AI for Adult Literacy and ESL Programs
Adult literacy and English language programs sit at an awkward intersection: the technology that could most obviously help them is also the technology most likely to be inaccessible to their learners. This guide separates the applications that genuinely extend a volunteer tutor's reach from the ones that quietly replace the human relationship the program depends on, and sets out what to put in place before either.
Adult literacy organizations operate under a constraint that most education nonprofits do not share. Their learners are adults with jobs, children, and commutes, which means instruction happens in the evening, attendance is irregular by necessity rather than by choice, and a learner who misses three weeks has not dropped out so much as had a difficult month. Progress is slow, dependent on consistency, and easily interrupted by circumstances that have nothing to do with motivation.
The need is large. According to the federal PIAAC assessment as summarized by APM Research Lab, roughly 130 million American adults have low literacy skills, a figure that includes both native and non-native English speakers and that has not improved in recent assessment cycles. Meanwhile the programs serving those adults are typically small, funded through a patchwork of state adult education dollars and local grants, and staffed by a handful of paid coordinators supporting a much larger corps of volunteer tutors.
Into this arrives a technology that is unusually good at exactly the things adult education is short of: generating leveled reading material, giving unlimited patient conversation practice, adapting to an individual learner's pace, and producing the lesson preparation that volunteer tutors have neither the training nor the time to do well. Funders have noticed. In Boston, First Literacy's most recent grant round included support for an AI-powered English learning program at the Boston Chinatown Neighborhood Center and an AI literacy pilot at another adult education provider, signalling that this is moving from experiment to funded practice.
It also arrives with a set of risks that are specific to this population and that generic education technology guidance does not address. This article covers where AI genuinely extends what a small program can do, where it undermines the learning it appears to support, the access questions that determine whether any of it reaches the learners who need it most, and what to put in place before handing tools to volunteer tutors.
The Tutor Preparation Problem
The highest-value application in most adult literacy programs is not learner-facing at all. It is lesson preparation, and it addresses a structural weakness that programs rarely discuss openly: volunteer tutors are usually well intentioned, reasonably educated, briefly trained, and largely unprepared for the specific work of teaching an adult to read.
A typical tutor training runs a day or two, covers the fundamentals of adult learning theory and the program's assessment framework, and then hands the volunteer a curriculum binder and a learner. What happens next depends almost entirely on how much time the tutor is willing to invest in preparation. Some prepare thoroughly. Most, being volunteers with their own jobs, arrive having skimmed the next unit in the car. The gap between the two is the largest single source of variation in learner outcomes, and no amount of curriculum investment closes it, because the binding constraint is unpaid preparation time.
This is where AI produces the most reliable value. A tutor can describe their learner in a sentence, state the goal for the session, and get a usable plan with a warm-up, a target skill, practice items, and an exit check in under five minutes. More importantly, they can get materials at the right level, which is the thing adult literacy programs never have enough of. Published adult materials are frequently either too childish in content or too advanced in language, and the perennial complaint of adult educators is the shortage of texts that treat an adult as an adult while using controlled vocabulary.
Generating a short passage about renting an apartment, reading a pay stub, or talking to a child's teacher, written at a specified reading level and about a topic the learner actually cares about, is a task these tools do well and that programs previously solved by photocopying whatever was to hand. The same applies to producing three versions of a worksheet at different levels so a mixed-ability class can work on the same content, which is a routine need that most volunteer tutors have no practical way to meet.
The caveat is that generated material needs a review step. Reading level claims from a model are approximate rather than measured, cultural assumptions creep into scenarios in ways that matter for immigrant learners, and factual content about benefits, immigration procedures, or employment rights can be wrong in dangerous ways. A program coordinator reviewing generated materials before they reach a classroom is a small ongoing cost that prevents a significant category of harm.
Preparation Tasks That Work Well
Behind the scenes, before the learner arrives
- Leveled reading passages on adult-relevant topics
- Three versions of one worksheet for a mixed-ability group
- Session plans built around a learner's stated goal
- Simplifying a real document the learner brought in
- Comprehension questions and quick exit checks
What the Review Step Catches
Why generated material should not go straight to a class
- Reading level that is claimed rather than measured
- Cultural assumptions that do not fit immigrant learners
- Wrong information about benefits, immigration, or employment
- Idioms and register that confuse rather than teach
- Content that talks down to an adult reader
Conversation Practice and the Thing Programs Cannot Supply
For English language learners specifically, the scarcest resource in any program is speaking time with a patient interlocutor. A weekly two-hour class divided among twelve learners produces a few minutes each of actual speaking, and the learners who most need practice are usually the ones least willing to take it in front of peers. Conversation groups help, but they depend on volunteers being available at times that work for people who finish shifts at seven.
