Skip to main content
NexoraLaunch home

Build log live

Privacy and Data-Use Statements for AI Tools

AI · 10 min read ·

What users need to know about how an AI tool handles their data: inputs, storage, training, sub-processors and deletion, written in plain words.

Illustration: A deep-violet data-flow diagram on a thin grid: user, product, model provider, storage, each a geometric block, with labelled arrows and a small padlock icon

When people type something into an AI tool, they often do so more freely than they would into a search box. They paste in contracts, customer emails, code, medical questions, half-finished plans. They are trusting the tool, and its makers, with material that may be sensitive. The least a product can do in return is say, plainly, what happens to it.

This guide describes how to write a privacy and data-use statement for an AI tool: the questions it must answer, how to word it and how to keep it true. It is general information, not legal advice; privacy law is complex and varies by place, so take qualified advice for your own situation.

Two documents, two jobs

Most products need two kinds of document.

The privacy policy is the formal, legally oriented description of how personal data is handled, as required by data protection law in many countries. It is comprehensive and usually long.

The plain-language data-use statement is a short page or panel, written for users, that answers the practical questions about their data in the AI feature. It links to the policy for the detail.

The second is the subject of this guide. It should never contradict the first; where they differ, the problem is in one of them.

What the law expects, in outline

In the UK and the EU, data protection law applies when personal data is processed, including by AI systems. The UK Information Commissioner's Office has published guidance on AI and data protection, which emphasises principles such as lawfulness, fairness, transparency, purpose limitation, data minimisation, accuracy, storage limitation and security, and the need for accountability and for explaining decisions where relevant.

You do not need to reproduce legal text on your statement. You do need to make sure that it supports compliance: that users are told, in clear language, what you do with their data, and that what you say matches what you do. Where you are unsure about roles, such as whether you are a controller or a processor of the data your users submit, or whether you need a data processing agreement, take advice.

The questions to answer

Answer each in a sentence or two. A table with a row per question works well.

What goes in?

What do you receive when someone uses the feature: typed text, uploaded files, audio, images, account details, usage data? Say what is sent to the model and what is not.

Where does it go?

Is processing done by you or by a third party? If a third-party model provider handles it, say so. Name the main sub-processors or link to the list, and state the region where processing takes place where you can.

How long is it kept?

State retention for inputs and outputs, in the product and with any provider. "Not stored after the response is returned" is a clear statement if true. "Kept for a stated number of days for abuse monitoring, then deleted" is also clear.

Is it used to train or improve models?

Say yes or no, plainly. If the answer differs by plan or setting, say how. If a third-party provider has its own terms about training on data submitted through their service, say what those terms provide for your customers, and link to them.

Who can see it?

Your staff, in what circumstances and with what controls. The provider's staff, if any. Reviewers, if human review is part of improving the service. Be specific about conditions, such as "only when you report a problem and give us permission".

What are the security measures?

A short statement of protections: encryption in transit and at rest, access controls and logging. Do not use absolute words such as "completely secure".

How can users control it?

Deleting data, exporting it, turning off optional uses and contacting you with questions. Describe how, and how long it takes.

What should users not put in?

Guidance on content that should not be submitted, such as special categories of personal data, credentials or confidential information of third parties, unless the product is designed and contracted for it.

What are the limits?

AI outputs may be inaccurate, and outputs may reflect what was submitted. Say whether outputs are stored and who owns them, per your terms.

Plain wording

Write for a tired reader on a busy day.

  • Short sentences, active voice.
  • Say "we" and "you."
  • One idea per sentence.
  • Avoid vague phrases such as "to improve our services" without saying how.
  • Use examples: "If you paste a customer email, we send its text to our model provider to write the reply. We do not keep it after the reply is generated."
  • Do not hide bad news in subordinate clauses.
  • Define terms such as "sub-processor" on first use.

Test the statement on someone who is not a lawyer or an engineer. If they cannot answer "will my data be used to train a model?" in ten seconds, rewrite.

Show it where it matters

Placement is as important as wording.

  • Beside the feature, in a sentence or an information panel, at the moment of use.
  • At sign-up, a short summary with a link.
  • In the settings, where users manage their data.
  • In documentation and sales material, the same facts.
  • In your listing, a line and a link.

Keep the content consistent everywhere. Update all places together.

