Further Reading: The Failure Modes of the Future

Tier 1: Verified Sources

Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (2021): 610-623. A foundational critique of large language models, addressing training data bias, environmental costs, and the risk of generating convincing but unreliable text. The paper's analysis of how language models can produce plausible-sounding outputs without understanding anticipates the "confidence laundering" concept.

O'Neil, Cathy. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy. Crown, 2016. Documents how algorithmic decision-making systems — in hiring, policing, lending, and education — encode and amplify human biases. O'Neil's analysis of how algorithms create feedback loops that reinforce inequality is directly relevant to the "training data as fossilized bias" failure mode.

Russell, Stuart. Human Compatible: Artificial Intelligence and the Problem of Control. Viking, 2019. A leading AI researcher's analysis of how AI systems can produce unintended outcomes and how to design AI that aligns with human values. Russell's framework for thinking about AI alignment complements this chapter's failure mode analysis.

Zuboff, Shoshana. The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power. PublicAffairs, 2019. Analyzes how technology companies extract and monetize behavioral data — creating incentive structures that prioritize engagement over truth. Directly relevant to the "consensus enforcement by algorithm" analysis.

Tier 2: Attributed Claims

The concept of "AI hallucination" — AI systems generating confident but factually wrong outputs — is well-documented in the AI research literature and has been observed across all major language models.

The "model monoculture" concern has been raised by multiple AI safety researchers, who note the concentration of foundational model development among a small number of companies.

The "epistemic pollution" / "model collapse" concept — AI systems trained on AI-generated data degrading in quality — has been documented in research papers examining the effects of training language models on synthetic data.

  1. Start with O'Neil (Weapons of Math Destruction) — for the current reality of algorithmic bias
  2. Then Bender et al. (2021) — for the structural critique of large language models
  3. Then Russell (Human Compatible) — for the AI alignment perspective
  4. Then Zuboff (Surveillance Capitalism) — for the incentive structures that shape AI deployment