4 Types of Work in Automated Persuasion
We organise the computational literature on persuasion into three capability classes, ordered by the direction of the information flow relative to the persuasion act:
| Capability | Direction | Goal |
|---|---|---|
| Descriptive | Observation → Explanation | Explain and measure persuasion in existing content |
| Simulative | Content → Predicted Response | Predict human persuasive response before it occurs |
| Generative | Target → Optimised Content | Create maximally persuasive content for a given audience |
4.1 Descriptive Capabilities
Descriptive systems explain the mechanics of persuasion in observed content. The core task is understanding what properties of content correlate with persuasive outcomes.
4.1.1 Content Capabilities
Understanding persuasive content requires parsing three intertwined signals:
- Content and Logic
- The quality of arguments, the presence of logical fallacies, the structure of reasoning chains. Computational argumentation mining extracts and classifies these automatically.
- Advertisements
- Visual and multimodal persuasion; predicting ad memorability, click-through, and brand recall from image and video features.
- Emotion and Affect
- Sentiment analysis, emotion recognition, and the mapping of affective content to persuasive outcomes across different demographics.
4.2 Simulative Capabilities
Simulative systems predict how a given person or population will respond to a given message before the message is deployed. This is hard because persuasive effects are moderated by audience characteristics that are rarely directly observed.
Representative lines of work:
- Can Language Models Recognise Convincing Arguments? — testing LLMs as judges of argument quality
- Measuring and Increasing Persuasiveness of Large Language Models [1]
- Memorability prediction from content and context
4.2.1 Single Person
Micro-targeting and personalisation — predicting individual response. Key challenge: the privacy vs. effectiveness trade-off.
4.2.2 Interpersonal Interaction
Section under active development.
4.2.3 Societal
- Propaganda detection
- Predicting opinion shifts in populations
4.2.4 Opinion Dynamics
Agent-based models of belief propagation; echo chambers and filter bubbles.
4.2.5 Content Recommendation
Given an audience, select the content that maximises a persuasive objective.
4.2.6 Audience Selection
Given a piece of content, identify the audience for which it will be most persuasive.
4.3 Generative Capabilities
Generative systems produce persuasive content optimised for a particular configuration of audience, time, channel, sender, and topic. This is the most powerful — and ethically fraught — capability class.
Large language models have dramatically lowered the cost of generating fluent, targeted persuasive text at scale. This raises questions that the research community is only beginning to confront: How persuasive are LLM-generated messages compared to human-written ones? Under what conditions do they outperform human writers? What disclosure obligations do deployers have?
If an AI system can generate a message that is more persuasive than any human writer could produce, targeted at a single individual, delivered via the channel and at the time of greatest receptivity — what governance structures should constrain its deployment?