AI tools at work, used with care.

What these tools get wrong

Generative AI tools write fluent, confident text. Fluent is not the same as true.

The UK National Cyber Security Centre lists the flaws plainly: these tools can get things wrong and present incorrect statements as facts (often called hallucination), can be biased, are often gullible when asked leading questions, and can be manipulated by text hidden in what they read, known as prompt injection.

• They can invent references, quotes, case law and statistics that look real.
• They can be out of date, and often do not say so.
• They reflect patterns in what they were trained on, including stereotypes.
• A tool that reads your email or documents can be steered by instructions planted in a message or web page.

What not to paste in

Once information is in someone else’s system, you no longer fully control it.

• Personal information about the people you serve, supporters, volunteers and staff.
• Health, case and safeguarding information, immigration status, criminal records, and anything about children.
• Passwords, keys and financial details: account numbers, card details, login codes.
• Confidential documents: contracts, board papers, partners’ unpublished material, anything under a confidentiality agreement.

Australia’s privacy regulator, the OAIC, recommends as best practice that organizations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools. It also points out that privacy obligations apply both to what you put in and to any personal information the tool produces, including information that is wrong.

Removing names is not enough. “The 14-year-old at our only rural site” identifies someone. See the de-identification section of data you can report honestly.

Uses that are usually fine

• Drafting from public information: a job advert, a social post about a public event.
• Rewriting your own non-confidential text for clarity or a different reading level.
• Summarising published reports and guidance, then checking the summary against the original.
• Help with spreadsheet formulas or code, using made-up example data.

Read the vendor’s terms

The most important question is what the vendor may do with what you type. The answer can change.

Ask these questions of any AI tool before staff use it for work, and get the answers from the terms or in writing:

1. Does the vendor use what we type, and what it produces, to train its models? Is that different for free, personal and business accounts?
2. Can we switch training off for our whole organization, and is that the default on the plan we are buying?
3. How long are our prompts and files kept, and can we delete them?
4. Can the vendor’s staff read our prompts, for example to review abuse, and under what controls?
5. Where is our data processed and stored, and which other companies handle it?
6. Do we get administrator controls: who can use it, which features, audit logs?
7. How will the vendor tell us before these terms change?

Terms change. The US Federal Trade Commission has warned that quietly changing terms of service or a privacy policy to allow more permissive use of data, such as using it for AI training, may be unfair or deceptive. That warning is not a guarantee it will not happen: re-check the terms when you renew and when the vendor announces changes. Our before you sign guide covers the wider contract.

Check every output

Treat it as a first draft from a confident stranger.

• Check facts against the original source. Figures, names, dates, quotes, references and anything legal or medical. If you cannot find the source, do not use the claim.
• A named person is responsible. Whoever sends it owns it, as if they had written every word.
• Never unread. Nothing goes to a client, funder, partner or the public without a person reading and correcting it.
• Keep people in charge of decisions about people. Eligibility, funding, hiring, safeguarding and complaints need human judgement. Several countries’ data protection laws have specific rules on decisions made about people by automated means.

Accessibility and bias

These tools can widen access, and they can quietly shut people out.

• Bias. Outputs can reproduce stereotypes about age, gender, race, disability and class. NIST’s work on bias in AI warns that biases remain common across technology processes and can cause harm regardless of intent. Read drafts about people with that in mind.
• Image descriptions and captions. Automatic alt text and captions are often wrong or miss the point. Good alt text depends on why the image is there, as the W3C’s images tutorial explains, so a person should check it.
• Translation. Have a fluent speaker check anything people will rely on: eligibility, consent, health or legal information.
• A human route. If you add a chatbot to your website, keep a clear way to reach a person, by phone, email or in person.
• Assistive use. Some staff use these tools for dictation, reading support or summarising. Your policy should allow that, within the same data rules.

A one-page addition to your acceptable use policy

Short enough to read, specific enough to follow. Adapt the words in square brackets.

It is written to sit inside the acceptable use policy in our templates. Keep the approved list next to it: each tool, the type of account, and what it may be used for.

Download the policy addition (.txt)

If you need to go further

For larger or higher-risk uses, these are the reference frameworks.

• NIST AI Risk Management Framework (January 2023): a voluntary framework for managing risks across the life of an AI system, and its Generative AI Profile, NIST AI 600-1 (July 2024).
• ICO guidance on AI and data protection: fairness, transparency, accuracy and security when AI systems use personal data, under UK law.
• OAIC guidance on commercially available AI products: choosing and using off-the-shelf AI tools under Australian privacy law.