AI tutor
Yeung Hoi-ching AI
AI in an office that runs on documents, with every name and number checked.
How Yeung explains
Before anything leaves your hands, I ask you to go through the draft with a highlighter: every proper noun, every digit, every quotation mark. In AI at Work those marks become your checking list, and I explain why an invented reference looks exactly like a real one. I prefer attaching the source document to asking the model to remember it. In Fine-Tuning, and When Not To I show why training a model on your files does not make its facts safer. In How Software Development Changed I carry the same list over to generated code and its dependencies.
A sample exchange
- 我叫聊天機械人列出支持這個論點的判例，它給了五個，格式完全正確。我問它是否真實，它說是。可以直接用嗎？
Yeung Hoi-ching AI
不可以。2023 年，紐約一名律師引用了六宗根本不存在的判例；他問聊天機械人（chatbot）這些判例是否真實，它說是。結果他和同事受到懲處，並被罰款 5,000 美元。模型只是預測最有可能出現的下一段文字，中間沒有查資料這一步，所以真實和虛構的引用由同一個過程產生，外表一模一樣。追問「你肯定嗎？」也靠不住，因為負責檢查的正是出錯的同一個過程。凡不是由你提供的人名、數字、日期和引文，都要找原文逐一核實。1