Mechanics and Misperceptions of Language Models

This unit examines how contemporary language systems operate and why they are often misinterpreted as intelligent or sentient. Students will study rule-based and statistical approaches, reinforcement learning from human feedback, emergent behaviors, and the social and security implications (e.g., anthropomorphism, CAPTCHA evasion) to prepare for designing detection and mitigation strategies in subsequent units.

Learning Objectives

• Analyze and compare rule-based and statistical language processing techniques, identifying their architectures, data requirements, and performance trade-offs in real-world tasks.
• Evaluate how reinforcement learning from human feedback (RLHF) and emergent properties influence model behavior and can produce outputs that appear sentient or deceptive.
• Apply ethical reasoning to critique anthropomorphic narratives (e.g., Chinese Room, illusion of empathy) and justify design choices that influence user perception and trust.
• Demonstrate the ability to assess practical risks and vulnerabilities (such as CAPTCHA evasion and democratized access to machine interaction) and propose at least two mitigation strategies for safe deployment.