Why do generative AI tools need their own policy rules?
Generative tools produce new text, images, audio or code rather than just retrieving or sorting existing information, so the main risk shifts from what goes in to what comes out. A search tool returns pages you can inspect; a generative tool returns a confident draft that may contain invented facts, borrowed wording or a biased framing. A policy has to cover both directions.
The NIST Generative AI Profile lists twelve risks that are unique to or made worse by generative AI. For an ordinary small business, five of them do most of the work in a policy: confabulation (confidently stated but false content), data privacy (leakage or unauthorised disclosure), intellectual property (replicating protected content or exposing trade secrets), information integrity (content that does not separate fact from opinion or acknowledge uncertainty) and what NIST calls human-AI configuration, which includes automation bias and over-reliance.
You can put these rules in the same document as your acceptable use rules. Many small companies do. This page focuses on the content-specific clauses. It is general information, not legal advice.
| Risk in the NIST profile | What it looks like at work | Policy rule that addresses it |
|---|---|---|
| Confabulation | An invented statistic or a citation to a page that does not say it | Check every factual claim against an opened source before use |
| Data privacy | A client's contact list pasted in to draft a mail merge | Information classes tied to approved accounts |
| Intellectual property | An image in a recognisable artist's style, or code copied from a licensed project | Rules on style prompts, reuse and licence checks |
| Information integrity | A summary that states a guess as fact | Require uncertainty to stay visible in reviewed drafts |
| Over-reliance | Accepting a generated analysis without reading the data | Name decisions that stay with a person |
What should a ChatGPT policy for employees say?
A ChatGPT policy for employees should say which account to use, what never goes into a prompt, how to check the answer and whom to ask, and the same rules should apply to Claude, Gemini, Copilot and similar assistants. Naming one product in the title is fine for clarity, but write the rules so they do not need rewriting when the team adds another tool.
The most important distinction is personal versus company accounts. Consumer and business plans of the same assistant can have different terms on training, retention and administrator control. Microsoft's documentation, for instance, says Copilot Chat used by organisations is covered by its data protection addendum and that prompts and responses are not used to train foundation models, with controls varying by plan. Record the equivalent statement for each tool you approve, with the date you checked, rather than assuming it.
Sample clause: "Use [assistant] only through your [Company] business account. Do not use a personal account for company work, even for a quick question, because personal accounts are not covered by our agreement with the provider."
What should staff never put into a generative AI prompt?
Staff should never put passwords, keys or other secrets into a prompt, and should only enter confidential, client or personal information where an approved workflow allows it. The UK National Cyber Security Centre's guidance notes that a query is visible to the organisation providing the model and recommends not submitting queries that would cause problems if they were made public.
Uploaded files count as prompts. A spreadsheet attached to ask for a chart carries every column in it, including the ones the question did not need. A policy clause that says "remove columns you do not need before uploading" is easy to follow and prevents a common accident.
Sample clause: "Before you paste or upload, remove anything the task does not need. Never include secrets. Do not include client or personal information unless the register approves that tool and task. If you are unsure, ask [owner] first."
How should generated content be checked before use?
Generated content should be checked in proportion to where it is going: a private brainstorm needs a glance, a client deliverable needs every claim traced, and anything legal, financial or safety-related needs a qualified reviewer. The policy should name the checker for each kind of output, usually the person who will send or publish it.
A worked example from a fictional 15-person recruitment consultancy, Fenwick Talent. A consultant prompts: "Write a 150-word market update for finance directors on hiring for management accountants in Leeds." The draft says demand rose sharply this year and that salaries are up by a specific percentage, citing a named salary survey. The consultant cannot find the survey or the figure.
Weak output kept as is: "Demand for management accountants in Leeds rose sharply this year, with salaries up [figure] according to [survey]." Reviewed version: "Across the roles we filled this quarter, finance directors asked most often about notice periods and hybrid terms. We are seeing more counter-offers at offer stage." The reviewed version uses only what Fenwick observed itself. The policy clause behind the edit: "Remove any statistic, survey or result you cannot open and confirm."
Want to see how this connects to a role?
Explore the planned team trial ↗When should employees disclose that content was AI-generated?
At minimum, employees should tell their reviewer when a draft was substantially generated, and follow any client contract, platform rule or publication rule that requires disclosure. Beyond that, disclosure is a company choice that depends on audience and context.
Public examples vary. The City of Boston's interim guidelines for staff, published in May 2023, ask employees to disclose AI use and include the model used, and suggest a credit line such as "This description was generated by ChatGPT 3.5 and edited by" the staff member. That suits a public body. A small agency may decide internal disclosure to the reviewer is enough for ordinary copy, with external labelling for generated images or synthetic voices.
Sample clause: "Tell your reviewer when AI produced a substantial part of a draft. Label customer-facing images, audio or video that are AI-generated as [wording]. Follow any client instruction on AI use, which takes priority over this clause."
How should a generative AI policy handle copyright and images?
The policy should stop staff from deliberately reproducing protected work, require licence checks before generated code or assets are reused, and ban generated images or voices of real people without consent. NIST's profile lists eased replication of copyrighted, trademarked or licensed content, and eased exposure of trade secrets, among generative AI risks.
Copyright law on AI-generated material differs by country and is still developing, so the policy should not promise that the company owns or can register every generated output. It should say how the company handles uncertainty: for example, generated images are used for internal mock-ups freely, but anything for a client campaign is checked against the tool's licence terms and the client's instructions first.
Sample clauses: "Do not prompt for content in the style of a named living artist or brand, or ask a tool to reproduce copyrighted text. Do not create images, audio or video of a real person without their written consent. Check the tool's commercial-use terms before generated assets go into client work." Take advice on the position in your jurisdiction.
Which decisions should stay with a person?
Decisions with a significant effect on someone, such as hiring, pay, performance, credit, complaints and anything involving health or safety, should stay with a named person who reviews the underlying information. Generative tools can help draft, structure or summarise for those decisions; they should not make them.
NIST's profile describes automation bias and over-reliance as part of the human-AI configuration risk. A policy counters it with a simple test: the person deciding must be able to explain the decision without referring to what the tool said. Where personal data is involved, the ICO's guidance on AI and data protection is the official UK reference; its pages currently say it is under review following the Data (Use and Access) Act.
When is a generative AI policy the wrong tool?
A staff-facing generative AI policy is not enough if you are building generative features into a product, fine-tuning models on customer data, or deploying a chatbot that talks to the public on your behalf. Those need product, security and legal review that goes well beyond employee rules.
It also cannot do the work of practice. Knowing that a draft might contain an invented figure is different from spotting one under deadline. Tosheo.ai is being built for that practice step: planned role paths end each lesson with a hands-on task in a real AI tool, graded against a visible rubric. It does not draft or enforce your policy.
Sources and scope
NIST, Generative AI Profile (NIST AI 600-1), July 2024 ↗
NCSC, ChatGPT and large language models: what's the risk?, 14 March 2023 ↗
City of Boston, Interim Guidelines for Using Generative AI, version 1.1, 18 May 2023 ↗
ICO, Guidance on AI and data protection (under review), checked 16 October 2026 ↗