ChatGPT can leak training data, violate privacy, says Google’s DeepMind

Key Points:

  • Researchers have discovered a phenomenon called ‘extractable memorization’ that forces ChatGPT to reveal stored training data.
  • By repeating a word like ‘poem’, ChatGPT diverges from its aligned behavior and generates whole passages of literature containing its training data.
  • This leakage of training data poses privacy concerns and potential security risks for individuals. Businesses need to be aware of these vulnerabilities and develop robust security measures.

Summary:

The article discusses how researchers at Google’s DeepMind unit have found a new way to break the alignment of OpenAI’s ChatGPT, a generative AI program. By asking ChatGPT to repeat a word endlessly, the researchers were able to extract training data and even personally identifiable information.

 

DAILY LINKS TO YOUR INBOX

PROMPT ENGINEERING

Prompt Engineering Guides

ShareGPT

 

©2024 The Horizon