Further Reading: Chapter 37 — AI-Generated Content and Synthetic Media

Foundational Research

Goldstein, J. A., Sastry, G., Musser, M., DiResta, R., Gentzel, M., & Sedova, K. (2023). "Generative language models and automated influence operations: Emerging threats and potential mitigations." arXiv preprint arXiv:2301.04246. Georgetown University Center for Security and Emerging Technology. The most rigorous capability assessment of LLMs for influence operation purposes. Methodologically careful about distinguishing demonstrated capability from documented deployment. Essential reading for understanding what is technically possible.

Chesney, R., & Citron, D. K. (2019). "Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security." California Law Review, 107(6), 1753–1820. Introduced the "liar's dividend" concept. Though focused on synthetic video, the epistemic implications for text-based synthetic content are equally applicable. One of the most influential legal-academic analyses of AI-generated content risks.

Buchanan, B., Lohn, A., Musser, M., & Sedova, K. (2021). "Truth, Lies, and Automation: How Language Models Could Change Disinformation." Center for Security and Emerging Technology, Georgetown University. A comprehensive policy-oriented analysis of LLM capabilities for disinformation, predating the GPT-4 era but providing foundational analysis of how language model capabilities translate to propaganda applications.

LLM Technology and Capabilities

Brown, T., et al. (2020). "Language Models are Few-Shot Learners." Advances in Neural Information Processing Systems, 33, 1877–1901. The GPT-3 paper. Reading the capabilities description alongside the propaganda implications developed in this chapter helps calibrate which capabilities matter for influence operations.

Wei, J., et al. (2022). "Emergent Abilities of Large Language Models." Transactions on Machine Learning Research. Examines the unexpected capabilities that emerge in LLMs at scale — capabilities not present in smaller models that appear suddenly as model size increases. Relevant for understanding why LLM propaganda capabilities were not fully anticipated from smaller predecessors.

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623. A critical analysis of large language models from researchers who raised early concerns about the social implications of LLM scale. Provides important counterweight to purely capability-focused analyses.

Detection and Attribution

Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W., & Feizi, S. (2023). "Can AI-Generated Text be Reliably Detected?" arXiv preprint arXiv:2303.11156. A rigorous mathematical analysis showing that reliable AI text detection faces fundamental theoretical limitations as model quality improves. Important for understanding the arms-race structure described in Section 37.6.

Kirchenbauer, J., Geiping, J., Wen, Y., Katz, J., Miers, I., & Goldstein, T. (2023). "A Watermark for Large Language Models." International Conference on Machine Learning. The primary academic analysis of statistical watermarking for LLMs. The most technically sound detection approach and the one most worth understanding — along with its limitations.

Liang, W., et al. (2023). "GPT Detectors are Biased Against Non-Native English Writers." Patterns, 4(7), 100779. Documents the disproportionate false positive rates of AI detection tools on writing by non-native English speakers. Critical for understanding the equity implications of AI detection deployment in academic and professional contexts.

Gehrmann, S., Strobelt, H., & Rush, A. M. (2019). "GLTR: Statistical Detection and Visualization of Generated Text." Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. An accessible early analysis of statistical approaches to AI text detection, including visualization of perplexity distributions. Useful pedagogically for understanding what detection tools are measuring.

Content Farms and Information Ecosystem

DiResta, R. (2020, September). "The supply of disinformation will soon be infinite." The Atlantic. An accessible and prescient analysis of what LLM-scale content generation would mean for disinformation. Written before GPT-3's public release, it accurately anticipates the central challenges.

Starbird, K., Arif, A., & Wilson, T. (2019). "Disinformation as Collaborative Work: Surfacing the Participatory Nature of Strategic Information Operations." Proceedings of the ACM on Human-Computer Interaction, 3(CSCW), 1–26. Analysis of how disinformation operates through networks of willing and unwilling participants. Relevant for understanding how AI-generated content seeds real human amplification networks.

Local news desert research: Multiple reports from the Hussman School of Journalism and Media at the University of North Carolina, particularly the "The Expanding News Desert" reports (2018, 2020), document the collapse of local journalism infrastructure that AI-generated local news exploits.

Scientific Misinformation and Manufactured Doubt

Oreskes, N., & Conway, E. M. (2010). Merchants of Doubt: How a Handful of Scientists Obscured the Truth on Issues from Tobacco Smoke to Global Warming. Bloomsbury Press. The foundational account of the manufactured doubt strategy as used by the tobacco and fossil fuel industries. Chapter 37's analysis of AI's implications for manufactured doubt is directly grounded in the strategy documented here.

