The 4 Things Every Charity Staff Member Needs to Understand About AI
Forget prompting tutorials. Every person on your charity payroll needs to understand four things about AI before they use it for anything. The plain-language briefing for staff who never had the training.
Forget the prompting tutorials. Every person on your charity's payroll needs to understand four things about AI before they use it for anything. Here they are, in plain language, for the staff member who has not had the training yet.
Who this is for
This article is for the charity staff member who has been told their team should "be more AI literate" and has no idea what that means.
You might be a fundraiser, a comms officer, a service manager, a finance officer, a volunteer coordinator, a CEO, or anything in between. You have probably used ChatGPT at least once. You may have used Microsoft Copilot, Claude, or Gemini. You may not.
You do not need to be a technologist to use AI safely in a charity context. You need to understand four specific things. Once you have those, you have what the EU AI Act calls a "sufficient level" of AI literacy for general use. Specific roles need more depth, but everyone needs these four.
This is not training. It is a primer. Read it and you will know what your role is in the AI conversation your charity is starting to have.
Thing 1: What AI actually is (and isn't)
Most charity AI conversations skip this and assume everyone knows. Most people do not.
What AI is, in practice. The AI tools your charity is using (ChatGPT, Microsoft Copilot, Claude, Gemini, Fireflies, Granola, and the rest) are large language models. They generate text by predicting which words are most likely to come next, based on patterns learned from enormous amounts of training data.
They are not databases. They do not look things up. They do not "know" things in the way a search engine does. They generate output that sounds plausible based on patterns.
This matters because it explains the failure modes.
What AI is good at. Things where pattern-matching produces useful output:
- Drafting and editing text
- Summarising long documents
- Translating between languages
- Suggesting alternatives ("rewrite this in a friendlier tone")
- Brainstorming and structured thinking
- Explaining concepts
What AI is not good at. Things where pattern-matching produces wrong but plausible output:
- Stating facts accurately, especially specific numbers, dates, names, or quotes
- Reasoning about anything not present in its training data
- Knowing what it does not know (it confidently makes things up)
- Anything requiring genuine current information unless connected to live data
The most important distinction. Search engines retrieve. AI generates. When you ask Google for the population of Manchester, it shows you a source. When you ask ChatGPT, it generates an answer based on patterns. The answer might be right. It might be plausibly wrong. You cannot tell from the answer itself.
This is the foundation of everything else.
Thing 2: Where it fails (and what that looks like in charity work)
AI's failure modes are not random. They are predictable. Understanding them is the difference between using AI safely and being caught out by it.
Hallucinations. AI generates content that sounds confident but is factually wrong. It invents statistics, citations, names, quotes, dates, references. The hallucinations are not flagged as such; they are produced with the same confidence as accurate output.
Charity example: a fundraiser asks AI to draft a grant application referencing the prevalence of a condition in a target area. The AI generates a specific statistic ("affecting 1 in 8 people in the borough"). The statistic is fabricated. The funder spots it. The application is declined and the relationship damaged.
Bias. AI systems reflect the biases in their training data. Outputs can systematically under-represent or stereotype certain groups, particularly those underrepresented in the training data.
Charity example: a charity uses AI to summarise case notes for handover meetings. The summaries consistently soften the urgency of cases involving certain demographic groups, reflecting bias in the underlying model's training. The pattern is invisible until someone audits the summaries against the original notes.
Data leakage. When you put information into an AI tool, that information may be stored, processed, or used in ways you did not expect. Free AI tools in particular often retain inputs, sometimes for model training, sometimes for indefinite storage.
Charity example: a service manager pastes a case note into a free AI tool to summarise it. The case note includes the beneficiary's name, address, and circumstances. That data is now stored by the AI provider. Depending on the tool, it may be used to train future models or accessed by the provider's staff. The charity has just breached its data protection obligations without realising.
Sycophancy. AI is trained to be helpful. That training can lead it to agree with you, validate your assumptions, or tell you what you want to hear rather than what is accurate.
Charity example: a CEO uses AI to "stress test" a strategy. The AI provides a plausible critique but ultimately endorses the direction. The CEO presents this to trustees as external validation. The endorsement is the AI being agreeable, not the AI being analytical.
Out-of-date answers. AI models have a training cut-off date. Anything that has happened after that date is unknown to the model unless it has live web access. Even then, the model will often confidently answer questions about current events with information that is wrong.
Charity example: a comms officer asks AI for the current advice on charity fundraising regulations. The AI confidently provides advice that is six months out of date. The comms officer publishes content that misrepresents current regulator expectations.
Security risks. AI tools are increasingly the targets of various forms of attack. Prompt injection (where instructions hidden in input data hijack the AI's behaviour) is a growing concern, particularly for any AI that processes external content (emails, documents, web pages).
Knowing these six failure modes does not eliminate them. It does mean you can recognise them when they show up, and avoid the worst consequences.
Thing 3: Your role and what's allowed
This is where AI literacy becomes specific to your charity. The general principles above are universal. The specific rules depend on your role and your charity's policy.
Every charity using AI should have answers to these questions. If yours does not, that itself is a sign of a governance gap.
Which tools are approved? Your charity should have an approved tools list. Anything not on the list, you do not use for charity work. If you are not sure what is approved, ask your manager or your CEO.
