AI and the Election

Types of Misuse

A major aim of research for SCS identifies the different ways generative AI can be misused, Heidari said. “This technology has been made publicly available without adequate guardrails,” she said. “We have already seen numerous examples of how it can be used in malicious
or otherwise questionable ways.”

It’s important to recognize that AI itself is not harmful, Carley explained; but people can use AI in harmful ways. Researchers categorize the misuse of generative AI into two main areas. The first is disinformation, which refers to inaccurate information purposely spread by people or robots with the intent to mislead. Secondly, there is misinformation, which is also inaccurate, but the sender doesn’t realize it and they spread these falsehoods without malicious intent.

A special class of disinformation known as deepfakes involves content that looks and sounds real, but has been completely generated by a computer. Deepfakes come in the form of audio, video, images, memes and even news stories that contain disinformation but appear to be completely true. Researchers have found evidence that online actors are using AI, sometimes in the form of bots, to spread disinformation, including deepfakes, on a daily basis. For example, clear data shows that foreign entities are using generative AI on quasi-news and social media platforms to influence U.S. voters.

Carley, whose research combines cognitive science, social networks and computer science, is currently conducting a global investigation to analyze a type of deepfake called “pink slime,” or hyper-localized fake news websites that contain completely made-up stories.

“With the demise of many local newspapers, we’ve seen a sharp increase in these fake local news sites,” said Carley.

“You will see news websites in two completely different towns that use the exact same story and words, but just change the photos and quotes as if they came from local officials. Our project is trying to identify these sites and then quantify how much misinformation they are spreading.

“Unfortunately, we have seen them have an impact in getting people to donate money to fake causes, and also in getting people to believe a candidate has more support than they actually do,” she said.

This project is just one example of how CMU researchers are working to identify and quantify the effects of AI misuse — an important first step to ensuring that AI is used fairly and productively.

Will AI Affect Election Results?

Despite specific examples of misuse, it is extremely difficult to quantify AI’s overall effect on elections and the democratic process because currently there is no concrete way to measure its cumulative impact over time.

“If you pick out any one instance of GenAI misuse, one could argue it is unlikely to have changed the outcome of the election,” Heidari explained. “But over time and as instances accumulate, the technology can contribute to the erosion of trust in the democratic process. We may get to a point where many voters won’t believe what they see or hear, simply relying on their gut feelings and emotions, instead of facts and reason, to decide how to vote.”

“Ideally, we want the democratic process to be a rigorous debate and exchange of information about candidates and their policies, rather than an uninformed majority vote.”

Heidari compares the impact of generative AI on elections to how social media has led to declining mental health of adolescents over the past decade. “In both cases, we are dealing with diffuse harm,” she said.“ With this type of harm, it’s difficult to pinpoint whether any one interaction has tipped the scale. There aren’t many identifiable victims, yet over time, we observe a huge impact.”

Broadly, Heidari’s research addresses issues of accountability and governance of AI, and specifically how to evaluate the negative impacts of AI tools on individuals and society. For instance, she studies a technique called red-teaming for GenAI, which involves stress-testing the model to assess the risk of it producing problematic content, such as misleading information about the electoral process.

“The results of the risk assessment can then be utilized to decide, ‘How can I prevent harms by adding appropriate guardrails?’” Heidari said. These guardrails can be technological, such as prompt-based filters, or policy-based, such as guidelines and regulations surrounding the use of GenAI in elections.

Both are important in ensuring the responsible use of AI, said Heidari. “Both are important, and they work in complementary ways,” she said.

Generative AI for Good

Although there is ample potential for the misuse of generative AI, researchers agree it’s critically important to highlight the positive uses for the technology as well.

“The goal is to create robots that are actually helpful to society,” said Yonatan Bisk, assistant professor of computer science at the Language Technologies Institute. Bisk’s research focuses on how robots use language to communicate with humans.

There are many examples of generative AI content used for good, such as alerting communities when a wildfire is approaching. In elections, generative AI can help candidates reach voters who might otherwise be disenfranchised by translating campaign messages into different languages or condensing policy proposals into summaries that are easier to understand.

“The problem is, we’re not seeing it used for good as much,” Carley said. “I know a lot of political consultants are afraid to use generative AI because they view it as complicated, and they don’t want to get it wrong.”

Carly explained that this hesitancy tips the scales in the wrong direction because bad actors are using AI freely.

“For researchers, the question becomes, what brakes can we create that help people to use this technology in good ways?” — Yonatin Bisk

To get there, technology creators must first understand all of the applications — both good and bad — for the technologies they create. According to Bisk, this often happens through a process called threat-modeling, where researchers try to imagine unintended and malicious uses for the systems they build.

Understanding the potential threats of AI systems is an important first step in regulating them, said Bisk. “The interesting questions are who is inventing, who is legislating, who is regulating that space? My biggest goal is to ensure that our elected officials are well-educated on the implications of the projects they are funding, and are constructing the appropriate policies to regulate these systems.”