What never goes in the chat box | Master AI Automation in 4 hours Master AI Automation in 4 hours Course About Ayush Modules Sample chapter Toolbox The Microcap Minute Module 11: Safety, Privacy & Nonsense Detection / Chapter 1 What never goes in the chat box Watch first, then read.
Same lesson, your pace.
What you will learn - The four data classes and their rules - Why "I'll delete it later" isn't protection - Practical substitutes when you genuinely need AI on sensitive work The floor Everything pasted into a cloud AI travels to servers you don't control, possibly logged for abuse-review, possibly retained per settings, potentially exposed in any future breach.
Assume pasted = permanent .
From that assumption, the floor writes itself: Class Examples Rule Credentials passwords, API keys (except via .env patterns), OTPs, card numbers NEVER, any tool, any tier Identity documents Aadhaar/PAN/passport photos, full birth details, medical reports with identifiers Never paste; redact or keep local Other people's private info friends' messages, classmates' marks, family disputes Never, their privacy isn't yours to spend Confidential obligations school unreleased papers, employer NDAs, anything watermarked "confidential" Never without explicit permission Note what's not banned: your own notes, public documents, anonymised problems ("a friend's marks: 88, 92..."), code with secrets swapped for placeholders.
The floor blocks identification and credentials, not usefulness.
Delete-later is a myth Chats may persist in provider logs beyond your delete click; abuse-review queues can hold flagged content; breaches expose whatever existed.
Prevention beats deletion the way locks beat apologies.
Also: free-tier chats sometimes train/improve services depending on settings, check each platform's data controls once (ten minutes, Module 1 cockpit day), turn off training where offered.
Needing AI on sensitive work anyway Legitimate needs have legitimate routes: Redact first , replace names/IDs with A, B, C before pasting; map restored after Local models , Ollama-class tools run capable models fully offline; nothing leaves the machine Enterprise/school tiers with no-training contracts exist where institutions must comply Do it yourself , some five-minute tasks cost less than the exposure Try it yourself Audit yourself honestly: search your last month of chats for anything from the table above (most people find one).
Note what should have been redacted and how.
Then write your personal one-line policy ("I will never paste ___ into any AI") into learn/my-privacy-floor.md , and s