Teaching AI Skills to the People You Serve
Most nonprofit conversations about AI literacy are about staff. There is a second conversation, much less developed, about whether AI skills belong in what you deliver to participants. A workforce program whose graduates cannot use the tools employers now assume, a legal aid client facing an AI-screened application, a parent whose child's school just deployed an AI tutor: these are program problems, not IT problems. Deciding whether to take them on is a real strategic choice, and it deserves better than a one-off workshop.
There is a pattern that repeats whenever a technology becomes economically consequential. First it changes what employers expect. Then it changes what public systems assume of applicants. Then, several years later, community organizations start teaching it, usually after the gap has already done its damage. It happened with basic computer skills in the nineties, with smartphones and app-based services in the twenty-tens, and it is happening now with AI.
Nonprofits are unusually well positioned to close that gap, and unusually likely to do it badly. Well positioned, because you already have trust, physical space, and standing relationships with the people furthest from the technology. Likely to do it badly, because the easiest version of this program is a two-hour workshop where someone demonstrates a chatbot, everyone nods, and nothing measurable changes. That version is cheap, popular with funders, and close to useless.
The difference between a real AI literacy program and a demonstration is the difference between teaching a tool and teaching a capability. Tools change every few months. The underlying capabilities, which are knowing when a machine-generated answer is likely to be wrong, understanding what happens to information you type into a system you do not control, being able to specify what you want precisely enough to get it, and recognizing when an automated decision has been made about you, are durable. A curriculum built on the second set survives the next model release. A curriculum built on the first is obsolete before the grant report is due.
This article works through the decision and the design. Whether AI literacy actually belongs in your program portfolio, which of the three very different programs hiding under that label you would be running, how to find out what your community is actually asking for rather than assuming, what a curriculum looks like that does not expire, how to fund and staff it, how to measure something a funder will accept, and the ethical traps that catch well-intentioned organizations. If your interest is in what other organizations have already tried, our survey of AI literacy programs for underserved communities covers the landscape.
The Question Is Not Whether It Matters. It Is Whether It Is Yours.
AI literacy is important for almost everyone, which is precisely why it is a dangerous thing to add to a program portfolio. Universal importance is not the same as mission fit, and nonprofits have a long history of adding programs because a topic is urgent and funding is briefly available. The result is usually a small underfunded offering that competes with the core work for staff attention and quietly ends when the grant does.
The useful test is whether AI is already interfering with the outcome you exist to produce. A workforce development organization whose placement rate depends on candidates clearing automated resume screening has a direct interest. A legal aid organization whose clients are receiving benefits determinations influenced by automated systems has a direct interest. A senior services organization watching participants get targeted by increasingly convincing voice-cloned scam calls has a direct interest. An arts organization probably does not, however much its participants might personally benefit from a class.
The second test is whether you would still be the right provider if a better-resourced institution showed up. Community colleges, public libraries, and workforce boards are all expanding into this space, often with dedicated funding streams and instructional staff you cannot match. If the local library is already running free AI classes on Saturdays, the highest-value thing your organization can do may be to get your participants there and provide the childcare and transportation that makes attendance possible. That is a less impressive line in an annual report and frequently a better use of your capacity.
Where nonprofits have a genuine and hard-to-replicate advantage is with populations that institutional providers reach poorly. People who will not walk into a college building, people who need instruction in a language the community college does not offer, people whose immigration status makes them cautious about formal enrollment, people managing a disability that standard classes do not accommodate. If your participants are in that group, the case for running it yourself is strong, and it is essentially the same case that justified your other direct services.
Five questions before you add the program
If you cannot answer these, you are not ready to fund it
- Which outcome we already report on would improve if participants had these skills?
- Who else in our area already offers this, and can we refer instead of build?
- Do our participants have the devices and connectivity to practice between sessions?
- Who on staff can teach this credibly, and what does it cost to get them ready?
- If the grant funding this ends in eighteen months, does the program end with it?
