GPT-2 Able to Recite PII in Training Data

February 14, 2019

OpenAI's GPT-2, a powerful language model, was found to memorize and recite verbatim instances of its training data, including sensitive personally identifiable information (PII) such as names, emails, Twitter handles, and phone numbers. This underscores the need for safe and secure AI practices.

For those interested in shaping the future of responsible AI governance and preventing incidents like these, join us at HISPI Project Cerebellum TAIM. Here you can Govern, Map, Measure, or Manage AI incidents to ensure a trustworthy AI ecosystem. JOIN US

Matched TAIM controls

Suggested mapping from embedding similarity (not a formal assessment). Browse all TAIM controls

MEASURE 2.6 — similarity 0.699, rank 1. TAIM detail and related incidents →
MEASURE 2.10 — similarity 0.696, rank 2. TAIM detail and related incidents →
GOVERN 2.2 — similarity 0.696, rank 3. TAIM detail and related incidents →

Alleged deployer: openai
Alleged developer: openai
Alleged harmed parties: openai, people-having-personal-data-in-gpt-2's-training-data

Source

Data from the AI Incident Database (AIID). Cite this incident: https://incidentdatabase.ai/cite/357

Data source

Incident data is from the AI Incident Database (AIID).

When citing the database as a whole, please use:

McGregor, S. (2021) Preventing Repeated Real World AI Failures by Cataloging Incidents: The AI Incident Database. In Proceedings of the Thirty-Third Annual Conference on Innovative Applications of Artificial Intelligence (IAAI-21). Virtual Conference.

Pre-print on arXiv · Database snapshots & citation guide

We use weekly snapshots of the AIID for stable reference. For the official suggested citation of a specific incident, use the “Cite this incident” link on each incident page.