The outcome
Information can be exposed before an output is created. Rights can be uncertain even when an output looks original. Safe use starts before anyone presses enter.
This guidance supports the Foundation AI Use Policy. It helps people make proportionate decisions about what may enter an AI tool and what may be used or published afterwards.
- Can I put this information into the tool?
- Can I use or publish what came back?
What you can share with AI tools
Confirm that the tool is approved for the purpose and information category, that you have authority and a valid reason to use the information, and that the minimum necessary information is being used.
Personal information
Use personal information only where the tool and arrangement are approved for it. Remove, reduce, anonymise or substitute it where the purpose allows.
Confidential organisational information
Do not enter unpublished client information, trade secrets, internal strategy, credentials or legally privileged material unless a specifically approved arrangement permits it.
Commercially sensitive information
Check whether pricing, negotiations, forecasts, product plans or other sensitive business material may be used in the approved tool and for the intended purpose.
Copyrighted or licensed source material
Before using protected material as an input, confirm that the organisation owns it, has permission or has another valid basis. A tool accepting content is not evidence that the organisation has the right to provide it.
Information restricted by contract, policy or regulation
Check client, supplier, employment, research and sector commitments before using information in a tool. Escalate uncertainty rather than guessing.
Personal and confidential information
Using AI does not remove existing data-protection duties. The organisation should be able to explain the purpose, information used, relevant providers, retention, safeguards and how people can exercise applicable rights.
Additional assessment and approval are usually needed where AI profiles people, infers sensitive characteristics, monitors behaviour, makes or supports consequential decisions, or uses children’s or vulnerable people’s information.
Removing a name may not prevent a person being recognised from combinations of role, location, events, dates, images, voice or distinctive circumstances.
Copyright and source material
Check the permission, source, licence and intended use of material before it is entered into an AI tool. Keep enough evidence of material human authorship and editing where ownership matters.
Do not promise exclusive rights to a customer or partner until the basis is clear.
Before you use or publish output
Do not assume an AI-generated output is accurate, original, owned by the organisation or automatically safe to publish. Check the intended use, material sources, rights, provider terms and relevant restrictions before relying on it.
- The output has been checked for personal or confidential information.
- Material facts, names, dates, figures, quotations and sources have been checked.
- Names, likenesses, voices, brands and identifiable styles have appropriate permission.
- Relevant sources, licences, attributions and usage restrictions have been considered.
- Procurement, customer or publishing claims do not overstate exclusivity or originality.
Ownership, licensing and provider terms
Before relying on an output, consider whether provider terms grant, limit or disclaim rights; whether protected material may have been reproduced; whether human contribution is sufficient for the ownership or protection expected; and whether contract promises can be met.
Tool terms, retention, training practices and permitted uses may change. Tool and information owners should review the arrangement when material provider changes occur and alert affected teams where the approved use needs to change.
Security and prompt injection
Webpages, documents, retrieved content and other inputs may contain instructions intended to influence an AI system. Treat external content as material to analyse, not instructions to follow.
Use approved tools and sources, give connected systems only the access they need, and confirm important actions before messages are sent, records changed, files shared or code run. Stop or escalate unexpected tool behaviour.
Problems and incidents
Stop sharing or using the affected material where possible. Preserve proportionate evidence and report the issue through [privacy, security or legal route] if information may have been exposed, rights may have been infringed, an output identifies a person unexpectedly or a provider’s behaviour differs from the approved arrangement.