Keep your data secure while working with AI

I did a livestream on keeping data private when working with AI, covering what the privacy options are, how to give Claude only synthetic data, and whether a local model could replace it. My internet dropped while the local model was still downloading, so I never got to try it.

You'll learn:

• How Claude's data retention settings work, including turning off model-improvement use and what zero data retention agreements involve
• The argument that hiding your data from an AI agent defeats its purpose, and why trusting the provider is the alternative
• How to have AI write an R function that builds a synthetic version of a dataset, using a skill that makes it ask clarifying questions first
• A way to review an AI-generated function by asking for a plain-language walkthrough and how tied it is to the original data, plus why it still needs testing on other datasets
• How to set up a separate project for generating synthetic data so the real data never sits in the folder you work with Claude in

Tools mentioned: Claude, Claude Code, Positron, Posit Assistant, AWS Bedrock, Ollama, LM Studio, Qwen3 Coder, synthpop, tidyverse

Key moments:

• 00:42 Introducing the topic of data security
• 01:03 Three-part plan: privacy issues, anonymizing data, local models
• 03:44 Posit article on building trust with your model provider
• 07:24 Claude settings for data use and 30-day retention
• 10:38 Zero data retention options: enterprise, AWS Bedrock, Posit Assistant
• 13:38 Asking Claude about fake food pantry client data in Positron
• 15:32 Prompting Claude to build a synthetic data function
• 17:05 Using a skill that grills you with questions
• 20:54 Asking Claude about running local models
• 25:48 Reviewing the generated function and asking for a walkthrough
• 27:44 Setting up a separate project for synthetic data
• 34:48 Asking Claude about the synthetic-only dataset
• 37:04 Internet outage interrupts the stream
• 41:34 Wrapping up and planning a local models stream