Skip to content

AI privacy worries: cheatsheet

WorryWhat it means in plain language
SurveillanceEvery interaction is observed. Text travels through real systems; the people maintaining those systems can read it.
Storage and leak riskWhat the tool keeps can be exposed later, by a breach or by a human reviewer inside the company.
Vendor lock-inThe rules you signed up under can change after you have already trusted the tool, and getting out is harder than getting in.

Surveillance: the postcard test

Ask yourself: would I write this on a postcard and leave it on a park bench? If no, the sensitivity level is too high for a tool whose data path you have not verified. Short-form: if it is sensitive, do not paste it into an unchecked tool.

Storage and leak risk: the what-if-leaked-tomorrow test

Ask yourself: if this conversation appeared in a news article next year with my name attached, what is the consequence? Nothing serious: paste. Job loss, client complaint, or a parent's grievance: do not paste. Same test you already apply to email; the tool is a new surface, the test is old.

Vendor lock-in: the architecture-vs-promise test

Ask yourself: is this tool's privacy protection built into how it works (it cannot retain certain data by design) or is it a promise (the company says it will not)? Promises change; architectures are harder to change. Prefer architecture over promise when the choice is available.


PitfallReality
"Privacy mode" is a privacy guaranteeRead what each mode actually promises. A "temporary chat" or "incognito" mode typically limits training use and visible history but still stores the conversation for some period. The specific protections are real; they are narrower than the word suggests.
"They cannot identify me" = "they cannot reach me"Reducing identifying information is not the same as eliminating risk. Data about how you use a tool can still be exposed in a breach even when your legal name is not attached. These are two different threats managed two different ways.
The first setting I find is the whole pictureMultiple layers interact: account-level settings, app-level settings, browser-level settings, and tool-specific feature toggles (training opt-out, conversation memory). Never assume the visible toggle is the only one.

  1. Postcard test. Would I write this on a postcard? If no, stop.
  2. Leak test. If this appears in a news article next year, what happens? If the answer is serious, stop.
  3. Policy check. Have I verified this tool's data-handling for this kind of content? If no, check first or choose a different tool.

If all three clear, paste.


Where each worry gets resolved in the track

Section titled “Where each worry gets resolved in the track”
WorryWhat to carry forward
SurveillancePhase 2 (data-flow trace): maps every hop your text takes before the model sees it.
Storage and leak riskKeep breach risk and training-data extraction separate in your own thinking: they have different defenses.
Vendor lock-inRead the policy before you commit, then re-read it every few months.
All three, togetherPrefer architectural protections (the system cannot do the thing) over promises (the vendor says it will not).

  • Surveillance: the ongoing observation of interactions as they travel through systems. Not a conspiracy; how networked software works.
  • Storage retention: how long a vendor keeps conversation data and under what terms.
  • Training opt-out: a setting that tells the vendor not to use your conversations to train future model versions. Available on some tools and plans; not universal.
  • Architectural privacy: privacy protection built into how a system is built (what it cannot retain by design), as opposed to promise-based privacy (what the vendor says it will not do).
  • Threat model: a description of what you are protecting, from whom, and what the consequences of exposure would be. Lesson 1.2 is where you build your first one.