Three Different Programs Hide Under One Label
A great deal of confusion in this area comes from treating AI literacy as a single subject. It is at least three, they serve different populations, they require different instructors, and they produce different outcomes. Organizations that do not choose one tend to produce a survey course that does all three shallowly and none of them well.
The first is economic. This is AI as a job skill, aimed at people who need to be employable in a labor market where employers increasingly assume familiarity. The content is practical and tool-adjacent: using an assistant to draft and revise, understanding what tasks in a given occupation are changing, and being able to speak about the technology in an interview without either overclaiming or freezing. The outcome is employment or wage progression, which is measurable and fundable, and this version pairs naturally with existing workforce programming.
The second is protective. This is AI as a threat surface, aimed at people who are being acted upon by these systems rather than using them. Recognizing synthetic voice and video in scams, understanding that a rental application or a benefits determination may have been screened algorithmically, knowing what recourse exists, and understanding what happens to personal information entered into a free tool. The population here skews older and more vulnerable, and the outcome is harm avoided, which is real but notoriously hard to measure.
The third is civic. This is AI as something communities have a right to shape, aimed at residents, advocates, and organizers who need enough fluency to participate in decisions being made about deployment in their schools, their housing systems, and their local government. The content is less about using tools and more about how these systems are procured, what questions to ask at a public meeting, and where accountability actually sits. The outcome is participation, and this version fits organizations that already do advocacy and organizing rather than direct instruction.
Most organizations should pick one and let the others show up as brief modules. The economic version is the easiest to fund and the most competitive. The protective version is the most needed and the hardest to sustain, because prevention rarely produces the numbers funders reward. The civic version is the least crowded and the best fit for organizations whose credibility rests on advocacy rather than service delivery.
Economic
AI as a job skill
Job seekers and incumbent workers. Practical tool use tied to specific occupations. Outcome is placement, retention, or wage gain. Funds most easily through workforce channels.
Protective
AI as a threat surface
Older adults, benefits recipients, tenants, and applicants. Scam recognition, automated decisions, data exposure, and recourse. Outcome is harm avoided, which resists easy counting.
Civic
AI as a public decision
Residents, parents, and organizers. How systems get procured and deployed locally, what to ask, and where accountability sits. Outcome is participation in decisions already being made.
Access Comes Before Literacy, and It Is Frequently the Real Problem
A recurring failure in digital inclusion work is teaching a skill that participants cannot practice. Someone attends a well-run session, understands it, goes home to a shared phone with a limited data plan, and has no opportunity to use anything they learned before it fades. Two months later the program reports thirty people trained and the participants report that they do not really use it. Nothing about the instruction was wrong. The precondition was missing.
Before designing curriculum, find out what your participants actually have. Whether they have a personal device or a shared one, whether they have home broadband or only mobile data, whether the device is new enough to run current applications, whether they have an email address they control and can reset, and whether they are comfortable creating an account with a phone number. That last item stops more people than any conceptual difficulty, particularly among participants who are cautious about giving identifying information to any system.
Language is the other precondition that gets assumed away. Most consumer AI tools handle major languages competently and handle less-resourced languages, regional dialects, and code-switching much less well. A participant whose first language is Hmong, Haitian Creole, or an Indigenous language may find the tool itself performs poorly for them, which is a substantive lesson worth teaching rather than an inconvenience to work around. Our discussion of serving communities on both sides of the digital divide goes deeper on designing for uneven access.
Where access is the binding constraint, the honest program design is different. Device lending, a supervised lab with drop-in hours, connectivity subsidies, and enrollment assistance for low-cost broadband programs may deliver more benefit than any curriculum. This is not a lesser program. It is the intervention that matches the actual barrier, and it makes the eventual instruction stick when you get to it.
Ask before you design, not after
A short conversation with fifteen current participants about what they already do with their phones, what confuses them, and what has gone wrong will reshape a curriculum more than a month of planning. It also frequently reveals that the topic they most want is not the one you assumed.
The most common surprise is that participants are less interested in productivity and much more interested in protection. People want to know whether the call from their grandchild was real.