Voice-capable AI assistants change the supply of this specific thing more than any previous technology. A learner can practise ordering at a pharmacy, calling a landlord about a repair, or answering interview questions, repeatedly, at eleven at night, with something that does not sigh, does not correct them in front of anybody, and does not become visibly impatient on the fourth attempt. The absence of social risk is not a minor feature. For adults who are embarrassed about their English, it is frequently the whole barrier.
What this does not do is teach. Practice consolidates what has been learned and builds fluency and confidence, and it works when it is aimed at something specific the learner is working on. Left unstructured, it becomes conversation for its own sake, which feels productive and produces little. Programs that get value from this frame it explicitly as homework attached to a lesson: this week you practised making appointments, here is how to practise it ten more times before we meet again.
Accent and dialect handling deserves scrutiny before recommending any particular tool. Speech recognition performs unevenly across accents, and a learner whose speech is repeatedly not recognized experiences that as personal failure rather than as a limitation of the software. This is a real risk of discouragement, and it falls hardest on exactly the learners who are earliest in their language learning. Any tool a program recommends should be tested by staff with learners representing the language backgrounds the program actually serves, not adopted on the strength of a demonstration.
There is a related question about which English is being taught. Research on digital language learning has repeatedly noted that automated systems tend to treat standardized varieties as correct and to flag regional or culturally specific expressions as errors. In an adult program serving a particular community, that bias can quietly work against the way learners actually need to speak in their own neighbourhoods and workplaces. Related tensions in translation and language support are covered in the guide to AI translation quality review for nonprofits.
The Access Question That Determines Everything Else
Every application discussed so far assumes the learner has a device, a data connection, and enough digital confidence to use an unfamiliar application without support. For a meaningful share of adult literacy learners, at least one of those assumptions fails, and the ones for whom all three fail are usually the learners furthest from where the program wants them to be.
Research on educational technology in language learning has been consistent on this point: benefits accrue unevenly, and platforms that depend on reliable connectivity and digital literacy can widen the gap they were introduced to close. The mechanism is straightforward. Learners with a smartphone, home internet, and prior experience of apps get an additional hour of practice each evening. Learners without get the same instruction they had before, while the program's reference point for normal progress shifts upward around them.
This is not an argument against adoption. It is an argument for measuring the gap deliberately rather than discovering it later. A short intake conversation covering what device the learner has, whether they have data at home, and how comfortable they feel installing an application gives a program the information it needs to avoid designing around an assumption. Many programs find the picture better than feared on devices and worse than feared on confidence, which points toward different interventions than a device lending scheme alone.
The practical response has three parts. First, keep the core program complete without any AI component, so that a learner who cannot or will not use the tools is not receiving a diminished version of the service. Second, treat digital skills as content rather than as a prerequisite, since being able to use a phone for something other than messaging is itself a life outcome many adult learners want. Third, where a tool genuinely improves outcomes, budget for the access rather than assuming it, whether that means devices, data, or supervised time on program computers. The broader framing is covered in the discussions of the digital divide among the people nonprofits serve and practical approaches to closing it.
Questions Worth Adding to Intake
Five minutes that prevent designing around a false assumption
- What device do you use most, and is it shared with anyone?
- Do you have internet at home, or do you rely on mobile data?
- Have you installed an app yourself before, and how did it go?
- Is there a quiet place where you could practise speaking aloud?
- Would you rather practise alone or with someone present?
Where AI Substitutes for Learning Rather Than Supporting It
There is a category of application that looks helpful and is corrosive, and adult literacy is unusually exposed to it because the same tools that teach reading and writing can also perform reading and writing on the learner's behalf.
A learner who needs to write a letter to their landlord can be taught to write it, or can be handed a letter. A learner who cannot read a benefits notice can be taught to decode it, or can have it summarized aloud. In both cases the immediate problem is solved faster by substitution, and in both cases the learner is no more capable than before. For a program whose entire purpose is capability rather than task completion, this distinction is the central pedagogical question the technology raises.
It is not a simple line, because substitution is sometimes the right answer. An adult facing an eviction notice this week needs the notice understood now, and a program that insists on teaching rather than helping has confused its method with its mission. The honest framing is that both are legitimate services and they are different services. Reading a document to a learner in crisis is advocacy and it is valuable. It is not instruction, and counting it as instruction produces attendance figures that do not correspond to any learning.
