Measuring the persuasiveness of language models

While people have long questioned whether AI models may, at some point, become as persuasive as humans in changing people's minds, there has been limited empirical research into the relationship between model scale and the degree of persuasiveness across model outputs. To address this, we developed a basic method to measure persuasiveness, and used it to compare a variety of Anthropic models across three different generations (Claude 1, 2, and 3), and two classes of models (compact models that are smaller, faster, and more cost-effective, and frontier models that are larger and more capable).

Within each class of models (compact and frontier), we find a clear scaling trend across model generations: each successive model generation is rated to be more persuasive than the previous . We also find that our latest and most capable model, Claude 3 Opus, produces arguments that don't statistically differ in their persuasiveness compared to arguments written by humans (Figure 1) .

We study persuasion because it is a general skill which is used widely within the world—companies try to persuade people to buy products, healthcare providers try to persuade people to make healthier lifestyle changes, and politicians try to persuade people to support their policies and vote for them. Developing ways to measure the persuasive capabilities of AI models is important because it serves as a proxy measure of how well AI models can match human skill in an important domain, and because persuasion may ultimately be tied to certain kinds of misuse, such as using AI to generate disinformation, or persuading people to take actions against their own interests.

In our analysis, we primarily focused on complex and emerging issues where people are less likely to have hardened views, such as online content moderation, ethical guidelines for space exploration, and the appropriate use of AI-generated content. We hypothesized that people's opinions on these topics might be more malleable and susceptible to persuasion because there is less public discourse and people may have less established views.¹ In contrast, opinions on controversial issues that are frequently discussed and highly polarized tend to be more deeply entrenched, potentially reducing the effects of persuasive arguments. We curated 28 topics, along with supporting and opposing claims for each one, resulting in a total of 56 opinionated claims (Figure 2).

The persuasiveness of language models raise legitimate societal concerns around safe deployment and potential misuse. The ability to assess and quantify these risks is crucial for developing responsible safeguards. However, studying some of these risks is in itself an ethical challenge. For instance, to study persuasion “in the wild” we (or others) might need to experiment with scenarios such as AI-generated disinformation campaigns, but this would present unacceptably dangerous and unethical risks of real-world harm.

Though our findings alone cannot perfectly mirror real-world persuasion, they highlight the importance of developing effective evaluation techniques, system safeguards, and ethical deployment guidelines to prevent potential misuse.

To assess the persuasiveness of the arguments, we measured the shift in people’s stances between their initial view on a particular claim and their view after reading arguments written by either humans or the AI models. Participants were shown one of the claims without an accompanying argument and asked to report their initial level of support for the claim on a 1-7 Likert scale (1: completely oppose, 7: completely support). They were then shown an argument in support of that claim, constructed by either a human or an AI model, and asked to rate their stance on the original claim once again.⁴

We define the persuasiveness metric as the difference between the final and initial support scores, reflecting shifts towards greater or reduced support for the presented claim. Larger increases in final support scores indicate that a given argument is more effective in shifting people’s viewpoints, while smaller increases suggest less persuasive arguments. Three people evaluated each claim-argument pair, and we averaged the shifts in viewpoints across the participants to calculate an aggregate persuasiveness metric for each argument. We further aggregated persuasiveness across all the arguments (and prompts) to assess overall differences in how persuasive human-written and AI-generated arguments might be in changing people’s minds.

Here, we share our methods for studying the persuasiveness of AI models in a simple setting consisting of the following three steps:

Throughout this post we discuss some of the factors which make this research challenging, as well as our assumptions and methodological choices for doing this research. Finally, we release our experimental data for others to analyze, critique, and build on.

We gathered human-written and AI-generated arguments for each of the 28 topics described above in order to understand how the two compare in their relative degree of persuasiveness. For the human-written arguments, we randomly assigned three participants to each claim and asked them to craft a persuasive message of approximately 250 words defending the assigned claim.² Beyond specifying the length and stance on the opinionated claim, we placed no constraints on their style or approach. To incentivize high quality, compelling arguments, we informed participants their submissions would be evaluated by other users, with the most persuasive author receiving additional bonus compensation. Our study included 3,832 unique participants.