Building a Curriculum That Survives the Tool Changing
The design constraint that matters most is obsolescence. Any curriculum organized around the interface of a specific product will be wrong within a year, and rewriting it annually is not feasible for a program with one part-time instructor. The way out is to organize around durable competencies and treat the specific tool as the example rather than the subject, in the same way that a financial literacy class teaches budgeting rather than a particular banking app.
Four competencies have held up so far and are unlikely to change soon. The first is verification, meaning the habit of treating a generated answer as a claim requiring a check rather than as a result. This is the single highest-value thing to teach, and it is teachable through practice rather than explanation: have participants deliberately ask about something they know well and watch the model make a confident error. Nothing produces appropriate skepticism faster than seeing a system be fluently wrong about your own neighborhood.
The second is specification, which is the skill of describing what you want with enough context and constraint to get it. This is genuinely a communication skill rather than a technical one, and it transfers to writing emails, briefing a contractor, and explaining a problem to a caseworker. The third is disclosure judgment, meaning an understanding of what happens to information typed into a system and a working rule about what not to enter. The fourth is provenance, meaning the ability to reason about whether a piece of media or a message is what it claims to be.
For programs that need an external framework to point at, the Department of Labor has published an AI literacy framework built around foundational areas that map reasonably well onto community instruction, and we walked through its five foundational areas as a training curriculum in a previous piece. Aligning to a recognized framework is worth doing mainly because it makes the program legible to workforce funders, who are often required to fund against established standards rather than a curriculum you wrote yourself.
Session structure should be short, repeated, and hands-on. A single three-hour workshop is the least effective format and the most common, because it is the easiest to schedule. Four ninety-minute sessions across four weeks, with a small task to attempt in between, produces retention that a one-off cannot. Where attendance is unreliable, which it usually is, design each session to stand alone rather than building sequentially, and accept that some participants will attend two of four.
The four durable competencies
Teach these, and let the specific tool be an example
- Verification. Treat every generated answer as a claim to be checked, and know two ways to check it.
- Specification. Describe the task, the audience, the constraints, and the format you want back.
- Disclosure judgment. Know what a free tool does with what you type, and hold a personal rule about what never goes in.
- Provenance. Ask whether a message, voice, or image is what it claims to be, and know how to confirm through a second channel.
The Teaching Problem Is That Your Staff Are Not Ready Either
The uncomfortable constraint on most community AI literacy programs is instructor capacity. The people who know your participants and hold their trust are program staff, and most program staff are themselves early in their own understanding of these tools. The people who understand the technology well are typically external and lack the relationship that makes community instruction work. Neither group can deliver the program alone.
The pairing that works is a program staff member as the primary instructor with a technically fluent person as preparation support rather than as co-teacher. The staff member spends time with the material until they can teach it in their own words and answer the obvious questions. The technical partner builds the exercises, prepares answers to the hard questions, and is reachable when something unexpected comes up. Putting the technical person in front of the room usually goes worse than expected, because the failure mode is answering questions nobody asked at a level nobody needed.
Instructor preparation should be budgeted explicitly rather than absorbed. A realistic figure is that a staff member needs several hours of practice for each hour they will teach, front-loaded before the first cohort and reducing sharply after. Organizations that skip this end up with an instructor reading slides they do not understand, which participants detect immediately and which does more damage to credibility than not running the program. Our guidance on building an AI training program when you are not technical yourself addresses this preparation gap directly.
Peer instruction is worth considering once you have run a few cohorts. Participants who completed the program and can teach it in the community's own language, with the credibility of having recently learned it, are frequently better instructors than staff. This model requires a stipend, real support, and a willingness to accept a somewhat less polished delivery, and it tends to produce better attendance and better retention than anything a staff member can achieve on their own.
What instructor preparation actually involves
- Using the tools daily for their own work for at least a month before teaching anything.
- Deliberately producing failures so they can demonstrate them on request rather than hoping one occurs.
- A written answer to the ten questions participants ask most, including the ones about jobs disappearing.
- Explicit permission to say they do not know, plus a route to find out before the next session.