What data can I put in? Your charity should have data classifications. Common categories:
- Public information (newsletters, published reports): generally fine
- Internal business information (strategy documents, internal communications): tool-dependent, usually fine in approved tools
- Personal data of staff or volunteers: only in tools with appropriate data protection guarantees
- Beneficiary data: usually off-limits for general AI tools, only in specifically approved systems
- Safeguarding-relevant information: almost always off-limits for general AI tools, no exceptions
- Donor financial data: regulated separately, off-limits for general AI tools
If your charity has not defined these categories, default to caution. Anything you would not paste into a public LinkedIn post probably does not belong in a free AI tool.
When do I need someone to review the output? Your charity should have guidance on review thresholds. Common standards:
- Any external communication: always reviewed
- Any content involving beneficiaries: always reviewed
- Any factual claims: always verified
- Internal drafts: lighter review acceptable
- Personal use: your own judgement
Who do I go to with questions? Your charity should have a named person for AI questions. It is usually the operations lead, the digital lead, or the CEO in smaller charities. Know who it is. Use them.
What do I do if something goes wrong? Your charity should have an incident process. If you spot a hallucination that went out the door, or realised you put data somewhere you should not have, or noticed AI behaving unexpectedly, there should be a clear escalation pathway. Use it. The cost of reporting a near-miss is small. The cost of hiding one until it becomes a problem is enormous.
Thing 4: When to escalate
This is the shortest section and the most important.
Escalate to your manager or your CEO if:
You see AI output that looks wrong. Better to ask than assume. False alarms are fine. Missed errors are not.
You realise data has gone somewhere it should not. Quickly. The first hour matters. Reporting fast lets the charity assess and contain. Hiding it makes everything worse.
A new AI tool is being suggested or rolled out. Ask whether it has been reviewed. The [7-Lens Tool Review Framework] exists for this reason. If your charity uses the Tool Review Agent, the answer to "has it been reviewed" is documented and quick to find.
You see a colleague doing something that worries you. Quietly, supportively, without blame. Most AI mistakes are made by people who did not know better. Most can be corrected with a conversation. The charity needs to know what is happening.
You are asked to do something with AI that does not feel right. Trust the instinct. Charities operate on trust and judgement. If a use case feels uncomfortable to you, that is worth raising. It might be fine. It might not. The conversation is what matters.
The cultural shift this requires
The four things above are not technical. They are cultural.
A charity where staff hide AI mistakes is a charity that does not learn. A charity where AI use happens in the shadows is a charity that cannot govern. A charity where staff are afraid to ask "is this approved" is a charity where the gap between policy and practice widens until something breaks.
Your role, if you are a charity staff member, is to make AI literacy a normal conversation in your team. Ask the questions. Raise the issues. Report the near-misses. Read the policy. Know who to escalate to.
None of this requires technical expertise. It requires the same kind of professional judgement charities apply to safeguarding, data protection, and finance. AI sits alongside those now.
Where this fits in the bigger picture
This article covers what every charity staff member needs to know. It is the foundation layer of the [4Ps Framework] applied at individual level.
Specific roles need more. A fundraiser needs depth on funder expectations and copy review. A service manager needs depth on beneficiary data and dignity considerations. A trustee needs depth on governance and oversight. Our guide on [why generic AI training fails] explains why role-specific layers exist on top of this foundation.
The foundation matters most. A charity where every staff member understands these four things has a working baseline. Without that baseline, role-specific training has nothing to build on.
Three things to do this week
If you are a charity staff member reading this:
Find out what your charity's AI policy says. If you have not seen one, ask. The answer to "we have one" matters less than the answer to "where can I read it."
Identify the approved tools. Use those. Stop using anything not approved for charity work.
Know who to ask. When AI questions come up, you should not be figuring it out alone. Find the named person for your charity.
If your charity has none of these in place, the issue is not your AI literacy. It is your charity's governance. The conversation needs to move up the chain.
The honest read
AI is not going away. The pressure to use it will increase. The standards expected of charity AI use will rise, both from regulators and from funders.
You do not need to become an AI expert. You need to be the kind of charity professional who treats AI as a tool that requires the same judgement you bring to everything else.
Four things. Read them. Apply them. Ask questions when you are not sure. That is what literacy looks like.
If you would like to share this with your team, you are welcome to. If your charity would benefit from structured AI literacy training, [book a call]. We work with charities of all sizes to build literacy that actually lands.
This article is the foundation layer of the GoodAgents 4Ps AI Literacy Framework. It is suitable for general charity staff orientation and is not a substitute for role-specific training where higher-stakes AI use is involved. Charities are welcome to use this article in their own onboarding and training materials with attribution.
Similar Posts
Twenty-eight per cent of charity boards admit poor digital skills. Most have had no briefing on AI. With Article 4 enforcement weeks away, this is the plain-English briefing your trustees should have had a year ago.
Two-hour ChatGPT masterclasses are the default in the charity sector — and the most common reason AI literacy fails to land. Why generic doesn't work and what role-specific training looks like.
When EU AI Act enforcement begins, regulators will ask how you decide which AI tools staff can use. A working tool approval process — and the evidence trail it produces — is your compliance defence.