Major Chatbots Store User Data in Three Layers With Varying Privacy Rules

Overview

• OpenAI counts at least 900 million active ChatGPT users, yet researchers warn about data memorization in large models and note that more than 225,000 stolen chatbot logins surfaced on Dark Web forums in 2024.
• ESET outlines a first layer of storage on company servers where chat histories can be kept for limited periods, often up to 30 days, to monitor security and abuse.
• A second layer involves using conversations to train models, which OpenAI enables for both free and paid users but allows consumers to disable, while some enterprise and health deployments include training use by default.
• A third layer relies on human review, where trained specialists read short chat fragments to rate quality and help block harmful responses.
• Policies differ across providers: Google’s Gemini follows Google privacy rules and uses human reviews, while Anthropic’s Claude promotes minimal data retention under its “Constitutional AI” approach.