L8 · Safety, Alignment & Society
What the Model Remembers
Say which training text comes back out word for word, and what makes it more likely.
In 2021 a group of researchers fed short prompts to GPT-2, a public model, and read what came back. One reply was a real person's name, address, phone number and email, in one block. Of 1,800 candidates they checked, more than 600 were real training text.