Glossary
Data privacy (in AI)
Data privacy in AI is the question of what happens to the text, files and recordings you put into a tool: who can see them, how long they are kept, and whether anything is trained on them.
In plain terms
Everything typed into one of these tools goes somewhere. The habit worth forming is asking three questions before anything confidential goes in: how long does it stay, who at the provider can read it, and does anything learn from it. All three have documented answers. They frequently differ between the free version and the paid business version of the same product, and checking which one you are on takes a minute.
Why it matters
This is the easiest way an AI pilot creates a genuine problem for an organisation, and it is almost entirely preventable by reading rather than by engineering. Somebody pastes a client contract into a consumer account because it was the tool on their laptop, and the questions that follow are contractual rather than technical. The reason to treat it as a working habit instead of a policy document is that the person making the decision is whoever has the tab open.
How it works
Three separate questions get run together and should not be. Retention asks how long your material is kept. Access asks who at the provider can look at it and under what circumstances. Training asks whether it is used to improve a model. A product can give you an entirely satisfactory answer on one of those and an uncomfortable one on another, so a reassurance about any single question is not an answer about your privacy.
The consumer plan and the business agreement for the same product frequently differ on all three, and that difference is usually the whole answer to whether a tool is usable at work. Checking which arrangement you are actually on takes a minute and is more informative than any amount of reading about the company in general.
Where the servers physically sit matters for regulated material and for contracts specifying a region. This has a definite answer, it is often configurable on business plans, and you want it established before rather than after somebody asks you to demonstrate compliance.
Derived material deserves the same treatment as the original. Transcripts, stored positions used for search, memory entries and ordinary logs are all produced from your content, and stored positions in particular are not readable but are also not anonymous. Anything derived from a confidential document should be protected as though it were that document.
The practical controls are unglamorous and effective. Keep a separate account for confidential work, and remove names and figures before sending where the task does not need them. Switch off history and memory features that are not earning their place, and where the material warrants it, prefer an arrangement in which it never leaves your boundary at all.
Three questions, asked once, before anything confidential goes in
Seen in the wild
Find the data section of a tool's own documentation and read the retention, access and training answers before the first confidential file goes in.
ClaudeRun a model entirely on your own machine, where the question of where your text travels has the shortest possible answer.
OllamaKeep a document collection inside a self-hosted assistant, so both the files and everything derived from them stay within your own network.
AnythingLLM
Common misconceptions
People assume
They say they do not train on our data, so we are fine.
In fact
That answers one of the three questions. How long material is retained, and who at the provider may look at it and when, are separate matters with separate answers. A product can decline to train on anything and still keep your material for a period, and both facts are usually documented.
People assume
Deleting the conversation removes it.
In fact
Deleting it from the interface and removing it from backups, logs and any review process are different things on different timescales. Providers document what deletion actually covers. It is worth reading once, because the gap between the two is exactly the period somebody will ask you about.
People assume
The business tier means it is safe.
In fact
It means different terms, and those terms are still terms you have to read. Business agreements are generally more favourable on all three questions, which is not the same as being satisfactory for your particular obligation. The tier narrows what you need to check rather than removing the need to check.
Telling them apart
Data privacy vs Opt-out
Data privacy
The whole question: where material goes, how long it stays, who sees it and what is learnt from it.
One control within it, usually a setting declining to have your content used for training.
Using the control is worth doing and answers a third of the question. Treating it as the answer is the mistake, because retention and access are untouched by it.
Data privacy vs Self-hosted
Data privacy
The question you have to answer for every tool your organisation uses, including the ones nobody approved.
One strong answer to it for a particular deployment, which costs money and operational effort to hold.
Running things yourself removes the question for that system and not for the sixteen other tools people already have open. Most organisations need the reading habit more than they need the servers.
Questions
- How do I find out what a tool does with our material?
- Look for the data or privacy section of the provider's own documentation rather than the marketing page, and read the three answers specifically: retention period, who may access it, and whether it is used for training. Note which plan those answers apply to, because that qualifier is doing a lot of work.
- Is the paid version genuinely different?
- Frequently yes, on all three questions, and it is the first thing worth checking rather than the last. Many organisations are running confidential work through a consumer account with quite different terms simply because that is what somebody signed up for, and correcting it is an administrative task rather than a project.
- What should we tell staff?
- Something short enough to be remembered: which tools are approved for confidential material, which are fine for public information, and that anything covered by a client agreement goes only into the approved ones. A rule people can recall beats a policy nobody opens, because the decision happens at the keyboard.
- What about the transcripts and search indexes built from our documents?
- Treat them as the documents. Anything derived from confidential material carries some of it, including stored positions used for meaning-based search, which are unreadable and not anonymous. They need the same access controls, the same retention decisions and the same attention in any audit.
Key takeaways
- Three separate questions: how long it is kept, who can see it, and whether anything trains on it.
- The consumer plan and the business agreement for one product often differ on all three.
- Deletion in the interface and removal from logs and backups are different things.
- Anything derived from your material, including search indexes and transcripts, deserves the same protection.
- The decision is made by whoever has the tab open, so the rule has to be short enough to remember.
Tools that use this
- Claude
Documentation with the retention, access and training answers set out.
- Ollama
Locally run, so the travel question has the shortest possible answer.
- AnythingLLM
Files and everything derived from them stay inside your network.
Last checked July 2026