I deleted 5 things from a sensitive file before AI read it. It still found what would break my launch.

Source: Nate’s Newsletter Date: 2026-07-24 URL: https://natesnewsletter.substack.com/p/use-ai-sensitive-files

Summary

A companion piece to the same Nate’s Newsletter post on AI and sensitive files (see the sibling stub on this URL): the author manually redacts a document — name, address, price — before running it through AI, only to have the model still infer the sensitive detail that would have compromised a launch. The anecdote is used to argue that manual redaction is unreliable, since a “sanitized” copy can’t anticipate every inference an AI model can draw from context.

Implications

• Distrust-descends-the-stack / untrusted-config thread: sharpens the trust-boundary point — even deliberate, careful manual sanitization fails against inference, meaning the real fix has to be structural (tooling/routing) rather than relying on individual diligence.
• Work-AI-adoption thread: undercuts the common assumption that “just remove the obvious PII first” is a sufficient control for using consumer AI tools on sensitive work material.