipulative use of AI.
“Generative AI can be used to scale up direct-to-voter candidate communication,” he said, adding that it could be used to improve “covert influence operations by creating more plausible text that sounds like a native speaker and to create diverse versions to use to cover a potential campaign.”
Easier to deceive voters
When asked about the prospects of a possible orchestrated attack in the context of the upcoming US elections, Gregory referred to recent reports by Microsoft, which that AI has been used at an increasing rate for “meme and content creation.”
He said the US-based tech giant had studied the “integration of generative AI into campaigns in advance of the Taiwanese elections, as well as in the US,” also uncovering the “potential use of deceptive audio and videos dubbed deceptively and with lipsync matching.”
Pointing to research by the UK-based Center for Countering Digital Hate and rights group WITNESS, the expert emphasized “how easy it is” with AI-based tools to “create the types of images and scenes that are the center of conspiracy theories on elections and could be used for voter suppression or to confuse voters.”
“We have also seen how AI-generated images can rapidly spread and be contextualized to frame issues. For example, recent images showing former President Donald Trump appearing with Black voters that were originally created to illustrate articles or as satire but have been recycled as deceptive images implying they are real,” he said.
Gregory further added that such acts will “play on the same vulnerabilities as previous attempts, compounded by the questions of access to adequate detection tools and resources in the most vulnerable communities.”
“Although we focus most in deepfake discussions on the idea of the big, flashy fake that shifts an election, a more significant risk might be a volume of smaller-scale content that suppresses voting and reduces voter commitment and enthusiasm,” he concluded.