Paying for It Without Becoming a Grant Chaser
The good news is that funding for this work exists and has broadened considerably. The less good news is that most of it is short-term, project-based, and attached to outcome requirements that a thoughtfully designed program may not naturally produce. Understanding which pot you are drawing from shapes the program more than most organizations expect, and it is better to know that going in than to discover it in the first report.
State and local digital equity funding is often the most natural fit, particularly for programs housed in or partnered with libraries. The California State Library, for example, administers digital literacy and access grants that support public digital literacy training including AI and internet safety instruction. Several state labor departments have run digital literacy training grants aimed at community-based organizations delivering skills training with connections to employment. These programs vary substantially by state and change from year to year, so the practical step is checking your own state's library agency, labor department, and broadband office rather than assuming national availability.
Workforce funding is the largest and most demanding source. Money flowing through local workforce boards typically requires participants to meet eligibility criteria, requires documented placement or credential outcomes, and requires reporting infrastructure that a small organization may not have. It is worth pursuing when your program is genuinely the economic version described earlier and worth avoiding when it is not, because the reporting burden will quietly reshape a protective or civic program into a job-training program that serves your original population poorly.
Corporate and sector-led initiatives are the third channel and have expanded quickly. The AI for Nonprofits Sprint, backed by a coalition of technology companies and foundations, has set out to bring large numbers of nonprofit staff to baseline AI literacy, and similar efforts exist from several technology providers. These are more useful for building your own staff capacity than for funding community delivery, but the capacity is a genuine precondition and the programs are usually free. Our roundup of free AI training resources every nonprofit should use lists options that cost nothing but time.
Whatever the source, the sustainability question deserves an honest answer before launch. A program funded entirely by a two-year grant will end in two years unless someone has a plan, and community trust is damaged more by a service that appears and disappears than by one that never existed. If the realistic answer is that this is a time-limited pilot, say so to participants at the outset rather than letting them assume permanence.
Digital equity and library funding
Best fit for protective and general literacy
State library agencies, broadband offices, and municipal digital inclusion funds. Outcome expectations are usually participation-based, which suits programs whose value is harm avoided. Check your own state, since availability varies widely.
Workforce funding
Largest source, heaviest requirements
Local workforce boards and state labor departments. Expect eligibility screening, credential or placement outcomes, and substantial reporting. Take it only if your program really is job-focused.
Measuring Something Other Than Attendance
Nearly every AI literacy program reports the same three numbers: sessions held, people trained, and satisfaction score. All three are outputs, none tells you whether the program worked, and funders have started noticing. Building a measurement approach that captures actual capability change is not difficult, but it has to be designed before the first session rather than assembled at reporting time.
The most practical instrument is a short scenario-based assessment given at intake and again at the end. Not a knowledge quiz, which measures recall, but a handful of situations with a right answer that depends on the competencies you taught. Show a fabricated but plausible generated answer containing an error and ask what the participant would do next. Present a message from a supposed family member requesting money urgently and ask how they would confirm. Give a task and ask them to write what they would type. Scored consistently, this produces a defensible before-and-after that means something.
Behavioral follow-up at sixty or ninety days is the second component and the one most programs skip because it is logistically annoying. A five-minute phone call asking what they have used since, what worked, and what they gave up on tells you more about program effect than any exit survey. It also surfaces the access problems described earlier, because the most common answer from a program that looked successful on paper is that the participant has not used it since, and that is the finding you need.
For the protective version, measurement is genuinely hard and honesty is the best policy. You cannot count scams that did not happen. What you can capture is self-reported incidents encountered and what the participant did, which at least tells you whether the recognition skill is being exercised. Reporting this transparently, with an explanation of why harm avoidance resists counting, generally lands better with thoughtful funders than a made-up proxy metric. If your organization is building measurement practice more broadly, our piece on using AI in strategic planning covers how program metrics connect to organizational goals.
A measurement set that fits a small program
Four instruments, none requiring an evaluator
- Scenario assessment at intake and completion, five situations, scored against a fixed rubric.
