What are the main risks?
AI assistants help with drafting, summarizing and translating. They also bring four risks that every NGO should name.
- Wrong answers. The assistant can produce fluent text that is false. It may invent a number, a source or a rule.
- Data exposure. Text you type is sent to a provider. If it contains personal or sensitive data, you have shared that data.
- Bias and tone. Output can reflect stereotypes or use words that hurt the people you serve.
- Over-trust. Busy staff accept a polished draft without checking it.
None of these means you must avoid AI. They mean you need rules.
The risks are not equal. Wrong answers are common but easy to catch if you check. Data exposure is rare but hard to undo, because you cannot take a prompt back once it is sent. So spend most of your effort on what goes into a prompt, and keep a steady habit of checking what comes out.
What should never go into a prompt?
Make a short list and put it where staff will see it. Unless you have approved the provider, its terms and your legal basis, keep these out:
- names, photos and identity numbers of beneficiaries,
- safeguarding, protection and complaint details,
- health information,
- staff disciplinary and salary records,
- passwords, keys and bank details,
- anything covered by a confidentiality clause in a donor or partner agreement.
When you need help with a sensitive text, replace names with codes and remove locations before you paste it. Read the result again to check that the story itself does not identify someone, because a rare event in a small village can point to a person even without a name. Our article on handling sensitive program data explains how to classify data.
How do you check the output?
Use a short routine every time.
- Check every number against the source record.
- Check every claim about a rule, a donor requirement or a law. Do not accept an invented citation.
- Read for tone. Does it respect the people described?
- Edit in your own voice. A draft is a starting point.
- Keep a person accountable. The staff member who uses the text owns it, not the tool.
For reports sent to donors, add a second reader who has not seen the prompt.
Translation needs the same care. A machine translation may sound fluent and still change the meaning of a key word, such as a legal or medical term. For anything that affects a decision or a person, ask a fluent speaker to check it. Say so openly when a document was translated or drafted with AI help, if your donors or partners expect that.
What does a good use policy say?
Keep it to one page. It should cover:
- Who may use it, and on which tools.
- For what tasks, such as drafting, summarizing public text or translating a non-sensitive note.
- Which data is banned, using the list above.
- How output is checked and who signs off.
- Where to report a mistake or a data slip.
Review the policy every six months at first, because tools and risks change quickly. Roll it out like any policy and keep a record that staff have read it. See our article on rolling out policies.
What should you ask your vendor?
If your software has an AI assistant, ask:
- Which provider processes the prompts, and where?
- Can we use our own provider account or key instead?
- Is our content used to train models?
- Is the assistant limited to the data each user is allowed to see?
- What limits and costs apply per user?
- Which providers and terms apply if we switch provider?
Ask for the answers in writing, and keep them with your data processing agreement. Ask too how the assistant treats permissions: it should not show a user records that the person could not open in the system. Test this yourself with a low-access account before you let staff rely on it. Law differs by country, so check your duties with a qualified adviser.
Where EMPOBase fits
EMPOBase includes a project-aware AI assistant that can help with indicator suggestions, report drafting and summaries. AI features can send prompts and selected context to Google. Supported customer-configured provider keys can send requests to Anthropic or OpenAI instead, and that option is on every plan. An AI provider key is not an encryption key. We do not use customer content to train models. AI output can be inaccurate, so review it against your records, and review the provider and its terms before you send sensitive information. See the documentation and the privacy page.