Abstract:As generative foundation models improve, they also tend to become more persuasive, raising concerns that AI automation will enable governments, firms, and other actors to manipulate beliefs with unprecedented scale and effectiveness at virtually no cost. The full economic and social ramifications of this trend have been difficult to foresee, however, given that we currently lack a complete theoretical understanding of why persuasion is costly for human labor to produce in the first place. This paper places human and AI agents on a common conceptual footing by formalizing informational persuasion as a mathematical decision problem and characterizing its computational complexity. A novel proof establishes that persuasive messages are challenging to discover (NP-Hard) but easy to adopt if supplied by others (NP). This asymmetry helps explain why people are susceptible to persuasion, even in contexts where all relevant information is publicly available. The result also illuminates why litigation, strategic communication, and other persuasion-oriented activities have historically been so human capital intensive, and it provides a new theoretical basis for studying how AI will impact various industries.

What problem does this paper attempt to address?

The problem this paper attempts to address is: Why is persuasion in economic and social activities both costly and difficult? Specifically, the author formalizes the problem of information persuasion and studies its computational complexity to explain why persuasion historically required significant human capital investment and to explore how AI automation will impact persuasion activities and their effects on the social knowledge landscape. ### Main Contributions 1. **Formalizing the Persuasion Problem**: The author defines information persuasion as a mathematical decision problem, which involves selecting a set of facts or using a model to frame public information to increase the other party's confidence in a focal event. 2. **Computational Complexity Analysis**: By proving that the information persuasion problem is NP-complete, the author reveals the computational difficulty of persuasion. This means that finding effective persuasion strategies is difficult, but verifying existing persuasion strategies is relatively easy. 3. **Explaining Human Susceptibility to Persuasion**: The paper explains from the perspective of computational complexity why people are easily persuaded even when all relevant information is publicly available. This is because generating effective persuasion strategies requires significant computational resources, while accepting others' persuasion is relatively simple. ### Background and Motivation - **Dual Impact of Technological Advances**: With the advancement of generative foundational models (such as large language models), these models can not only accelerate scientific discovery but also make existing knowledge more widespread. However, this also raises concerns about AI being used for large-scale manipulation and misinformation. - **Economic and Social Impact**: Persuasion activities hold a significant position in modern society, with many industries (such as law, advertising, politics, etc.) relying on persuasion capabilities. Understanding the cost and difficulty of persuasion is crucial for assessing the impact of AI automation on these industries. ### Research Methods - **Formal Definition**: The author defines a probability space and a series of fact sets, along with a target event and a belief threshold, then studies whether there exists a subset of facts that can increase the receiver's belief in the target event beyond the threshold. - **Computational Complexity Proof**: By reducing the information persuasion problem to known NP-complete problems (such as the exact cover problem), the author proves that the information persuasion problem is NP-complete. ### Conclusion - **Computational Cost of Persuasion**: The computational cost of information persuasion grows exponentially with the number of facts considered in the worst-case scenario, explaining why persuasion activities historically required significant human capital. - **Impact of AI**: Although AI can generate efficient persuasion strategies, the cost and difficulty still exist. This provides a theoretical basis for studying how AI will reshape the persuasion industry and its impact on the social knowledge landscape. Through this research, the author provides a new theoretical framework for understanding the cost and difficulty of persuasion and offers directions for future research.

When and Why is Persuasion Hard? A Computational Complexity Result

Artificial Influence: An Analysis Of AI-Driven Persuasion

Undermining Mental Proof: How AI Can Make Cooperation Harder by Making Thinking Easier

A Mechanism-Based Approach to Mitigating Harms from Persuasive Generative AI

Bayesian explanations for persuasion

Towards Understanding Persuasion in Computational Argumentation

The potential of generative AI for personalized persuasion at scale

Algorithmic Bayesian Persuasion with Combinatorial Actions

Large Language Models are as persuasive as humans, but how? About the cognitive effort and moral-emotional language of LLM arguments

Computational Aspects of Private Bayesian Persuasion

Deception and Manipulation in Generative AI

The Persuasive Power of Large Language Models

Non-Bayesian Persuasion

Algorithmic Persuasion with Evidence

Learning to Persuade on the Fly: Robustness Against Ignorance

Measuring and Improving Persuasiveness of Large Language Models

On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial

Is computational complexity a barrier to manipulation?

Counterfactual Reasoning Using Predicted Latent Personality Dimensions for Optimizing Persuasion Outcome

What Makes a Message Persuasive? Identifying Adaptations Towards Persuasiveness in Nine Exploratory Case Studies

Is artificial intelligence more persuasive than humans? A meta-analysis