Peer-reviewed research
Extracting Training Data from Large Language Models
Nicholas Carlini et al., USENIX Security Symposium 2021, 2021
Read the originalRead on October 17, 2026
- What it is
- A security study that attacks GPT-2, a large language model trained on text scraped from the public Internet, and recovers verbatim sequences from its training data. The paper was presented at USENIX Security 2021, a peer-reviewed security venue.
- How much weight it carries
- Peer-reviewed and widely cited. It tests one older model on public web text. It shows that large models can memorize and repeat training text, not how often today's dictation services keep or repeat user speech.
Facts taken from this source
- In a 2021 study, an attack on the GPT-2 language model recovered hundreds of verbatim text sequences from the model's training data.
“are able to extract hundreds of verbatim text sequences from the model's training data” (Abstract)