Introduction
When AI can pass, what proves learning?
An impressive essay can now conceal almost no student thinking. Bowen and Watson examine how teachers can preserve the effort that builds expertise while preparing students for work where AI assistance is expected.
1. AI basics
Fluent answers still need supervision
Generative AI produces plausible new content from learned patterns. That makes it useful for invention and vulnerable to fabricated facts. Its confident voice gives readers little help distinguishing between the two. Training data and human feedback also shape whose assumptions appear in its answers.
The book illustrates bias with Stable Diffusion: a cited analysis found women represented only 7% of generated doctor images, compared with 39% of US doctors. Attempts to correct representation can introduce historical inaccuracies instead.
Search and retrieval-augmented generation, or RAG, can ground responses in sources. NotebookLM, for example, works with uploaded materials. These tools improve the conditions for accuracy, but teachers still need to check outputs and understand which sources a system is using.