AI in Adult and Worker Education

UFT-CWE workshops belong to the chapter. The chapter builds them. The funding we're using comes from a contract we negotiated — NYS taxpayer EPE dollars allocated to us as a chapter, unencumbered as to subject matter so long as the work serves the work. That clause didn't write itself. Years of bargaining made it possible to gather a room of adult educators around a question of our own choosing, on time we are paid for, with materials we hand each other rather than receive from above. I want to name that out loud at the start. The day we are planning is not a gift. It is what we secured and deployed. In the upcoming workshop on artificial intelligence (AI) for which this reader is an introduction, there is an analogous theme and a persistent question: even at this very time when national and global politics are so disturbing, and collective action seems impossible, AI and automation are bringing the potential for radical abundance in the world. Virtually limitless availability of energy, food, shelter, healthcare, clean drinking water at costs declining and eventually approaching zero is no longer fantasy. It is within practical reach; but for whom? Whose abundance? If it has been built with our data and the accumulated knowledge, art, and wisdom of millennia, is it not a natural right of all the people everywhere? Is abundance to be locked up in the hands of a frontier company, or enclosed by a superpower? Or is it to be shared by all the people? And if to be shared, by students and teachers, parents and children alike, by all communities everywhere, then how? How might we advocate for entitlement and agency, just as this Chapter has always done?

This question we've chosen is how to envision artificial intelligence for the people — not as gadget, not as curriculum supplement, not as another thing to learn before lunch, but as the technology now restructuring the labor we and our students perform. Some of what's coming has already arrived. Many of you have students who use ChatGPT or Claude or Gemini for homework, for cover letters, for translations, for the hundred small literacies of life in this city. Some of you use these tools yourselves. Others have stayed away by choice or circumstance. The chapter is a microcosm.

What I want this introductory session to do is modest and serious at once. Modest, because we will not in three or six hours become experts on a field that even its insiders cannot keep up with. Serious, because adult educators in a union setting are exactly the people who should be thinking together about what these tools mean for our students, our classrooms, our wages, and the wider economy our students enter when they leave us. We are not going to start from somebody else's framing.

The premise of the day, said plainly:

The packet you have in hand contains short pieces on the threads we couldn't fully cover in the time we have. A primer on the frontier companies and what distinguishes them. A piece on AI and labor that names the question without dressing it up. A practical guide to prompting these tools for educational work. A piece on assessment now that students have access to the same tools we do. A short reading on bias and language — particularly relevant for those of us teaching ESL, where what counts as "correct" English has always been a contested matter. A piece on privacy and data for students whose lives include reasons to keep certain details out of corporate databases. Each one is meant to be readable in twenty minutes or so. Each one is meant to seed a conversation, not close it.

What this day will not do is hand you a method. The room contains too much variety of subject, level, population, and pedagogy for one method to fit. What I hope it will do is put you in working contact with the technology, give you a frame for thinking about it as workers and as teachers, and offer enough material that the conversations we begin today can continue in the weeks and months ahead.

Which brings me to the second purpose of the day. This session is the opening of a chapter study group, or several. I would like the people in this room to choose what they want to keep working on — practical classroom applications, the labor and ownership questions, the technical literacy of prompting, the assessment problem, whatever calls them — and to take that work into smaller groups that meet across the spring. The chapter has the funding to support that. We have the people. We have the politics. What we need is to use what we have.

The written materials for our upcoming workshop draw on work I've been doing on a project called The People's Share, which takes up the question of democratic ownership of AI infrastructure in plain terms — written in its aspirations for everyone. The pieces in your packet stand on their own. I name the source so you know where the through-line comes from, and so that those who want to read further know where to look.

A note on Claude. You will see Claude — the AI assistant — show up in this work. I have used Claude as a writing and thinking partner in producing much of what you are reading. We will spend some time today looking at what that collaboration actually looks like, in real time, with the screen shared. I want you to see the seams. The point is not to demonstrate that AI is impressive. The point is to demonstrate what working with it looks like when a worker is doing the working.

That is the day. Greetings, with a big smile to those of you I've known, and advocated, and taught alongside for years. Welcome, to those I'm meeting for the first time. Let's get to it.

There is a story being told about artificial intelligence, and it goes like this. A new technology has arrived. It is powerful. It will change everything. Some jobs will disappear. New ones will emerge. Workers will need to adapt. Those who learn to use the tool will thrive. Those who don't will fall behind. The technology itself is neutral; what matters is how we use it.

This story is wrong in a specific way, and the specific way matters.

It is not wrong that the technology is powerful, or that jobs will change, or that learning to use these tools is worth doing. All of that is true. The story is wrong in what it leaves out, which is the question of ownership. Who built this technology. Who paid for it. Who owns the infrastructure that runs it. Who collects the revenue when it is used. Who decides what it can and cannot do. Who is allowed to shape its direction, and who is offered only the choice to use it or fall behind.

These are not secondary questions. They are the questions. And they have answers. The major AI systems in the world right now were built by a handful of corporations, financed by another handful of investors, running on infrastructure owned by still another handful, and trained on the collective written and visual output of essentially everyone who has ever posted anything online — which is to say, on work the corporations did not pay for and did not ask permission to use. The economic arrangement is one of the most concentrated in modern history. A small number of firms are positioning themselves to collect rents from a tool built on the unpaid contributions of billions of people, and the public conversation about this tool is largely confined to whether individual workers will adapt fast enough to keep their jobs.

That is the frame we are being handed. This piece is an argument that we should not accept it.

Adult and worker education has always operated on a particular wager — that workers, given the time and the room and the materials, will produce their own analysis of the conditions they live under. The CWE network exists on this wager. Forty-plus training providers, thirty-six union locals, thirty-four thousand learners last year — that is not a service-delivery operation. It is a standing infrastructure for working-class education, built and sustained by the labor movement of this city. The teachers in this room work inside it.

Consider what your students bring to the GED, ESL, computer skills, college prep, etc., classroom. They bring labor history — their own, often unrecognized. They come from 1199, TWU, Workers United, IBEW, UAW, the Actors' Guild locals, the building trades; from NYCC and the other community-based organizations in the network; from jobs that were once unionized and are no longer, and from jobs whose unions are still fighting to hold the line. Many work in occupations that have been restructured by previous waves of technology in ways that benefited owners and not workers. Home health aides whose schedules are now algorithmically determined. Warehouse workers whose pace is set by software. Drivers whose routes are optimized for the platform's profit and not their fuel costs. Office workers whose keystrokes are logged. The technology arrived. The workers were told they would need to adapt. They adapted. The owners kept the gains.

AI is the next iteration of this pattern, with one important difference. The previous waves of technology took particular skills — driving, sorting, scheduling — and routed them through software. AI takes the general capacity for language, reasoning, and creative work and routes it through software. The category of labor it touches is much broader. Teachers, writers, translators, paralegals, illustrators, programmers, customer service workers, tutors, researchers — work that was assumed to be safe because it required judgment and language and care. Some of this work will disappear. More of it will be restructured: the worker remains, but the worker now operates a system that does much of what the worker used to do, and the wage adjusts accordingly. We have seen this before. We know how it ends if nobody intervenes.

The intervention is not "learn to use the tool." That is necessary, but it is not sufficient and it is not even the main thing. The intervention is ownership. The question is not whether your students will use AI. They already do. The question is whether the value generated by that use accrues to them, to the corporations that built the tool, or to some structure we have not yet built that distributes it differently.

There are precedents for what such a structure might look like. Worker cooperatives have run productive enterprises for over a century. Mondragón, in the Basque Country, employs over eighty thousand people across industrial, financial, and educational cooperatives, owned and governed by the workers themselves. Cooperative Home Care Associates, in the Bronx, is the largest worker cooperative in the United States, owned by its home health aides. The Park Slope Food Coop, in Brooklyn, has run for half a century on member labor and democratic governance. The Burlington Community Land Trust takes housing out of the speculative market and holds it in common. These are not utopias. They are working institutions, with disagreements and difficulties, that demonstrate something simple: economic activity can be organized so that the people doing the work also own and govern the enterprise.

The question this piece poses is whether the same logic can be applied to artificial intelligence. The infrastructure is currently held by a small number of private corporations. The training data was produced by all of us. The labor of using and improving these tools is increasingly distributed across the working population. There is no economic law that says this configuration is permanent. There is, however, a political project to make it seem permanent, to make any other arrangement seem unrealistic, and to confine our conversation to how individual workers can survive within the configuration as it stands.

Adult/worker educators are uniquely positioned in this conversation. Our students are the workers most exposed to displacement and most underserved by the public conversation about AI. They are also, in the long tradition of adult and worker education — from Highlander to Freire to the labor colleges and union halls of this city — the people whose political education has historically generated the analysis that changes things. The classroom is not separate from this question. It is one of the rooms where it gets answered.

What we are doing in this workshop is not, finally, about prompting techniques or curriculum integration, important as those are. It is about whether teachers and workers in adult education develop a position on AI that is theirs and not borrowed from a corporate communications department. It is about whether we read the rules ourselves.

The other pieces in this packet take up specific aspects of this — what the major companies are, how to use the tools well, how to think about assessment and bias and privacy. Take what is useful. The frame is what matters most. The frame is this: this is a labor question, and labor has answers when it organizes itself around them.

If you have heard people talk about "AI" as one thing, set that aside. There are several major systems, built by different companies with different histories and different stated purposes, and the differences among them matter — both for what they can do and for what they tell you about the industry building them. This piece is a short orientation to the field as it stands. Treat it as a map drawn at a moment in a fast-moving landscape; some details will shift between this writing and the workshop, but the structure will hold.

The major labs.

OpenAI is the company that put generative AI into public consciousness with the release of ChatGPT in late 2022. It was founded in 2015 as a nonprofit research organization with a stated mission of ensuring that artificial general intelligence benefits all of humanity. That mission and that legal structure have not survived contact with the technology's commercial value. OpenAI is now a capped-profit company in close partnership with Microsoft, which has invested tens of billions of dollars and embeds OpenAI's models throughout its product lines. The conversion from nonprofit to commercial entity has been ongoing and contested, with departures of senior staff and public disagreements about whether the original mission has been honored, abandoned, or restructured into something compatible with private capital at scale. OpenAI's flagship products are the GPT series of models, accessible through ChatGPT and through an API used by other companies to build their own AI products. The company is the largest by user base and has the deepest cultural penetration. Whether the original mission means anything in practice now is one of the live questions of the field.

Anthropic is the company that makes Claude, the assistant I prefer, one that, through the workshop activities, I think you'll find distinctive. Anthropic was founded in 2021 by former OpenAI staff who left over disagreements about safety and direction. Its stated focus is on building AI systems that are safe, steerable, and interpretable — meaning systems whose internal workings can be understood and whose behavior can be reliably shaped. Anthropic's models are widely regarded as strong on writing, reasoning, and long-form work. The company is smaller than OpenAI by user base but well-resourced, with major investment from Google and Amazon — meaning that the two largest cloud infrastructure providers are now financially committed to two different frontier labs. Although I am the author of this piece, it is fair to say that Claude is a working partner. As odd as that may sound, it is a practical actuality. These models have become phenomenally capable; that development will only continue, and we are collaborating with "tools" that might convince you that there's a ghost in the machine. There is not; and yet, something is evolving and emerging that we may not have words or categories for at this moment in history.

Google DeepMind is Google's AI research division, formed by merging Google Brain with the British research lab DeepMind that Google acquired in 2014. Its public-facing model is Gemini, integrated across Google's product ecosystem — Search, Workspace, Android. Google has decades of research depth in AI, enormous computing infrastructure, and the largest distribution channel of any of the major labs. Gemini is competitive at the frontier and is the model most adult learners will encounter without seeking it out, because it shows up inside tools they already use.

Meta (Facebook's parent company) released the Llama series of models from 2023 through 2025 with open weights — meaning the underlying model could be downloaded, modified, and run by anyone. For a stretch of years this made Llama the foundation of the open-source AI ecosystem and gave researchers, smaller developers, and cooperative projects access to frontier-grade infrastructure they could not otherwise have built. In April 2026, Meta retired the Llama line and replaced it with Muse Spark, a proprietary model accessible only through Meta's API and hosted services. The shift from open weights to closed commercial release was framed by the company as a competitive necessity — Chinese labs had absorbed the Llama architecture and were producing competitive models at lower cost — but the practical effect is that one of the major routes for non-corporate use of frontier AI has narrowed considerably. The open-source ecosystem now leans on the Chinese labs, on Mistral in France, and on smaller efforts whose resources do not match the frontier. The Llama-to-Muse-Spark transition is itself part of the story this packet is telling. When an asset becomes valuable enough, the owners enclose it.

xAI is Elon Musk's AI company, which makes Grok, the model integrated into the platform formerly known as Twitter. It is positioned as a less filtered alternative to the other major labs, and it is positioned that way deliberately, because its owner has reasons for wanting an AI assistant that produces fewer of the kinds of outputs the other labs constrain. The model itself has improved considerably since its launch and is now competitive on standard benchmarks. The disposition of Grok — what it will say, what it praises, what it dismisses — reflects choices made by its proprietor. This is true of every model on this list to some degree, but it is more visibly true here, and the visibility is itself the point. Worth knowing about because students and family members will encounter it, particularly through the X platform.

Mistral, based in Paris, is the most credible Western lab still committed to releasing open-weights models. It is much smaller than the American labs and operates at a different scale, but it has produced strong models and represents — for now — a European counterweight to the U.S.–China bipolar picture, with European regulatory support and a stated commitment to open release. Whether this position holds as commercial pressure intensifies is one of the things to watch.

The Chinese labs. DeepSeek, Alibaba's Qwen, Baidu's Ernie, Moonshot's Kimi, Zhipu, and others. These are serious players, in some cases producing models that match or exceed Western frontier performance on specific benchmarks at lower training cost. DeepSeek's release of a frontier-competitive model with open weights at low cost in early 2025 was one of the year's significant events and accelerated Meta's decision to abandon the open-weights strategy a year later. The geopolitical dimension of the field — U.S. export controls on advanced chips, Chinese state investment, the question of whether AI development becomes another arena of great-power competition — runs through this part of the picture. The competition between U.S. and Chinese labs is increasingly the structural fact shaping decisions made by every other actor on this list.

What distinguishes them, practically.

Across the major models, basic capability is converging. All of the leading systems can write reasonably well, reason through most everyday problems, summarize, translate, code, and converse. The differences that matter for working teachers are subtler:

Writing texture. Different models have different prose tendencies. Claude tends toward measured, paragraph-form responses; ChatGPT tends toward heavier formatting, bullet points and headers; Gemini varies by version. None of these is right or wrong — they reflect design choices. For producing teaching materials, the model whose default texture is closer to what you want will require less correction.

Reasoning depth. The newer models from each lab can be set to "think longer" before answering, which produces better results on hard problems at the cost of speed. Useful for math walk-throughs, multi-step explanations, careful analysis of a passage. Worth knowing about even if you don't use it constantly.

Context window. How much text a model can hold in its working memory at once. The leading models now handle hundreds of pages of text in a single conversation. This matters when you want to upload a long article, a textbook chapter, or a set of student writing samples and have the model work across all of it.

Refusals and disposition. The models differ in what they will and won't do. Some are more cautious; some are less. For most adult education work this is not a concern, but it surfaces in unexpected places — discussing certain political topics, generating realistic role-play scenarios, working with sensitive content drawn from students' lives.

Voice and multimodal. Most major models now accept images and produce voice output. The quality varies. For ESL pronunciation practice and listening comprehension, the voice models are increasingly useful.

What this picture leaves out.

Five corporations and a handful of national research efforts are not the entire AI ecosystem. There are smaller labs, open-source projects built on Llama and other open-weight models, cooperative and academic research efforts, and a growing world of specialized models built for particular tasks — medical, legal, scientific. The picture above is the commercial frontier. The wider field is more diverse, and some of the most interesting work in it is being done outside the major labs.

The point of the orientation is not to pick a favorite. It is to be able to read the conversation. When someone says "I tried AI and it couldn't do X," the question is which AI, in what configuration, asked how. When a school district adopts a particular tool, the question is which tool and on what terms, with whose data, paid by whom. When a news story discusses "AI" as one thing, the question is which thing and which company and which interest, and why this story is being told this way at this moment.

Five corporations and a handful of national efforts will not remain the entire picture forever, and the configuration of the field is more contested and more politically determined than the marketing suggests. The literacy this packet is trying to build begins with knowing the field has actors, not just artifacts — and that the actors have addresses, board members, financial backers, and political commitments. None of them is a neutral utility. None of them was built to serve the interests of adult learners in this city. That does not mean the tools are useless. It means the tools require literate users — which is where you come in.

The first thing to understand about prompting an AI assistant is that it is not a search query and it is not a command line. It is closer to a conversation with a very fast, very widely read, somewhat literal-minded colleague who has no memory of yesterday and will do exactly what you ask, including the parts you didn't mean. Most of what people call "prompt engineering" is just learning to give good instructions to such a colleague. You already know how to do this. You do it every time you explain a task to a substitute teacher. The skill transfers.

Here is the underlying principle. These systems generate responses by predicting what would plausibly come next given everything you have written. What you write determines what you get. Vague input produces vague output. Specific input produces specific output. Context matters more than cleverness. There is no secret syntax. There are no magic words. There is only the quality of the instructions you give.

A practical translation of this for classroom work:

Tell it who it is talking to and what for. "Write a paragraph about photosynthesis" produces a generic paragraph. "Write a paragraph about photosynthesis for an adult ESL learner at an intermediate level, using simple vocabulary, present-tense sentences, and a concrete example from a kitchen garden" produces something you can actually use. The second prompt is not more clever than the first. It is just more honest about what you actually need.

Give it your actual constraints. If a passage needs to be at a sixth-grade reading level, say so. If it needs to fit on one page, say so. If it needs to avoid specific vocabulary your students haven't learned yet, list the vocabulary. If it needs to support a particular learning objective, name the objective. The system will not infer your constraints. It will produce something competent and generic unless you tell it the specifics that make your context yours.

Show it what good looks like. If you have a sample passage that works for your students — something at the right level, in the right tone, with the right structure — paste it in and say "produce something like this on the topic of X." This is more powerful than any list of adjectives. The example carries information that words about the example cannot.

Iterate. Do not expect the first answer to be the answer. Real prompting is a back-and-forth. The first output gives you something to react to. Tell the system what is wrong with it. "This is too abstract — give me three concrete examples instead." "This is too long — cut it in half." "The vocabulary in the second paragraph is above my students' level — simplify." Each round narrows toward what you need. