AI acceptable use policy: five decisions, with real examples

01

Which data employees may put into which tools

A tighter rule keeps unpublished sensitive information out of AI tools. In its policy for generative AI in news and editorial work, WEHCO Media writes: "We prohibit staff from putting confidential or sensitive information, including source information, into AI tools unless it has already been published."

The more open version ties the data to the tool. Ringier's guidelines keep confidential information, company secrets and personal data out of AI tools "unless the tool has been internally approved for use with such data categories." The company secrets they name include strategies, financial data and source code. The UK Government's AI Playbook says never to enter official information into public AI applications unless it is already public or cleared for publication. It says exceptions may apply for specific applications with different data handling terms under commercial licences, naming Microsoft Copilot as an example.

Choose which data must stay out of AI tools and which may be used in a tool approved for that category. Make the permission specific enough that an employee can tell whether it covers the document in front of them. If it does not, they need a decision before uploading it. Data rules also have to fit the law and the contracts that apply to your company, so have a lawyer check them.

02

How a tool gets approved

Name who approves a new AI tool and what they check before staff use it. The BBC's guidance says external AI tools "must normally be authorised in line with existing BBC software authorisation or procurement processes before being used". The BBC keeps a list of authorised tools for staff. Agile Collective's staff handbook says it only uses approved AI tools with client projects.

Oxide Computer Company's RFD 576 sets a condition for uploading a document to a hosted model. There must be assurance that the model will not use the document to train future versions. It notes that a tool may train on uploads by default, and that this can generally be changed in its preferences.

You can require staff to use an approved list, or let them proceed when stated conditions have been checked. In either case, say who checks the terms and settings and how approval is recorded. A privacy condition for uploading a document does not, on its own, settle what else the tool may do.

03

What AI output may change or decide

Write down which systems an AI tool may act on, which actions need approval each time, and which can proceed within agreed limits. In its editorial workflow, Ars Technica lets AI tools suggest edits but reserves the changes for people: "These tools can recommend changes; only humans can make them."

The UK Playbook calls for human validation of high-risk decisions influenced by AI. For applications that need instant responses, such as chatbots, it recognises that human review may not be possible in real time. In those cases, it calls for human control at other stages of development and deployment, and recommends testing before deployment and regular checks of the live tool.

A small experiment with limited access can give people room to learn without giving it authority over customer records or business-critical processes. For uses without approval of every action, name who checks the system before it starts and who can intervene or stop it.

04

Who answers for the result

The person who uses an AI tool stays responsible for what it produces. At Ars Technica, anyone using AI in the editorial workflow remains responsible for the accuracy and integrity of the work. Its policy adds: "This responsibility cannot be transferred to colleagues, editors, or the tools themselves." Oxide says the output of a model is the responsibility of the human using it.

The BBC and the UK Government add a named owner. Wherever the BBC uses AI for editorial purposes, a senior editorial figure must be responsible and accountable for overseeing its deployment and continuing use. The UK Playbook asks teams to nominate a senior responsible owner who is accountable for the use of AI in a project.

Choose which uses need a separate owner for deployment and continued use, as well as someone responsible for each result. Write down who checks the work and who can decide that the tool should no longer be used. For a presales team, AI training for presales works through what a person checks before an AI draft reaches a customer.

05

How an employee tries a new tool or use

Write the route as a name and a step: who a request goes to, what they check, and where a trial can run with limited access.

At the BBC, a proposal to use AI must first go to a senior editorial figure, who should consult Editorial Policy. Editorial Policy may refer it to an AI Risk Advisory Group, with experts from legal, data protection, commercial and business affairs, information security and editorial policy. Exemptions from its rule against AI-generated news content, such as graphics, may be considered through a piloting process. Ringier requires early coordination with its legal, data protection and information security teams when a project brings in a new AI tool or integrates one permanently.

The City of Boston's interim guidelines encourage staff to try the tools. The guidelines direct staff whose proposed use needs sensitive information to the city's Department of Innovation and Technology, which helps "provision access to enterprise secure resources". The UK Playbook suggests sandboxes where data, analytics and digital professionals can experiment with AI securely.

Employees also need practice working inside the policy. In AI training, participants practise in the tools the company has approved and learn where a person must decide.