What This Covers
This page is for staff who use generative AI assistants at work, not for people building AI systems. An AI system is a machine-based system that infers from its input how to generate outputs such as predictions, content, recommendations or decisions. One risk to know about is hallucination (also called confabulation), the confident production of false content that can mislead you. Below you will find the everyday tasks worth trying, what must stay out of public tools, how to check output, and the AI use policy to have in place before your team adopts anything.
Start Here
- Draft and edit text. Grant proposals, newsletters, thank-you notes and press releases are common uses. Get a first draft, then review it as a person. See AI for grant writing without inventing facts and using AI for donor reports.
- Pick a first use. Begin with a pain point. Time-consuming, extremely repetitive tasks are good candidates.
- Decide what never goes into a public tool. Personal and sensitive information stays out. In development and humanitarian work, disclosure can lead to vulnerable people and communities being further harmed or exploited. With free public models there is also the risk that staff paste in private donor information, which the provider can then store and read.
- Check every output. Treat it as a first draft and fact-check it by hand. The verification prompt helps you find invented facts.
- Run a small pilot. Keep it a time-limited prototype that staff assess. If you use a translation tool, test it thoroughly first so you know translations into your languages are accurate.
A person stays accountable and makes the final call.
Key Concepts
- Hallucination (confabulation): confident false content. It is a natural result of how generative models are designed, not a rare glitch.
- Automation bias: excessive deference to automated systems, which can make hallucination and bias worse.
- Performance gaps: generative AI can perform differently across sub-groups or languages, possibly because training data was not representative, which leads to wrong assumptions about how well it works for your users.
- Data responsibility: the safe, ethical and effective management of personal and non-personal data for operational response.
- Provider visibility: a query in a public chatbot is visible to the organization providing it. As of the source's writing, a model did not automatically add your query to what it tells other users, but stored queries may be read by the provider or its partners. Check your tool's current terms.
- Prompt injection: one of the attacks that generative AI tools are themselves vulnerable to.
- Human oversight: any system making critical decisions without human involvement needs ongoing human review, with documented processes and audit trails.
Templates
- AI use policy template: staff ground rules, including using output only as a first draft, fact-checking responses by hand, and sharing no confidential information with public models.
- Product due-diligence checklist (build your own): ask first whether you need the product at all. Then check it was tested for your intended use, how human oversight fits in, the privacy and security risks, and who can access the data. For cloud private models, ask whether vendor staff or partners can see data, and whether it is pooled with other organizations.
- Privacy or data impact assessment: build one before adopting a tool, as part of privacy by design. One guidance note on cash and voucher assistance lists data impact assessments among its recommended data-responsibility actions.
AI for This Topic
An assistant earns its place on repetitive tasks and bottlenecks, and on drafting and editing text. Before asking a public tool anything sensitive, read its terms of use and privacy policy; and remember that these systems are probabilistic and do not understand the data they handle.