Characterizing AI Manipulation Risks in Brazilian YouTube Climate Discourse
Study analyzes Brazilian YouTube climate discourse using psycholinguistic methods, revealing generative AI manipulation risks; dataset includes 226K videos and 2.7M comments.
Key Findings
Methodology
The study employs a psycholinguistic framework combining 10 persuasion strategies and 7 Theory of Mind (ToM) categories. Data was annotated with GPT-4.1, topics clustered using BERTopic, and comment popularity predicted via the Bradley-Terry model.
Key Results
- Emotional and moral persuasion strategies significantly increased user engagement, with emotional content boosting like ratios by 2.1% on average, while logical and statistical appeals were less effective.
- Generative AI can easily produce climate denial comments; experiments showed an extreme denial model generating highly misleading comments with high semantic similarity.
- Moral persuasion in short videos has grown more effective over time, particularly post-2023, with authority-based strategies also gaining traction in short-form content.
Significance
The study highlights the potential risks of generative AI in manipulating climate discourse, especially in Brazil, a key player in global climate policy. It introduces a psycholinguistic framework and a public dataset, laying the groundwork for future research on digital climate communication and ethical risks of generative media.
Technical Contribution
Proposes a novel framework integrating psycholinguistics and generative AI for climate discourse analysis. Combines persuasion strategies with ToM categories and introduces GPT-4.1-based annotation and BERTopic clustering.
Novelty
This is the first systematic study of psycholinguistic persuasion strategies and generative AI manipulation risks in Brazilian YouTube climate discourse, offering a unique perspective and dataset.
Limitations
- The dataset's increasing proportion of short videos may limit comprehensive analysis of long-form content.
- Generative AI experiments are limited to Brazilian Portuguese and lack multi-language generalization.
- The study does not deeply explore the impact of different publishing channels on persuasion effectiveness.
Future Work
Future work could expand to multilingual and cross-cultural scenarios, investigating generative AI manipulation risks across diverse languages and cultural contexts, and exploring more complex persuasion strategies.
AI Executive Summary
Climate change poses a global challenge, and public perception is heavily influenced by social media. Brazil, as a key player in global climate policy, offers a unique lens to study climate discourse on YouTube.
This study analyzed 226,775 Brazilian YouTube videos and 2,756,165 comments from 2019 to 2025 using a psycholinguistic framework. It revealed that emotional and moral persuasion strategies significantly drive user engagement, while logical appeals are less effective. The study also demonstrated how generative AI could produce misleading climate denial comments, raising ethical concerns.
By providing an open dataset and analytical framework, this research lays the foundation for future studies on digital climate communication and the ethical risks of generative media. Despite limitations in language and channel-specific analyses, it offers critical insights into the intersection of AI and climate discourse.
Deep Analysis
Background
Climate change threatens global health, food security, and economic stability. Social media increasingly shapes public discourse, making it critical to understand its role in climate communication. Brazil, with its geopolitical significance and high YouTube penetration, provides a valuable case study.
Core Problem
The core problem is quantifying the impact of psycholinguistic persuasion strategies on climate discourse and assessing generative AI's potential for manipulation. This challenge spans psychology, linguistics, and computational science.
Innovation
Key innovations include: 1) Combining persuasion strategies with ToM categories for climate discourse analysis; 2) Introducing a GPT-4.1-based annotation framework; 3) Using the Bradley-Terry model to quantify comment popularity.
Methodology
- �� Data collection: 226,775 videos and 2,756,165 comments via YouTube Data API v3.
- �� Annotation: GPT-4.1 annotated 10 persuasion strategies and 7 ToM categories.
- �� Topic clustering: BERTopic analyzed thematic distributions.
- �� Modeling: Bradley-Terry model predicted comment popularity.
Experiments
Experiments included: 1) Analyzing user engagement based on persuasion strategies and ToM categories; 2) Generating climate denial comments using generative AI; 3) Comparing persuasion effects across video lengths and publishing channels.
Results
Results showed emotional and moral persuasion significantly increased engagement, with emotional content boosting like ratios by 2.1%. Moral persuasion in short videos grew more effective over time, while logical appeals were less impactful.
Applications
Findings can optimize climate communication strategies and support tools to detect and counteract generative AI manipulation.
Limitations & Outlook
Limitations include the dataset's growing proportion of short videos, language constraints in generative AI experiments, and limited exploration of channel-specific impacts.
Plain Language Accessible to non-experts
Think of this study as a 'psychological lab' analyzing how people react to climate videos on YouTube. Researchers found that emotional videos get more likes, while logical ones don’t perform as well. They also tested AI's ability to create fake comments like 'Global warming is a hoax' and found it could influence opinions. This raises ethical questions about how AI might manipulate public views online.
ELI14 Explained like you're 14
Imagine watching YouTube videos about climate change. Some make you feel emotional, so you hit like, while others are boring. Scientists studied why this happens and found emotional videos are more popular! They even used AI to create fake comments like 'Global warming is fake.' Scary, right? It shows how AI could trick people online, so we need to be careful about what we believe!
Glossary
Persuasion Strategies
Techniques like emotional or logical appeals used to influence opinions.
Analyzed to understand video engagement.
Theory of Mind (ToM)
The ability to understand others' mental states, like beliefs or intentions.
Used to classify user comments.
Generative AI
AI that creates text or images, like GPT models.
Used to generate climate denial comments.
BERTopic
An unsupervised topic clustering algorithm.
Used to analyze video themes.
Bradley-Terry Model
A statistical model for pairwise comparisons.
Used to predict comment popularity.
Open Questions Unanswered questions from this research
- 1 How can the psycholinguistic framework be generalized to multilingual contexts?
- 2 Are generative AI manipulation risks consistent across cultures?
Applications
Immediate Applications
Optimizing Climate Communication
Design more effective public education content using findings.
AI Manipulation Detection
Develop tools to identify misleading AI-generated content.
Long-term Vision
Global Climate Communication
Expand to multilingual contexts to enhance global climate education.
Abstract
Climate change poses a global threat to public health, food security, and economic stability. Addressing it requires evidence-based policies and a nuanced understanding of how the threat is perceived by the public, particularly within visual social media, where narratives quickly evolve through voices of individuals, politicians, NGOs, and institutions. This study investigates climate-related discourse on YouTube within the Brazilian context, a geopolitically significant nation in global environmental negotiations. Through three case studies, we examine (1) which psychological content traits most effectively drive audience engagement, (2) the extent to which these traits influence content popularity, and (3) whether such insights can inform the design of persuasive synthetic campaigns--such as climate denialism--using recent generative language models. Another contribution of this work is the release of a large publicly available dataset of 226K Brazilian YouTube videos and 2.7M user comments on climate change. The dataset includes fine-grained annotations of persuasive strategies, theory-of-mind categorizations in user responses, and typologies of content creators. This resource can help support future research on digital climate communication and the ethical risk of algorithmically amplified narratives and generative media.