Merging AI Incidents Research with Political Misinformation Research: Introducing the Political Deepfakes Incidents Database

Christina P. Walker, Daniel S. Schiff, Kaylyn Jackson SchiffAAAI 2024 · DOI

TL;DR

This paper introduces the Political Deepfakes Incidents Database (PDID), a novel collection of politically-salient deepfakes and "cheapfakes." The PDID aims to bridge the gap between AI incidents research and political misinformation studies, providing a crucial resource for understanding the prevalence, trends, and impact of synthetic media in politics.

What Problem Does This Paper Address?

The rapid advancement of generative AI has made it increasingly easy to create hyper-realistic synthetic media, including deepfakes and simpler "cheapfakes." While these technologies have diverse applications, their use in politics is particularly concerning, with instances of deepfakes being leveraged to influence elections, spread scandals, and even impact geopolitical events. Existing research on deepfakes often relies on idealized laboratory studies using researcher-created content, which doesn't fully capture the complexities of real-world political misinformation. Furthermore, while AI incident databases exist, they often lack a specific focus on politically charged synthetic media and the nuanced harms they can inflict. This paper addresses the need for a comprehensive, publicly available resource that connects AI incidents research with political misinformation research, enabling a deeper understanding of the phenomenon and its societal implications.

Methods

The core contribution of this paper is the creation and introduction of the Political Deepfakes Incidents Database (PDID). The methodology involves:

  • Data Collection: Gathering politically-salient deepfakes (videos, images) and "cheapfakes" primarily from English-language social media and news websites, with data collection beginning in June 2023 and ongoing. Incidents date back to 2017.
  • Manual Coding: A detailed codebook was developed and used for manual coding of each deepfake incident. This codebook was informed by existing taxonomies from the AI Incidents Database, Center for Security and Emerging Technology, and AI Vulnerabilities Database, with modifications to capture politically relevant aspects.
  • Codebook Development: The codebook was iteratively developed based on an initial sample of prominent political deepfakes. Variables were chosen for their ability to parse features with sufficient variation and granularity, their reliability and replicability, and their theoretical importance for misinformation researchers.
  • Database Structure: The PDID incorporates variables from existing AI incident databases but is tailored to capture theoretically relevant aspects of political deepfakes, including less tangible political and social harms.

Datasets

The paper introduces the Political Deepfakes Incidents Database (PDID) as its primary contribution. This is a newly created dataset.

  • PDID: A collection of politically-salient deepfakes and "cheapfakes," encompassing synthetically-created videos, images, and less-sophisticated content. It includes metadata, verification data, entity descriptors, social media and sharing information, context and content details, real-world connections, and politics and policy-related variables. The database is intended to be open-access.

Key Insights and Ideas

  • Bridging Disciplinary Gaps: The PDID explicitly merges AI incidents research with political misinformation research, recognizing that a holistic understanding requires insights from both fields.
  • Beyond Detection: While deepfake detection is important, the PDID focuses on the broader context, prevalence, and impact of political deepfakes, moving beyond purely technical challenges.
  • Nuanced Harm Typology: The database goes beyond physical harms to categorize a wider range of potential harms, including human rights violations, discrimination, and environmental concerns, reflecting the multifaceted dangers of political deepfakes.
  • Understanding Communication Goals: The PDID acknowledges that deepfakes can have diverse communication goals, including satire and education, not just malicious intent, allowing for a more comprehensive analysis.
  • Connecting to Real-World Events: A key feature is the linkage of deepfakes to contemporaneous events (scandals, elections, wars) and reported real-world harms, enabling the study of causal relationships.
  • Policy and Narrative Analysis: The database includes variables for policy sectors, framing, and narrative elements, allowing researchers to investigate how deepfakes are used to shape public opinion and policy discourse.

Key Results

The paper does not present quantitative results from the analysis of the PDID itself, as it is an introduction to the database. Instead, it highlights the potential for future research and analysis enabled by the PDID. The key "results" are the establishment of the database and its comprehensive structure, which allows for:

  • Empirical Studies: Filling the gap in current research by enabling studies that link the use of political deepfakes to real-world outcomes.
  • Trend Analysis: Revealing the prevalence and trends of political deepfake usage.
  • Stakeholder Benefits: Providing valuable insights for policymakers, researchers, journalists, fact-checkers, and the public.

Limitations and Future Work

Limitations:

  • Initial Scope: The initial phase of the PDID focuses on video and image content (omitting audio-only) and is primarily limited to the U.S., with intentions to scale gradually.
  • Subjectivity in Coding: Determining communication goals and harm types involves subjectivity and inference.
  • Ethical Considerations: Curating and storing deepfakes carries the potential to increase access to harmful content. The authors are carefully weighing the benefits and risks of public release.

Future Work:

  • Sustaining the Database: Ensuring the PDID remains findable, accessible, interoperable, and reusable (FAIR principles) requires sustainable infrastructure, funding, and staffing.
  • Public Incident Reporting: Developing mechanisms for public reporting to scale data collection and ensure shared governance.
  • Ecosystem Interoperability: Making the PDID interoperable with other AI incident and vulnerability reporting tools.
  • Expanding Coverage: Increasing regional and linguistic coverage, and potentially incorporating other media formats like audio.
  • Shared Governance: Engaging with a broad range of stakeholders to ensure the PDID's utility and effectiveness.
  • Expanding Dataset Coverage: Iterating on coding categories and exploring automated techniques to improve coverage and efficiency.

Summary updated

BibTeX
@inproceedings{walker2024merging,
  author = {Christina P. Walker and Daniel S. Schiff and Kaylyn Jackson Schiff},
  title = {{Merging AI Incidents Research with Political Misinformation Research: Introducing the Political Deepfakes Incidents Database}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i21.30349},
}

Loading related content...

Figure