AI Is Becoming More Persuasive Than Humans

Source: Hana/Unsplash

Two radically different means of personalization have claimed two more radically different scalps, from an AI-driven lab experiment to the real world of British electoral politics.

In the lab, a new study showcases the ability of AI-driven large language models (LLMs) to personalize their arguments and persuade. Salvi, Ribeiro, Gallotti, and West (2024), in a working paper, report that a personalizing LLM was significantly more persuasive than humans within an online setting, by more than 80% (p<0.01). That is, when faced with an LLM that has access to demographic information allowing it to personalize its argument, humans are 81.7% more likely to agree with the arguments when compared with a human adversary. The study cites a number of published experiments from last year that have reported that LLMs were as able as human participants, and even professional propagandists, to write persuasive text (Bai et al., 2023; Palmer & Spirling, 2023; Goldstein et al., 2023; Karinshak et al., 2023).

The discussion topics were purposely chosen to be accessible, and participants were selected into either a pro or against condition, detracting a little from the study's real-world validity. Topics included the ethics of research on animals and race as a factor in college admissions, however, so they were by no means topics removed from real-world debates. In the anonymized conditions, either humans or the LLM had access to demographic data on the person they were debating.

Personalizing to Win Elections

Far from the Swiss labs that hosted the experiment in question, the real-world implications of personalizing were being demonstrated in the English town of Rochdale. In the midst of a by-election, former Labour MP George Galloway sent certain areas of his new constituency letters that emphasized certain foreign policy stances, while sending other areas letters focused instead on social values. This was especially controversial, as some commentators felt one was a message for the Muslim community, while the other was for the White community. Despite how coarse and rudimentary the "technology" that was deployed, Galloway won his race. While the extent to which his dual-messaging system contributed to his victory is extremely debatable, his win is not an isolated example of the power of effective and varied messaging. Other instances of this kind of political maneuvering include Cambridge Analytica's infamous involvement in the 2016 U.S. presidential election. (Less discussed is Barack Obama's apparent use of personalizing in his own path to reelection in 2012.) In these examples, we find an application to the real world that might elude those only reading research papers.

Unknown Horizons

Salvi, Ribeiro, Gallotti, and West (2024) note on a number of occasions their concerns with these tools, and their potential to be deployed maleficently. A storm continues to brew, amidst increasingly dextrous LLMs with access to evermore amounts of personal data; more human behavior taking place within online spaces; and increasing polarization around key political and societal issues. We might have a sense that these developments will continue to shape a broad array of human behavior, including voting and consumption behaviors, but we do not yet know just how radical they will be as factors for change. Perhaps we cannot know, given they are likely to be AI-led and increasingly divorced from human cognition at an impending point in time.