Fact · Voice privacy and security
In a 2021 study, an attack on the GPT-2 language model recovered hundreds of verbatim text sequences from the model's training data.
Checked against its source on October 17, 2026
In the source's own words
“are able to extract hundreds of verbatim text sequences from the model's training data”
Extracting Training Data from Large Language Models, Nicholas Carlini et al., USENIX Security Symposium 2021, 2021, Abstract.
Open the original- What it measures
- The authors queried GPT-2, a language model trained on scrapes of the public Internet, and tested whether its outputs matched text in the training data.
- What it does not show
- One older model trained on public web text. The study does not measure how often a dictation service's models keep or repeat user speech.