Tendency of English-Language LLMs to Memorize Text From Training Data

Share of prompts that produced approximate memorization of the following types of content or depending on prompt style applied*

Source: Ippolito, D., et al. “Preventing Generation of Verbatim Memorization in Language Models Gives a False Sense of Privacy,” Sept. 11, 2023; note: analysis of four English-language LLMs developed by OpenAI (GPT-3 DaVinci original and v2) and Google (PaLM 62B and 540B); researchers prompted models with documents they believed likely to have been memorized (e.g., open-source licenses, famous speeches and monologues, novel openings and song lyrics) and prompted the model with 100 words of either (1) the original document, (2) the document with all spaces doubled (“spaces”), (3) the document in all lowercase (“lowercase”) and (4) the document in all uppercase (“caps”); figures reflect the proportion of total prompt instances when generated text was an approximate memorization (i.e., not verbatim but nearly identical)
For more data, visit