AI can be more persuasive than humans themselves

AI can be more persuasive than humans themselves! An experiment proved this. Two radically different means of personalization have achieved equally different results, from a laboratory experiment driven by artificial intelligence to the real world of British electoral politics.

In the laboratory, a new study highlights the ability to large-scale language models (LLMs) powered by AI to personalize your arguments and persuade. Salvi, Ribeiro, Gallotti, and West (2024), in a working paper, report that a personalized LLM was significantly more persuasive than humans in an online environment, by over 80% (p<0.01). In other words, present a difference when compared to an LLM that has access to demographic information. Thus, it allows that when LLM personalizes its argument, humans are 81.7% more likely to agree with the arguments compared to a human adversary.

Thus, the study cites a series of experiments published last year. So much so that they reported that LLMs were as capable as human participants, and even professional propagandists, of writing persuasive texts (Bai et al., 2023; Palmer & Spirling, 2023; Goldstein et al., 2023; Karinshak et al. , 2023).

Discussion topics are purposely formatted to be accessible. So participants were selected in pro or con conditions, somewhat decreasing the study’s real-world validity. Thus, topics included the ethics of animal research and race as a factor in college admissions, however. So by no means were they topics removed from real-world debates. In the anonymized conditions, humans or the LLM had access to demographic data about 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 personalization were being demonstrated in the English city of Rochdale. In the midst of a by-election, former Labor MP George Galloway sent letters to certain areas of his new constituency that emphasized certain foreign policy stances. On the other hand, he sent letters to other areas focused on social values. This was especially controversial as some commentators felt that one message was for the Muslim community while the other was for the white community.

Despite how rudimentary and crude the “technology” implemented was, Galloway won his race. So his victory is not an isolated example of the power of effective and varied messaging. In fact, even though it is extremely debatable how much his dual messaging system contributed to his victory. Other examples of this type of political maneuvering include the infamous involvement of Cambridge Analytica in the presidential election of USA in 2016. (Less discussed is the apparent use of personalization by Barack Obama on his own path to re-election in 2012). In these examples, we find a real-world application that may escape those who only read research articles.

Unknown Horizons

Salvi, Ribeiro, Gallotti and West (2024) note on several occasions their concerns with these tools and their potential for their harmful use. A storm continues to brew, amid increasingly agile LLMs with access to ever-increasing amounts of personal data; more human behavior occurring in online spaces; and growing polarization around key political and social issues.

We may have an idea that these developments will continue to shape a wide range of human behaviors, including voting and consumption behaviors, but we do not yet know how radical they will be as drivers of change. We may not know, given that they will likely be under AI leadership and increasingly divorced from human cognition at an imminent point in time.