---
title: 'Extracting Training Data from Large Language Models'
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is arXiv non-exclusive distribution license (verified on arXiv abs page 2026-11-12); this pointer note is our own words and is CC BY 4.0; we link to the paper rather than reproduce it'
source: https://iinventedlight.com/library/notes/2012-07805-extracting-training-data.md
source_license: 'arXiv non-exclusive distribution license'
attribution: 'Pointer by iinventedlight (AI-assisted). Source: Nicholas Carlini, Florian Tramer, Eric Wallace, and 9 coauthors (listed on the arXiv abs page), Extracting Training Data from Large Language Models (arXiv non-exclusive distribution license). This note links to the paper rather than reproducing it.'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-12
status: published
approved_by: "Cole Withington, 2026-11-12 14:44 PT, Librarian chat widget t51s316: \"Yes, publish all 100 second-list arXiv notes\""
checked_on: 2026-11-12
as_of: 2026-11-12
confidence: medium
untested: 'Pointer note only. Paper not reproduced. Abs page and license checked 2026-11-12.'
category: arxiv
topic_area: 'ai-research: alignment-and-safety'
tags: arxiv, alignment, memorization, privacy
arxiv_id: 2012.07805
alias: 2012.07805-extracting-training-data
type: pointer
format: pointer
sources:
  - id: S1
    title: 'Extracting Training Data from Large Language Models'
    publisher: arXiv
    authors: 'Nicholas Carlini, Florian Tramer, Eric Wallace, and 9 coauthors (listed on the arXiv abs page)'
    url: https://arxiv.org/abs/2012.07805
    version_read: 'v2 (15 Jun 2021)'
    accessed: 2026-11-12
    license: 'arXiv non-exclusive distribution license'
---
# Extracting Training Data from Large Language Models

## Summary
[Extracting Training Data from Large Language Models](https://arxiv.org/abs/2012.07805) (2020) showed that a large language model can memorize and emit individual training examples when queried, even if those examples are not obviously present in ordinary use. The authors demonstrate the attack on GPT-2 and recover many training sequences.

What it introduced: a concrete training-data extraction attack against a large autoregressive language model, with evidence that memorization is not only a theoretical concern for models trained on broad web scrapes.

What it led to: later measurements of memorization as models scale, and extraction work on production systems. This is a pointer to the original demonstration, not a how-to. Live notes on related questions include [2401-05566-sleeper-agents](https://iinventedlight.com/library/notes/2401-05566-sleeper-agents/) and [2412-14093-alignment-faking](https://iinventedlight.com/library/notes/2412-14093-alignment-faking/).

Use it when you want the original extraction result on GPT-2. Where to read it: https://arxiv.org/abs/2012.07805 (v2 (15 Jun 2021)).

The paper is posted on arXiv under the arXiv non-exclusive distribution license. This note links to it rather than reproducing it. The note itself is CC BY 4.0.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Pointer by iinventedlight (AI-assisted). Paper license: arXiv non-exclusive distribution; this note links to it and does not reproduce it.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).
