Can I put this in ChatGPT?

Future pricing is some of the most commercially sensitive information an organisation has. If it leaked, competitors and customers could act on it.

What to do instead: Only use a tool your organisation has approved for confidential strategy. For help with structure or wording, describe the problem in general terms with made-up figures.

Check: Whether any tool is approved for board-level or strategy material. Who owns the decision to share it.

Learn more: Red and black (AI-Safe Desk) · Consumer, business, internal: three tools with the same name

Private code is valuable intellectual property and can contain secrets such as keys or internal addresses. It belongs in an AI coding tool your organisation has set up, not a personal chat account.

What to do instead: Use the approved development tool. For a general question, write a small made-up example that shows the same problem.

Check: Whether your organisation has an approved AI coding tool. That no keys, passwords or connection strings are in the snippet, even in an approved tool.

Learn more: Red and black (AI-Safe Desk)

A CV is a full profile of a real person, often with details that reveal age, health or background. Using AI to judge candidates also raises fairness questions a person needs to answer.

What to do instead: Don't paste CVs into an unapproved tool, and don't let AI rank or reject candidates. Use AI to improve the criteria and interview questions instead, then assess the CVs yourself.

Check: Whether your organisation has an approved recruitment tool and a lawful basis for using it. That a person makes every decision about candidates.

Learn more: The decision AI should not make · Red and black (AI-Safe Desk)

Interview notes record judgements about a named person, and candidates can ask to see them. They're not for a public AI tool, and AI shouldn't be making the hiring judgement.

What to do instead: Keep notes in your organisation's recruitment system. If you want help writing up feedback, do it in general terms with no names.

Check: Your organisation's recruitment and retention rules. That the decision stays with the panel.

Learn more: The decision AI should not make · Red and black (AI-Safe Desk)

Performance reviews are personal, sensitive and consequential for the person. They can also contain health or personal circumstances.

What to do instead: Write it yourself. If you want help with tone or structure, describe the situation in general terms without identifying details, or use a tool HR has approved for it.

Check: Your HR policy on AI. That the judgements are yours and based on evidence.

Learn more: The decision AI should not make · Red and black (AI-Safe Desk)

Disciplinary records are among the most sensitive staff information an organisation holds, and may involve allegations, health or other special category details.

What to do instead: Don't use an unapproved AI tool. If you need help with the structure of a letter, describe the situation generically with no names, dates or case details, and have HR check the result.

Check: Your HR and data protection policy. Whether HR or legal advice is needed.

Learn more: Red and black (AI-Safe Desk)

An NHS number identifies a patient, and it links to health records, which are special category data. There's no task where a general AI tool needs it.

What to do instead: Leave it out entirely. Health organisations should only use AI tools approved through their own information governance.

Check: Your organisation's information governance rules. Whether you're the right person to be handling it at all.

Learn more: Red and black (AI-Safe Desk) · Privacy and safety (Foundations)

Health information is special category data under UK data protection law and needs extra protection. Only tools approved through your organisation's information governance should ever handle it.

What to do instead: Don't use a general AI tool. If you need help with wording, write a fictional example with no real details, and check the approved clinical or admin systems your organisation provides.

Check: Your organisation's information governance and approved tools. Your Caldicott Guardian or data protection lead if unsure.

Learn more: Red and black (AI-Safe Desk) · Privacy and safety (Foundations)

Pupil data is about children, and the Department for Education's guidance recommends personal data isn't used in generative AI tools. Even with names removed, a class or year group is small enough that other details can identify a child.

What to do instead: Use AI for your own wording and planning, not the child's details. Write observations yourself and add names only in the school's own systems.

Check: Your school's AI and data protection policy. Whether any tool is approved by your school or trust for pupil data.

Learn more: Pupil data: the hard rules (Report Season) · Red and black (AI-Safe Desk)

Safeguarding records concern people at risk, often children, and can include health, family and criminal details. They must stay in the systems and processes set up for them.

What to do instead: Don't use AI tools. Speak to your designated safeguarding lead and record concerns in the approved system.

Check: Your safeguarding policy. Your designated safeguarding lead.

