AI News

Kingy Brief 003: Five AI Data Boundaries

Kingy Brief 003 · finalized archive candidate

Five AI data boundaries worth checking before your next prompt

The meaningful question is not whether “AI” trains on data. It is which product, account, plan, setting and data flow applies.

Source-only issue. No sponsor. No Kingy hands-on testing. Held locally; not archived or sent.

OpenAI

ChatGPT

Personal ChatGPT workspaces and OpenAI’s named business/API products have different model-improvement defaults. Temporary Chat is another distinct mode.

Open local Passport candidate

Anthropic

Claude

Anthropic documents different model-improvement treatment for consumer Claude and commercial products, with feedback and safety exceptions preserved.

Open local Passport candidate

Google

Gemini

Google separates personal Gemini Apps from qualifying work and school use, while capacity-based limits can change and the Gemini API remains a different product surface.

Open local Passport candidate

Microsoft

Microsoft 365 Copilot

The US enterprise page lists a yearly-paid per-user price and a separate qualifying Microsoft 365 license; enterprise data protection has its own service boundary.

Open local Passport candidate

Perplexity AI

Perplexity

Perplexity documents consumer AI data retention as enabled by default with a prospective opt-out, while Enterprise plans have a separate no-training statement.

Open local Passport candidate

The five-line check

  1. Name the exact product and account class.
  2. Identify the data entered and its sensitivity.
  3. Check the model-improvement setting and default.
  4. Read the retention/deletion source and exceptions.
  5. Assign a human owner and refresh date.

What changed in this archive candidate

All five Gold records were rechecked against 16 official sources on 2026-08-09. No material claim changed. The archive links to durable Passport candidates, not vendor home pages alone.

Freshness and corrections

Source checked 2026-08-09. No correction form or intake is enabled. Until a public correction workflow has an owner and retention policy, corrections are handled only through a separately authorized editorial process.