GPT-2 dupes Medicaid
Occurred: December 2019
WIRED reports that a Harvard medical student has been able to fool volunteers into thinking automatically generated comments submitted about a Medicaid proposal were real.
Of the 1,000 or so comments on proposed changes to Idaho's Medicaid programme on the medicaid.gov website, half had been generated by student Max Weiss using OpenAI's GPT-2 large language model.
Troublingly, Weiss' research shows Medicaid users were unable to distinguish real comments from fake ones.
The stunt demonstrates how easily large language systems such as OpenAI's GPT-2 can be deployed to produce effective misinformation and disinformation campaigns.
Sector: Multiple; Govt - health
Purpose: Generate text
Technology: Large language model (LLM); NLP/text analysis; Neural networks; Deep learning; Machine learning
Issue: Mis/disinformation; Dual/multi-use
Transparency: Governance; Black box
OpenAI (2019). GPT-2: 1.5B release
OpenAI (2019). Better Language Models
GPT-2 research study (pdf)
Weiss M. (2019). Deepfake Bot Submissions to Federal Public Comment Websites Cannot Be Distinguished from Human Submissions
News, commentary, analysis
Published: January 2021