The practical guidance that works for volunteer tutors is a sequencing rule rather than a prohibition. The learner attempts first, the tool assists second, and the tutor's job is to make sure the attempt happens. A learner drafts the letter, then uses AI to check it and see alternatives, then discusses with the tutor why a particular change was suggested. That sequence uses the tool to accelerate feedback, which is genuinely scarce, rather than to bypass the productive struggle that constitutes the learning.
This has implications for how programs assess progress. If learners have access to tools that can write for them, then writing samples produced outside supervised sessions stop being evidence of ability. Programs that rely on portfolio evidence for funder reporting need to think this through before it becomes a discrepancy someone else notices, and the broader question of measuring learning honestly is explored in the piece on measuring student outcomes beyond test scores.
Supports the Learning
The learner does the work, the tool speeds the feedback
- Learner drafts first, then asks for corrections and reasons
- Repeated speaking practice on a skill taught in session
- Extra practice items at the level the learner is working on
- Explaining a word or rule the learner asked about
Substitutes for the Learning
Legitimate as help, not countable as instruction
- Producing a finished letter the learner did not attempt
- Summarizing a document instead of decoding it together
- Translating everything so English is never required
- Unsupervised written work used as evidence of progress
Privacy and the Particular Vulnerability of Adult Learners
Adult literacy and ESL programs hold information that is more sensitive than their size suggests. Enrolment records may indicate country of origin, length of time in the country, employment situation, family composition, and in some cases immigration status. Learners frequently disclose more in the course of instruction, because the material is personal and because a tutor relationship built over months invites confidence.
This makes casual use of consumer AI accounts by volunteer tutors a genuine concern rather than a theoretical one. A tutor pasting a learner's writing sample into a chatbot for feedback has transmitted whatever that sample contains, and adult learner writing is frequently about the learner's own life. A tutor asking for help drafting a reference for a learner has transmitted the learner's circumstances. Neither tutor did anything they would recognize as a privacy breach.
The rule that works is to teach de-identification as a habit rather than to rely on judgment about sensitivity. Remove names, replace specific places with generic ones, strip dates and case numbers, and never enter anything relating to immigration status, benefits eligibility, health, or a legal matter. Stated that plainly, most volunteer tutors comply, because the instruction is concrete rather than a principle they have to interpret.
Programs receiving federal or state adult education funding should also check whether their grant conditions or applicable student records rules constrain what may be shared with a third-party service. The specifics vary by funding stream and by whether the program operates within an educational institution, and the answer changes what tooling is permissible rather than merely advisable. The related regulatory territory is covered in the guide to student privacy and FERPA in AI environments, and the general policy framing in AI policies for small nonprofits.
There is a consent dimension that deserves more attention than it usually gets. Explaining data practices to someone whose English is at an early level, in a way that constitutes meaningful rather than nominal consent, is genuinely difficult and cannot be solved with a translated form. Programs that take this seriously tend to explain it verbally in the learner's first language, keep the explanation extremely short, and give a clear option to decline that carries no consequence for the service received.
Preparing Volunteer Tutors
A volunteer tutor corps is not a staff team, and training that works for employees does not transfer. Tutors have limited time, they joined to teach rather than to learn software, and any requirement that feels like an administrative burden will reduce the number of people willing to tutor. That constraint should shape the training rather than be treated as an obstacle to it.
The most effective approach observed in practice is to embed the tool in existing tutor training rather than to run a separate session about AI. Twenty minutes inside the standard onboarding, framed as how to prepare a session, is attended by everyone. A separate optional workshop on artificial intelligence is attended by the tutors who least need it. The framing matters as much as the content, because the tutors who would benefit most are frequently the ones least likely to opt into anything technological.
What that twenty minutes should cover is narrow. One worked example of generating a leveled passage on a topic a learner cares about. One worked example of adapting a real document the learner brought in. The de-identification rule, stated as a short list rather than a principle. The instruction that generated material is checked by a coordinator before use with a learner. And an explicit statement of where the program does not want AI used, which prevents well-meaning improvisation more effectively than any amount of encouragement shapes it.
Ongoing support matters more than initial training, and the cheapest effective version is a shared library of prompts and materials that worked. When a tutor produces a good passage on tenant rights at a low reading level, that passage should be available to every other tutor rather than being regenerated forty times. Programs that build this accumulate a genuinely valuable resource within a year, and the approach mirrors the broader practice described in AI for nonprofit knowledge management. The general mechanics of preparing volunteers are covered in the guide to streamlining volunteer onboarding and training.
Twenty Minutes Inside Standard Tutor Training
Narrow, concrete, and attended by everyone
One example of generating material
A leveled passage on a topic a real learner cares about, produced live, with the reading level checked rather than assumed.
One example of adapting a real document
A utility bill or school letter simplified into a teachable text. This is the use tutors will reach for most often.
The de-identification list
Names, places, dates, case numbers, immigration status, benefits, health, legal matters. Stated as items, not as a principle to interpret.
Where the program says no
An explicit boundary prevents improvisation more reliably than encouragement produces good practice.
Choosing Between Purpose-Built and General Tools
Programs face a choice between general assistants and platforms built specifically for adult education, and the answer depends on scale and on how much the program values alignment with the assessment frameworks it reports against.
Purpose-built adult education platforms have appeared that align content to the standards and assessments adult programs actually use, including the competency frameworks and test preparation that determine funding-relevant outcomes. That alignment is the genuine advantage, because a general assistant knows nothing about the specific assessment your state uses and cannot target instruction to it without a great deal of prompting. Some of these platforms publish learning gain comparisons for their tutoring features, though as with all vendor-reported outcome data these are worth reading with attention to how the comparison was constructed.
The disadvantages are cost, lock-in, and the implementation burden that small programs consistently underestimate. A platform that requires learner accounts, staff administration, and data migration is a project, and adult literacy programs run by two coordinators and forty volunteers have limited capacity for projects. A general assistant used well by tutors for preparation captures a substantial share of the available value at no licensing cost and with no implementation at all.
The sequence that makes sense for most programs is to start with tutor-facing use of general tools, because it is free, requires no learner accounts, and surfaces no privacy questions beyond de-identification. Only once the program can describe specifically what a purpose-built platform would add, and can name the person who will administer it, does the paid option become a reasonable proposition. Buying first and finding the use case afterward is the pattern that leaves small organizations with a subscription they cannot justify at renewal, and the general evaluation discipline is set out in the guide to AI in tutoring programs.
What Adult Learners Are Actually Asking For
There is a final consideration that is easy to miss when evaluating tools, which is that adult literacy learners increasingly have their own reasons for wanting to engage with this technology, and those reasons are not always the ones the program has in mind.
Adults enrolled in literacy and language programs are frequently in workplaces where AI tools are being introduced, applying for jobs through systems that screen applications automatically, and navigating public services that increasingly present a chatbot as the first point of contact. For these learners, being unable to use these tools is becoming a functional literacy problem in its own right, in the same way that being unable to fill in an online form became one two decades ago.
That suggests something beyond using AI to teach reading, which is teaching AI as part of what literacy now means. A short unit on what these tools are, how to ask them something usefully, how to recognize that an answer might be wrong, and what should never be typed into one is defensible content for an adult education program, and several funders have begun supporting exactly this kind of pilot. It also has the useful property of being immediately applicable to learners' working lives rather than being deferred benefit.
Programs that go this route should be careful about one thing. Teaching a population with limited literacy to rely on a tool that produces confident, fluent, sometimes wrong output carries an obvious risk, because the ability to detect that an answer is wrong depends on exactly the skills the learner is still building. Any such unit needs to place verification at the centre rather than treating it as a footnote, and to be honest that the tool is frequently wrong in ways that sound completely convincing. Related thinking on extending capability to the communities nonprofits serve appears in the discussion of reaching communities on the wrong side of the connectivity gap.
Conclusion
Adult literacy and ESL programs have a stronger case for AI than most nonprofit program areas, because the constraints they face are unusually well matched to what the technology supplies. Leveled material at scale, unlimited patient speaking practice, and lesson preparation for volunteers who have no time to prepare are genuine shortages, and they are shortages no amount of additional curriculum funding has resolved.
The value is most reliable behind the scenes. A volunteer tutor who arrives with a plan and a passage at the right level about something the learner cares about is a better tutor, and getting there costs five minutes rather than an hour. That single change, applied across a tutor corps, probably outweighs every learner-facing application in aggregate effect, and it requires no learner device, no account, and no consent conversation.
The risks are equally specific. Learners without devices or confidence fall further behind if the program assumes access it has not checked. Speech tools that fail on a learner's accent discourage the people who need practice most. Tools that write for a learner produce documents rather than capability. And a population that discloses sensitive circumstances in the course of learning deserves more careful handling than a volunteer with a consumer chatbot account can be assumed to provide.
Start with tutor preparation, add the de-identification rule to your standard onboarding, ask three access questions at intake, and keep the core program complete for learners who use none of it. That is a modest programme of work, it costs almost nothing, and it captures most of what is genuinely available while avoiding the failures that make this technology a net loss for the learners with the least.
Bring AI Into Your Program Without Losing the Teaching
We help education and literacy nonprofits work out which applications strengthen instruction, which quietly replace it, and how to prepare volunteer tutors for both.