Staff already use AI tools; a policy has to assume that

What the guidance documents actually say

The UK's Generative AI Framework for HMG, published 18 January 2024 and later withdrawn on 10 February 2025 in favour of a successor playbook, set out ten principles for civil servants, including knowing a tool's limitations, keeping meaningful human control at the right stage, and not letting private or sensitive data train a generative model without the data owner's knowledge or consent. The National Cyber Security Centre's blog post of 14 March 2023 takes a narrower, more technical angle: it warns that queries submitted to a public large language model are stored by the provider and may be used in model development, and advises organisations to keep sensitive information out of those queries or to consider a self-hosted or privately contracted model instead. The US National Institute of Standards and Technology's AI Risk Management Framework, released 26 January 2023, offers a voluntary structure for building trustworthiness considerations into how an organisation designs, develops and uses an AI system, rather than a specific workplace rule.

Why staff use these tools before any policy exists

None of these three documents were written because an employer asked first; they exist because assistants built into browsers, phones and productivity software are already available to an individual employee regardless of what an employer has approved. A policy written as though adoption is a future decision is already behind the actual behaviour in most organisations. The realistic starting assumption, drawn from all three documents, is that some staff are already pasting text into a public assistant, and a policy has to change that behaviour rather than merely permit or forbid a hypothetical future one.

What an acceptable-use policy needs to cover

Combining the three sources, a workable policy needs to state which categories of data must never go into a public tool's prompt box, who has authority to approve a private or self-hosted alternative for sensitive work, what human review is required before an AI-assisted output is used or sent externally, and how the policy itself will be revisited as tools change. It also needs a channel for staff to disclose the tools they are already using, since a policy nobody discloses against is unenforceable.

• Can a member of staff name what they are and are not allowed to paste into a public assistant, without checking a document first?
• Who signs off on a private or self-hosted alternative when a public tool's terms are not acceptable for a piece of work?
• Does anyone revisit the policy on a schedule, or only after an incident forces the question?

This is best read as an editorial synthesis of three separate documents rather than a single settled standard: government guidance in this area has already been withdrawn and replaced once, and a workplace policy modelled on it should expect to be revised on a similar timeline.