Lecture Notes: Moderating Synthetic Content: the Challenge of Generative AI

Abstract

• Key Concern: Artificially generated content significantly disrupts the public sphere.
• Threats: Spread of misinformation, political propaganda, and non-consensual deepfakes.
• Solution Proposal: Enforce general platform rules rather than creating new policies for synthetic content.

Introduction

• Examples of Synthetic Content:
  1. Fabricated text claims from large language models.
  2. Audio deepfakes affecting politics.
  3. Non-consensual intimate deepfakes.
  4. Artificially generated images in conflict scenarios.
• Technology: Generative AI uses large datasets to produce realistic content.
• Availability: Tools like GPTs and Stable Diffusion are widely accessible.
• Need for Governance: Technological advancements require proper governance and guardrails.

Current Approaches

1. Just Ban It: Synthetic Prohibitionism

• Proposal: Ban all synthetic content due to its potential harm.
• Challenges:
  • Weak detection systems for synthetic content.
  • Not all synthetic content is harmful; some have genuine value.
  • Free speech values are involved in the creation and consumption of synthetic content.

2. Sui Generis Policies for Synthetic Content

Current Policies

• Meta, X, TikTok, YouTube: Each has policies for synthetic content.
• Critiques:
  • No necessity for distinguishing between AI and human-generated content in moderation.
  • Harsh moderation for AI-generated content where humans might receive leniency.
• Example: Misinformation policies apply regardless of content origin.

Integrated Policy

• Principle: Apply the same rules to AI and human-generated content.
• Focus: Harmfulness of content, not its technological origin.
• Application:
  • False content: Existing misinformation policies.
  • Deepfakes: Apply rules ensuring electoral integrity and handling non-consensual content.

Complications

Reinterpreting Transparent Synthetic Content

• Context Change: Known AI-generated content might be perceived differently (e.g., satire).
• Text vs. Audio-Visual Content: Different implications for truth and believability.

Content Moderation Principles

• Rights and Duties: Platforms have responsibilities towards users’ freedom of speech.
• Human vs. Bot Distribution: Considerations differ due to lack of moral rights in bots.

Dealing with Bots

• Bots’ Content: Lacks speaker rights but can have audience value.
• Moderation Justification: Easier to justify restriction on bots’ harmful content.

Conclusion

• Feasibility and Desirability: Blanket bans are infeasible and undesirable.
• Proposed Solution: Technology-neutral content moderation focused on content harm.

References and Additional Information

• Authors and Affiliations: Sarah A. Fisher, Jeffrey W. Howard, Beatriz Kira.
• Funding and Ethics: Supported by UKRI, ethical considerations disclosed.