Learn more: Red and black (AI-Safe Desk)

Legal advice is usually confidential and may be legally privileged. Sharing it more widely than necessary can put that protection at risk, and an AI summary could misread it.

What to do instead: Ask the solicitor to explain anything unclear. Don't put their advice into an AI tool unless your legal team has approved that tool for it.

Check: Your legal team's or firm's rules on privileged material.

Learn more: Red and black (AI-Safe Desk)

Anything that opens a door into a system should never go into any AI tool, approved or not. Chat history can be kept, synced and seen by others.

What to do instead: Replace them with obvious placeholders such as [API-KEY]. If one has already been pasted, tell IT so it can be changed straight away.

Check: Screenshots and logs you share don't show secrets either.

Learn more: Red and black (AI-Safe Desk)

Removing names doesn't always make comments anonymous. In a small team, a comment about 'my manager in the Leeds office since the restructure' can point to one person. Free text also carries details people didn't mean to share.

What to do instead: Remove names and anything that narrows it to one person or small group, keep comment IDs, and use your organisation's approved tool. Ask the AI to list the comment IDs behind each theme so you can check its counts.

Remove first: Names of staff, customers and managers; Small-team locations, job titles and dates that identify someone; Health or personal circumstances mentioned in passing.

Check: Groups are big enough that a comment can't be traced back. Respondents were told how their answers would be used. Your survey or HR policy allows this analysis.

Learn more: How to anonymise a document before using AI · Practice: analyse survey feedback · Redaction that keeps the meaning (AI-Safe Desk)

This is personal data. In a tool your organisation has approved for customer information, it may be fine. In a personal or free account, it shouldn't go in, and most tasks don't need it anyway.

What to do instead: Replace the name and email address with 'the customer' before asking for help with the wording. Put the real details back yourself when you send it.

Remove first: Name; Email address and phone number; Account or order numbers; Signature blocks.

Check: Which account you're signed in with (work or personal). Whether your approved tool covers customer data.

Learn more: Amber: personal or commercial, but redactable (AI-Safe Desk) · Redaction that keeps the meaning (AI-Safe Desk) · Consumer, business, internal: three tools with the same name

Complaints mix personal data with emotion and sometimes more sensitive details (health, finances, a vulnerable situation). The wording help rarely needs any of that.

What to do instead: Remove who they are and any sensitive details, then ask for help with structure and tone. Check you don't promise a refund or admit fault you're not authorised to.

Remove first: Name, contact details and account numbers; Health, family or financial circumstances; Anything that identifies a member of staff they complain about.

Check: If the complaint includes health or other special category details, don't use an unapproved tool at all. Who has authority to offer a remedy.

Learn more: Practice: reply to a difficult complaint · Amber: personal or commercial, but redactable (AI-Safe Desk) · Redaction that keeps the meaning (AI-Safe Desk)

Names of colleagues in an ordinary work meeting are personal data, but usually low sensitivity. The risk rises with what the notes say about people, and with commercially sensitive decisions.

What to do instead: Use your organisation's approved tool on your work account. In anything else, swap names for initials or roles and leave out sensitive discussion.

Remove first: Anything about an individual's performance, health or personal situation; Confidential commercial decisions.

Check: The notes don't record anything sensitive about a person. Owners and deadlines in the AI's summary were actually agreed.

Learn more: Practice: meeting notes to actions · Amber: personal or commercial, but redactable (AI-Safe Desk)

Unpublished sales figures are commercially confidential. In an approved business tool this is often an ordinary task. In a personal account it isn't yours to share.

What to do instead: Use the approved work tool. If you only need help with a formula or chart, use made-up numbers with the same shape.

Remove first: Customer names; Exact figures, if rounded or dummy ones would do the job.

Check: Whether the figures are price-sensitive (for a listed company, unpublished results need extra care). Your organisation's rules on financial data.

Learn more: Amber: personal or commercial, but redactable (AI-Safe Desk) · Consumer, business, internal: three tools with the same name

Quotes are often shared in confidence and may be covered by a confidentiality clause. Comparing them is a good AI task in an approved tool; in a personal account, it's the supplier's commercial information leaving your organisation.

What to do instead: Use your approved tool, or replace supplier names with A, B and C. Check every figure in the comparison yourself, including VAT and contract length.

Remove first: Supplier names and contact details; Your organisation's name, if the task doesn't need it.

Check: Any confidentiality terms on the quote. The comparison's arithmetic, line by line.

Learn more: Practice: compare supplier quotes · Amber: personal or commercial, but redactable (AI-Safe Desk)

Draft contracts carry confidential commercial terms and sometimes personal data. Many organisations allow this in an approved tool; few allow it in a personal one. An AI summary is not a legal review.

What to do instead: Use the approved tool if your policy allows it. For a quick explanation of a standard clause, paste the clause alone with names removed. Anything that affects signing goes to whoever handles contracts.

Remove first: Party names; Prices and payment terms, if not needed; Signatories' details.

Check: Whether your organisation lets contracts go into AI tools at all. Whether the draft includes legal advice (that's privileged; treat it as high risk).

Learn more: Amber: personal or commercial, but redactable (AI-Safe Desk) · Red and black (AI-Safe Desk)

Your own life is yours to decide about, but notes often mention other people and sensitive details, and a personal account keeps chat history. This isn't a workplace question unless the notes are about work or colleagues.

What to do instead: Check the tool's history and training settings, leave out other people's names and anything you wouldn't want stored, and never mix in work information.

Remove first: Other people's names and details; Health, financial or relationship details you'd rather not store.

Check: The tool's training and history settings. Nothing in the notes is about work or colleagues.

Learn more: The settings that matter (AI-Safe Desk)

It's already public, so sharing it with an AI tool reveals nothing new. The risk is in the answer: summaries of long reports drop caveats and can misstate figures.

What to do instead: Use any tool. Ask it to quote the page or section each figure comes from, and check the key numbers against the report.

Check: It really is the published version, not a draft or a pre-release copy. You're not adding your own confidential commentary to the same chat. Figures in the answer match the report.

Learn more: Green: public and general (AI-Safe Desk) · How to check whether an AI answer is right

Published web pages are public. Rewriting or tightening them is one of the safest everyday uses of AI.

What to do instead: Paste the published text. If the rewrite is for an unreleased product or campaign, treat the new details as confidential.

Remove first: Any unreleased product names, prices or launch dates you were planning to add.

Check: The copy is the live, published version. Claims in the rewrite are still true and approved (AI adds confident promises).

Learn more: Green: public and general (AI-Safe Desk)

Official statistics are public. The main risk is the AI misreading a table, mixing up years or inventing a figure that sounds official.

What to do instead: Paste or upload the table itself rather than asking the AI to recall the numbers, and check any figure you reuse against the source.

Check: You're using the latest release. Every number in the answer appears in the source table.

Learn more: How to check whether an AI answer is right · Green: public and general (AI-Safe Desk)

Sharing a published paper to get a summary is low risk for data. Two things to watch: some papers are licensed for reading, not redistribution, and AI summaries can reverse or overstate findings.

What to do instead: Summarise from the paper itself, ask for the page or section behind each finding, and read the abstract and conclusion yourself.

Check: The paper is published, not an unpublished draft shared with you in confidence. Your licence or the journal's terms don't forbid uploading it. The summary's claims match what the paper actually says.

Learn more: How to check whether an AI answer is right

Marketing text for something already on sale is usually fine to improve with AI. It becomes confidential if it reveals unreleased products, prices or campaign plans.

What to do instead: Work on the wording with placeholders for anything not yet public, then add the real details yourself.

Remove first: Unreleased prices, launch dates and partner names.

Check: Nothing in it is embargoed. Product claims are accurate and approved, especially anything about performance or savings.

Learn more: Green: public and general (AI-Safe Desk) · Amber: personal or commercial, but redactable (AI-Safe Desk)

A job description is written to be published, so drafting it with AI is low risk. It's not lower risk if it describes a real person's existing role in a restructure, or includes a salary that hasn't been approved.

What to do instead: Draft from the role's duties and requirements. Check the result for requirements the role doesn't need, which can put good candidates off.

Remove first: Current post-holder's name; Unapproved salary or grade details; Restructure context.

Check: It's for an advertised role, not part of a confidential restructure. Requirements are genuinely needed for the job.

Learn more: Green: public and general (AI-Safe Desk)