Larson, H. J. (2020). Stuck: How Vaccine Rumors Start and Why They Don't Go Away. Oxford University Press. Analysis of how health misinformation spreads and persists. The application to AI-generated health misinformation — particularly the pre-print server vulnerability — extends directly from this foundational analysis.

Van der Linden, S. (2023). Foolproof: Why Misinformation Infects Our Minds and How to Build Immunity. W. W. Norton. Sander van der Linden's accessible synthesis of inoculation theory research, with direct applications to the AI content era. The inoculation design principles in Section 37.10 draw on this framework.

Regulation and Platform Policy

European Parliament and Council. (2024). Regulation (EU) 2024/1689 on Artificial Intelligence (AI Act). Brussels: European Union. The primary legislative text. Article 50 and the surrounding provisions on high-risk AI systems and general-purpose AI are most relevant to AI-generated content.

Fjeld, J., et al. (2020). "Principled Artificial Intelligence: Mapping Consensus in Ethical and Rights-Based Approaches to Principles for AI." Berkman Klein Center Research Publication. Maps the landscape of AI ethics principles across governments, companies, and civil society. Useful for understanding where regulatory proposals come from and where consensus exists.

Dafoe, A. (2018). "AI Governance: A Research Agenda." Future of Humanity Institute, University of Oxford. A foundational research agenda paper that anticipated many of the governance challenges now being confronted. Useful for understanding the policy response landscape.

Inoculation and Counter-Propaganda

Lewandowsky, S., Ecker, U. K. H., & Cook, J. (2017). "Beyond Misinformation: Understanding and Coping with the Post-Truth Era." Journal of Applied Research in Memory and Cognition, 6(4), 353–369. Analysis of why fact-checking alone is insufficient and what approaches are more effective. The foundation for the inoculation approach applied to AI-era disinformation.

Roozenbeek, J., & van der Linden, S. (2019). "Fake news game confers psychological resistance against online misinformation." Palgrave Communications, 5(1), 65. Research on the "Bad News" inoculation game and its effectiveness in building resistance to online misinformation techniques. Directly applicable to designing AI-era inoculation content.

Pennycook, G., & Rand, D. G. (2019). "Lazy, not biased: Susceptibility to partisan fake news is better explained by lack of reasoning than by motivated reasoning." Cognition, 188, 39–50. Challenges the assumption that political bias is the primary driver of misinformation susceptibility. Has significant implications for targeting AI-era inoculation interventions.

Advanced and Emerging Research

Zellers, R., et al. (2019). "Defending Against Neural Fake News." Advances in Neural Information Processing Systems, 32. Early research on the adversarial detection problem — generating synthetic news and simultaneously building detectors — that anticipated the arms-race dynamic described in Section 37.6.

Kreps, S., & Kriner, D. (2023). "The Artificial Intelligence Disinformation Threat." Perspectives on Politics. Political science analysis of how AI disinformation affects political behavior in experimental settings. Provides empirical grounding for capability assessments.

Lin, H., & Kerr, J. (2017). "On Cyber-Enabled Information Warfare and Information Operations." In Oxford Handbook of Cybersecurity. Oxford University Press. Situates AI-enabled information operations within the broader context of information warfare theory. Useful for connecting Chapter 37's analysis to the strategic studies literature.

Journalism and Practitioner Resources

NewsGuard's "AI-Generated News Sites" monitoring reports (ongoing, available at newsguardtech.com). The most current empirical tracking of AI-generated news sites in the wild.

Stanford Internet Observatory (io.stanford.edu). The leading academic research organization documenting real-world influence operations, including AI-assisted ones. Their case studies provide concrete examples of how the capabilities described in this chapter are actually deployed.

Poynter Institute's AI resources (poynter.org). Practitioner-focused coverage of AI in journalism and disinformation, including guidance for journalists investigating AI-generated content.

First Draft (firstdraftnews.org). Resources for journalists and researchers on verification and misinformation, with increasing coverage of AI-specific challenges.

Chapter 38 (Deepfakes and Synthetic Media: The Visual Dimension) continues the Part 7 examination of AI-generated content, focusing on synthetic video and audio and their applications in political disinformation.