Chapter 4 Further Reading: Trust Calibration
These resources deepen your understanding of AI reliability, trust calibration, verification practices, and the cognitive factors that affect how people evaluate AI-generated content.
On AI Reliability and Accuracy
1. "TruthfulQA: Measuring How Models Mimic Human Falsehoods" Stephanie Lin, Jacob Hilton, Owain Evans — arXiv, 2021
A foundational paper on evaluating truthfulness in language models. The research reveals that larger, more capable models are not necessarily more truthful and introduces a benchmark specifically designed to measure truthfulness rather than linguistic capability. Required reading for anyone who wants a technical grounding in why AI "knows wrong things."
2. "Hallucination is Inevitable: An Innate Limitation of Large Language Models" Ziwei Xu et al. — arXiv, 2024
A technical but accessible argument that hallucination is not a bug to be patched but a structural consequence of how language models work. Understanding why hallucination is fundamental rather than accidental informs how you should calibrate trust systemically.
3. "Do Large Language Models Know What They Don't Know?" Zhen Bi et al. — arXiv, 2023
Examines whether LLMs express appropriate uncertainty about claims they are likely to be wrong about. Spoiler: not reliably. Directly relevant to the confidence-calibration problem discussed in Chapter 4.
On Citation Hallucination Specifically
4. "Check Your Sources: Evaluating the Quality of AI-Generated References" Multiple authors, various publications 2023-2024
Several research teams have independently documented that AI tools hallucinate citations at alarming rates. Search for current research on "LLM citation hallucination rates" to find the most recent studies. The consistent finding: citation hallucination is one of the most dangerous and most common AI failure modes.
5. The Retraction Watch Database retractionwatch.com
Not about AI per se, but an important context-setter: even in peer-reviewed academic literature, a significant number of published papers are retracted for errors or misconduct. If even vetted human scholarship requires verification culture, AI-generated citations require it even more. Understanding Retraction Watch builds the right attitude toward all citations.
On Code Security and AI
6. "Do Users Write More Insecure Code with AI Assistants?" Neil Perry, Megha Srivastava, Deepak Kumar, Dan Boneh — ACM CCS 2023
A controlled study examining whether AI coding assistants lead developers to write less secure code. Findings suggest that developers using AI assistants were less likely to write secure code in some scenarios, potentially because they trusted the assistant's suggestions too much. Essential reading for developers calibrating trust for security-sensitive code.
7. OWASP AI Security and Privacy Guide owasp.org
The Open Web Application Security Project is the gold standard for practical security guidance. Their AI security materials cover both securing AI systems and the security implications of AI-generated code. Bookmark and use as a verification reference.
On Calibrated Judgment and Decision-Making
8. "Superforecasting: The Art and Science of Prediction" Philip Tetlock and Dan Gardner — Crown Publishers, 2015
The book on calibrated judgment from the world's leading researcher on forecasting accuracy. Tetlock's work on what distinguishes well-calibrated forecasters from overconfident ones directly maps to the trust calibration challenge in AI use. Specifically, the chapters on "thinking in probability" and "updating beliefs" translate directly to AI trust management.
9. "Thinking, Fast and Slow" Daniel Kahneman — Farrar, Straus and Giroux, 2011
Kahneman's synthesis of decades of behavioral economics research. Chapter 12 on overconfidence and Chapter 19 on the outside view are particularly relevant to trust calibration — specifically why people over-trust confident-sounding sources and how to counteract that tendency.
On Verification and Fact-Checking
10. "The Checklist Manifesto" Atul Gawande — Metropolitan Books, 2009
Not about AI, but a compelling argument for systematic verification checklists drawn from medicine and aviation. Gawande demonstrates that even experts with years of experience make catastrophic errors when operating from memory and habit rather than systematic checklists. The Trust Audit framework in Chapter 4 is directly inspired by this approach.
11. International Fact-Checking Network (IFCN) poynter.org/ifcn
The professional body for fact-checkers. Their training materials and methodological guides cover verification techniques applicable to any factual claim — including AI-generated claims. The IFCN's criteria for what constitutes verification are a useful standard.
On Human Oversight in AI Systems
12. "Human-in-the-Loop Machine Learning" Robert Monarch — Manning Publications, 2021
A technical book but accessible in its key concepts: when and how to incorporate human judgment in AI-assisted workflows, and what kinds of errors humans are and are not good at catching. Relevant for building systematic human oversight processes rather than ad hoc review.
13. "Automation Bias: A System Factor in Aviation and a Challenge to Oversight" Linda Skitka, Kathleen Mosier, Mark Burdick — International Journal of Aviation Psychology
Research on "automation bias" — the tendency to over-trust automated systems and under-apply critical judgment. Originally studied in aviation but widely applicable. Understanding the psychology of automation bias helps you design review processes that counteract it.
Ongoing Resources
14. AI Incident Database incidentdatabase.ai
A living database of documented AI system failures and harms. Browsing this database provides concrete, specific calibration inputs — real cases where AI systems failed in specific, documented ways. More useful than abstract discussions of AI limitations because it is grounded in actual events.
15. "AI Snake Oil" newsletter and blog aisnakeoil.com — Arvind Narayanan and Sayash Kapoor
A rigorous, research-grounded critical perspective on AI capability claims. Not anti-AI, but specifically focused on separating genuine capability from hype. Useful for calibrating which AI capability claims to trust and which to scrutinize.