Where you want to use user data for purposes beyond providing the service, such as improving models, think about whether you need a choice and how to offer it. Legal requirements vary, and pre-ticked boxes and bundled consent are problematic in many places. Provide an easy way to opt out, and honour it. Make the default the most protective option where practicable.

Third-party interfaces and agents

If your product calls other services on a user's behalf, such as search or other tools, tell users what is sent to them. If the product exposes a server through which assistants access user data, describe what an assistant can read and do, and how access is granted and revoked. OWASP's list of top risks for applications built on large language models highlights problems such as sensitive information disclosure and excessive agency; reviewing it helps you identify what to state and what to build safeguards for.

Keep it true

A statement that was accurate at launch can become false as the product changes.

  • Assign an owner for the statement.
  • Link it to your change process: any change that alters data flows, providers or retention triggers a review.
  • Check it against your contracts, including those with providers.
  • Date it and keep a changelog of changes.
  • Notify users of material changes in advance where possible.
  • Audit annually, or after any incident.

If you discover that the statement was wrong, correct it promptly and say so. If personal data was affected and the law requires notification, follow the rules and seek advice quickly.

A worked example

A team offers a writing assistant that improves text pasted by users. Their old statement read: "We take your privacy seriously and use your data to improve our services." They rewrite it as a one-page statement.

"What we receive: the text you paste and your account email. What happens: your text is sent to our language model provider to generate suggestions, then returned to you. Where: the provider processes text in a stated region. How long: we do not keep your text after the suggestions are returned; we keep a log that records when a request was made and its size, for thirty days, to fix errors and prevent abuse. Training: we do not use your text to train models, and our provider's terms, linked here, say that text sent through their business interface is not used to train their models. Who can see it: no one at our company reads your text unless you send it to us in a support request. Control: delete your account in settings and we delete your data within a stated period; backups age out within a further stated period. Please do not paste passwords or other people's confidential information. Last reviewed: date."

Before publishing, they check each sentence against the provider's contract and their own system's actual behaviour. The statement takes one screen to read and answers the questions buyers ask first.

Questions founders ask

What if I do not know what my provider does with data? Find out before you publish anything. Read the provider's terms and data-handling documentation, ask them in writing where it is unclear and record the answer. If you cannot get a clear answer, assume the worst case when writing and consider changing provider.

Can I say that data is anonymised? Only if you can explain how and are confident that people cannot be identified. Anonymisation is hard; the word is often used loosely. If you cannot stand behind it, say "we remove obvious identifiers" and no more.

What about outputs that contain personal data? Outputs may reflect what a user submitted. Treat them under the same rules as inputs, and say how long they are kept.

What if users are in different countries? Write for the strictest regime you serve, and take advice on transfers of personal data between countries. State where data is processed.

What should I do if a user asks me to delete their data? Have a process: verify the request, delete from live systems, ensure that providers also delete where applicable, record the date and confirm to the user.

A ten-point audit

Once a quarter, check ten things. Does the statement name every provider that sees user data? Does it state retention correctly for each? Does it say whether data trains models, and is that still true? Does it describe human review accurately? Are security claims supportable? Does it describe deletion as it really works? Does it match the privacy policy and contracts? Is the date current? Are staff trained to answer questions consistently? Have there been changes to features that touch data? Fix what fails, record the date and keep the last two versions on file. A statement that survives this audit is worth far more than one that merely looks reassuring.

Summary

An AI tool should tell users, in plain language, what data it receives, where it goes, how long it is kept, whether it trains models, who can see it, how it is protected and how users can control it. Keep it short, specific and consistent with your privacy policy and contracts, put it where users will see it, give real choices and keep it true as the product changes. Take qualified advice on your legal obligations. Trust in an AI product starts with honesty about the data.

Questions and answers

What should an AI privacy statement cover?
What data the tool receives, where it goes, how long it is kept, whether it is used to improve models, who else processes it and how users can control or delete it.
Is a privacy policy enough?
A legal policy is required in many places, but a short plain-language statement near the feature helps users understand what is happening before they use it.
Do I need to say whether data trains models?
Yes, plainly. It is one of the first things users and buyers ask, and the answer should match your contracts and your actual practice.
What about personal data in prompts?
Users may include personal data in what they type or upload. Tell them what happens to it and give guidance on what not to include.

Sources

Related guides

Ask a question