- Confidence self-rating on each taught competency, which is soft but tracks willingness to try.
- A ninety-day phone call on actual use, with an explicit question about what they stopped doing.
- Program-specific outcome tied to your existing reporting, such as placements or benefits retained.
The Traps That Catch Well-Meaning Programs
Vendor capture is the most common. A technology company offers curriculum, licenses, staff time, and sometimes funding, and the resulting program teaches that company's product as though it were the category. Participants leave able to use one interface and unable to evaluate alternatives, which serves the sponsor and underserves the community. Accepting corporate support is reasonable and often necessary, but the curriculum should cover at least two comparable tools and should include an explicit segment on how to evaluate a new one. If a sponsor objects to that, the objection tells you what the partnership is.
The second trap is teaching enthusiasm rather than judgment. Programs designed to reduce fear of technology can slide into promoting uncritical trust, which is a particular risk when the population being taught is one that automated systems already treat badly. Teaching a tenant to use an AI assistant while saying nothing about algorithmic tenant screening produces someone who is more comfortable with a technology that is being used to exclude them. The protective content is not a downer to be tacked on at the end. It is part of the competency.
The third is consent around participant material. Programs frequently want to use participant work, quotes, and images in reports and funder communications, and the request tends to be made informally in a room where declining is socially awkward. Written, specific, revocable consent obtained separately from the session, with a clear statement that participation does not depend on agreeing, is the minimum. This matters more than usual here because AI literacy programs often work with populations who have reason to avoid visibility.
The fourth is scope creep into advice you are not qualified to give. Participants will bring real problems: a benefits denial, an immigration question, a dispute with a landlord, a health concern raised by something a chatbot told them. An instructor who starts answering those questions has moved from teaching to practicing, and in several of those areas that is both risky and possibly unlawful. Decide the referral routes before the first session and give the instructor a script for redirecting, because it will come up.
The fifth is running the program on tools your own organization would not accept. If staff are barred from entering client information into a consumer chatbot, teaching participants to enter their own sensitive information into the same tool without discussing the tradeoff is inconsistent. The same reasoning that produced your internal policy should show up in the curriculum, in language participants can use for their own decisions.
A program that only creates users is not literacy
The test of whether you have taught literacy rather than adoption is simple. Can a graduate explain a situation in which they would decline to use these tools, and give a reason that is theirs rather than one you supplied?
If every graduate leaves enthusiastic and none leaves appropriately wary, the curriculum has a hole in it, whatever the satisfaction scores say.
Conclusion
Teaching AI skills to the people you serve is a program decision, not a technology decision, and it should be made the way you would make any other program decision. Does this connect to an outcome we already own? Are we the right provider, or should we be referring? Do the preconditions exist for participants to actually use what we teach? Can we staff it credibly and pay for it past the first grant? Organizations that answer those questions honestly will run better programs than organizations that answer the question of whether AI is important, which was never really in doubt.
Where the answer is yes, the design principles are reasonably settled. Choose one of the three programs rather than attempting all three. Build the curriculum around verification, specification, disclosure judgment, and provenance rather than around any product. Solve access before instruction, because a skill that cannot be practiced does not persist. Pair a trusted staff member with technical support rather than importing an expert. Measure capability change with scenarios and a follow-up call rather than counting attendance. Keep the protective content in even when the funder is paying for the economic version.
The organizations that do this well will not be the ones with the most sophisticated curriculum. They will be the ones whose participants come back a year later still using something they learned, and who can say plainly which situations they would not trust these systems with. That is a modest-sounding outcome and a genuinely hard one to produce, and it is worth considerably more than a large number of people trained.
If you are unsure whether this belongs in your portfolio, the cheapest possible test is a conversation with fifteen participants about what they already encounter and what worries them. The answers will tell you whether there is a program here, which of the three it is, and whether anyone is asking you to run it.
Thinking About Adding AI Literacy to Your Programs?
We help nonprofits decide whether this belongs in their portfolio, then design curriculum, instructor preparation, and measurement that hold up to a funder and to the community.