Tailored Truths: Persuasive Capabilities of LLMs

Feb 11, 2025·
Jasper Timm
,
Chetan Talele
Jacob Haimes
Jacob Haimes
Abstract
Large Language Models (LLMs) are becoming increasingly persuasive, demonstrating the ability to personalize arguments in conversation with humans by leveraging their personal data. This may have serious impacts on the scale and effectiveness of disinformation campaigns. We studied the persuasiveness of LLMs in a debate setting by having humans (n=33) engage with LLM-generated arguments intended to change the human's opinion. We quantified the LLM's effect by measuring human agreement with the debate's hypothesis pre- and post-debate and analyzing both the magnitude of opinion change, as well as the likelihood of an update in the LLM's direction. We compare persuasiveness across established persuasion strategies, including personalized arguments informed by user demographics and personality, appeal to fabricated statistics, and a mixed strategy utilizing both personalized arguments and fabricated statistics. We found that static arguments generated by humans and GPT-4o-mini have comparable persuasive power. However, the LLM outperformed static human-written arguments when leveraging the mixed strategy in an interactive debate setting. This approach had a 51% chance of persuading participants to modify their initial position, compared to 32% for the static human-written arguments. Our results highlight the concerning potential for LLMs to enable inexpensive and persuasive large-scale disinformation campaigns.
Type
Publication
In AI for Social Impact: Bridging Innovations in Finance, Social Media, and Crime Prevention workshop at the 39th Annual AAAI Conference on Artificial Intelligence

Main Motivation and Research Question:

  1. Large Language Models (LLMs) can argue effectively, but do humans find these arguments persuasive?
  2. Assuming threat actors weaponize LLMs for disinformation, what are the most persuasive strategies and how do they compare to human-written arguments?
Diagram describing the process flow for each interaction recorded.

Key Findings

Why does this matter? A Path to Scalable Disinformation

For just $100, GPT-4o-mini could debate 300,000 people—equivalent to the number of undecided swing-state voters in the 2024 US Election. With this level of scalability, if LLMs are also highly persuasive, they could become an incredibly effective tool for mass disinformation.

Likert and Loaded: Measuring AI’s Persuasive Punch

Human participants engaged in multiple debates with an LLM on a given topic. To measure opinion change, we compared their initial stance to their final stance using a 7-point Likert scale (Strongly Agree to Strongly Disagree). Each round featured a different persuasion strategy (see below). In some cases, participants read a static argument instead of engaging in a debate.

Details

Topics

When deciding on debate topics we drew inspiration from the Anthropic post on Measuring the Persuasiveness of Language Models. Issues were chosen to be less polarizing, focusing on: "complex and emerging issues where people are less likely to have hardened views".
Example topics
Prescription drug importation should be allowed to increase access and lower cost.
Genetic modification of unborn babies is unethical and dangerous.
Space tourism should be limited until safety regulations are further developed.
AI must be transparent and explainable in order to be widely accepted.
Internet access should be considered a basic human right.

Interaction Types

  • Static Arguments:
    • arg-hum: Paragraph written by a human to be read by participants
    • arg-llm: Paragraph written by an LLM to be read by participants
  • Simple: Basic debate with no additional persuasion instructions.
  • Stats: LLM uses (mostly) fabricated statistics to persuade.
  • Personalized: LLM tailors responses using user demographics and personality traits.
  • Mixed: Multi-agent approach combining personalized and stats agents, with an executive agent finalizing responses.

Key Terms

  • Likert Δ: Difference in the initial and final rating on the Likert scale. Changes in the direction which the LLM was arguing for are considered positive.
    Visual representation of Likert Δ. Two likert scales are show in top of eachother with the one value highlighted each, the 2 and the 5. An annotation is provided, showing that the difference between the final value, 5, and the initial value, 2, is the Likert Δ | Timm et al.
    Visual representation of Likert Δ.
  • P(+change): The likelihood of a positive Likert Δ. A positive Likert Δ means the opinion shifted in the direction the LLM argued for.
  • EMM: Estimated Marginal Mean, essentially the average.1
  • Debating Under Influence: Mixing a Persuasion Cocktail

    Personalization alone led to a Likert Δ of 0.479, showing a modest impact on opinion shifts. The simple approach performed better, with a Likert Δ of 0.782. The statistics-based method achieved a higher Likert Δ of 0.823, outperforming both - the personalization and the simple approach. However, the mixed approach had the greatest effect, reaching a Likert Δ of 1.146. Since a 1-point shift represents a full step on the scale, this result confirms that the mixed approach outperformed all other methods individually. This suggests that the right strategy is more effective than a simple debate prompt aimed at persuading the user. Specifically, personalizing fabricated statistics makes arguments significantly more convincing than either approach alone.

    Bar plot comparing the estimated marginal means of likert delta recorded during the study | Timm et al.
    Estimated marginal means for Likert Δ.
    Bar plot comparing the estimated marginal means of the probability of positive change recorded during the study | Timm et al.
    Estimated marginal means for P(+change).

    This led to an interesting observation when comparing LLMs and humans. The arg-llm type had a higher P(+change), while arg-hum had a higher Likert ∆. This suggests that while LLMs may often sway opinions, human arguments can sometimes be significantly more persuasive.

    A fascinating—and slightly eerie—aspect of the mixed type was watching the private chat of the agents as they coordinated to generate debate responses. They categorized users by demographics and personality traits, exchanging responses and debating which arguments and fabricated statistics would be most persuasive. It felt like observing an AI focus group fine-tune the perfect pitch, adjusting strategies on the fly to maximize influence.

    Diagram depicting the process used to generate the Mixed approach responses. Three agent responses are shown in succession, each labeled with a different colored robot emoji. Ther personalized agent provvides micro-targeting information based on the users demographics. This is then provided to a statistics agent, which generates the multiple fabricates statistics relevant to the subjects suggested by the first agent. Both responses are then provided to an executive agent, which outputs a synthesized debate response | Timm et al.
    Diagram depicting the process used to generate the Mixed approach responses. The messages seen here are excerpts from one interaction recorded during our experiments.

    The challenge ahead

    The low cost and high impact of AI-driven persuasion highlights the need for safeguards. Detecting AI-generated content in conversations is tough without clear markers, so improving detection, content verification, and platform safeguards is key to preventing misuse.

    Ethical considerations

    At the conclusion of the study, participants were informed that some of the models were instructed to make up falsified statistics in order to strengthen their arguments. They were also given a recommended reading list to better inform themselves about false information on the internet.

    Citation

    @misc{timm2025tailored,
      author = {Jasper Timm and Chetan Talele and Jacob Haimes},
      title = {Tailored Truths: Optimizing LLM Persuasion with Personalization and Fabricated Statistics}
      year = {2025},
      language = {en},
      month = {jan},
      eprint = {2501.17273},
      url = {https://arxiv.org/abs/2501.17273},
    }
    

    1. EMM provides a way to interpret the effects of categorical predictors while controlling for other variables in a statistical model. For more details on EMM, check out this explainer↩︎