A new study predicts that by mid-2024, artificial intelligence campaigns by "bad actors" causing harm online by spreading disinformation will be a daily occurrence. The findings are concerning given that more than 50 countries, including the United States, will hold national elections this year and the results will have global consequences.
Even before the latest iteration of the Generative Pretrained Transformer (GPT) system was released, AI experts predicted that by 2026, 90% of online content would be computer-generated without human intervention, leading to the spread of misinformation and disinformation.
There is an assumption that the large social media platforms with the most users should be regulated to control risks. This assumption is correct to a certain extent and also makes them the target of legislation such as the EU’s Digital Services Act and Artificial Intelligence Act. However, there are other, smaller "bad actors" - individuals, groups and countries who intentionally engage in behavior that causes harm to others - who are also abusing AI.
A new study led by researchers at George Washington University (GW) provides the first quantitative scientific analysis to examine how bad actors are abusing artificial intelligence and GPT systems to cause harm on social media platforms around the world, and what can be done about them.
"Everyone talks about the dangers of artificial intelligence, but before our study, there was no science behind it," said Neil Johnson, the study's lead author. "You can't win a battle without a deep understanding of the battlefield."
The researchers began by mapping the dynamic network of interconnected social media communities that make up the global online population. Users—sometimes a few people, sometimes millions—join these communities because of shared interests, which may include harm. The researchers focused on extreme "anti-X" communities, where two or more of the 20 most recent posts from each community contained explicit hate speech and/or extreme nationalism and/or racism. These anti-X communities include anti-American, anti-women, anti-abortion or anti-Semitic communities. Over time, the connections between these communities can form community clusters within and across different social media platforms.
"Any community A can create a link (i.e., a hyperlink) to any community B, as long as B's content is of interest to members of A. The link directs the attention of members of A to B, and members of A can add comments about B without members of B knowing the link -- thus, members of community B can be exposed to, and potentially influenced by, members of community A," the researchers said.
Using mathematical models, the researchers identified possible activity between bad actors and AI and why. Specifically, they found that the most basic GPT systems, such as GPT-2, were sufficient and more likely to attract bad actors than more sophisticated versions, such as GPT-3 or -4. This is because GPT-2 can easily replicate human style and content in extreme online communities, and "bad actors" can exploit basic tools like GPT-2 to produce more inflammatory output by subtly changing the form of online queries without changing their meaning. In contrast, GPT-3 and -4 include a filter that overrides responses to potentially controversial prompts, preventing such output.
The researchers say bad actor-AI activity is likely to flourish on the online "battleground" that is the "bad actor" community and the communities they directly link to, that is, the vulnerable mainstream community. Taken together, these communities amount to an online ecosystem of over a billion people, allowing villain-actor-artificial intelligence to flourish on a global scale. The researchers cited non-AI-generated hatred and extremism related to COVID-19 and the recent Russia-Ukraine and Israel-Hamas wars to illustrate their point.
They predict that by mid-2024, bad AI behavior will become an everyday occurrence. To determine this, they used proxy data from two historically similar incidents involving manipulation of online electronic systems: the 2008 automated algorithmic attack on U.S. financial markets and the 2013 Chinese cyberattack on U.S. infrastructure. By analyzing these data sets, they inferred the attack frequency of these two incidents in the context of current advances in artificial intelligence technology.
2024 is known as "the largest election year in history." More than 50 countries, including the United States, will hold national elections this year. From Russia, Taiwan, the United Kingdom, and India to El Salvador and South Africa, the results of some elections will have global ramifications and have huge consequences for human rights, the economy, international relations, and world peace. As a result, the threat of bad actors using artificial intelligence to spread and amplify disinformation during these elections is real, researchers say.
They recommend that social media companies adopt strategies to curb disinformation, rather than removing every piece of content generated by bad actors.
Given the rapidly changing nature of artificial intelligence, the researchers raised a note of caution about their findings. Still, the research highlights some of the significant challenges posed by bad actors exploiting artificial intelligence.
The researchers said: "Due to the rapid development of technology and the rapidly changing network environment, no one can accurately predict what will happen to bad actor AI in the future, so strictly speaking, the predictions in this article are speculative. However, these predictions are quantified, testable, and generalizable, thus providing a concrete starting point for strengthening discussions on bad actor-AI policies."
The research was published in the journal PNAS Nexus.