Using AI safely
AI can be genuinely useful — but treat what you upload as data you are handing to another service, and verify anything important that comes back.
Last reviewed: October 2026
AI tools are now useful for research, transcription, translation, summarising, drafting and many routine production tasks. The security problem is not that AI should be avoided. It is that an apparently private conversation can make it very easy to hand information to an external service without thinking about where it goes next.
Don’t paste first and think later
Before uploading text, documents, photographs, audio or video, ask whether you would be comfortable sending the same material to an outside supplier.
Take particular care with:
- confidential or embargoed material;
- unpublished investigations and rushes;
- information that identifies a sensitive source or contributor;
- passwords, API keys and other credentials;
- contracts, legal advice, personnel information and personal data;
- documents containing information about people who have not agreed to its use.
Removing a person’s name may not make material anonymous if the surrounding details still identify them.
Know which account you are using
Consumer, business and enterprise versions of the same AI service can have different arrangements for retention, training and administrative access. Do not assume that a protection advertised for one version applies to another.
For work involving sensitive information, use organisation-approved accounts and settings. Check the provider’s current privacy and data-use terms rather than relying on what was true when the service first launched.
Treat the output as unverified
AI systems can produce convincing information that is incomplete, misleading or simply wrong. They can also invent references and quotations.
Verify facts against reliable sources before publication or decision-making. For journalism, an AI-generated answer is a lead or working aid — not a source.
Be careful with connected AI and agents
The risk changes when an AI tool can read email, browse cloud files, access calendars, run code or take actions on your behalf. It may have access to much more than the information visible in the immediate conversation.
Give connected tools only the access they need. Review permissions periodically and remove integrations you no longer use. Be especially cautious about allowing an AI system to act automatically on material received from outside the organisation: malicious instructions hidden in webpages, documents or messages can sometimes manipulate AI systems, a problem known as prompt injection.
Don’t assume a chatbot is a confidential channel
A chatbot is not a replacement for an approved secure channel with a source, lawyer or colleague. Conversations may be stored, processed by other systems or accessible for safety, support or administration depending on the service and account.
If confidentiality matters, establish that the tool and account are approved for the material before using them.
Protect copyright and editorial judgement
Uploading material to an AI service may have contractual, copyright or contributor-consent implications as well as security consequences. Follow your organisation’s rules for unpublished and third-party material.
AI can accelerate work, but responsibility for what is published or delivered remains with the person and organisation using it.
When in doubt, minimise
You can often get the benefit without supplying the sensitive material itself: remove names and identifying details, use an extract, describe the problem generically or use dummy data.
The safest useful prompt is often the one that contains only what the system genuinely needs.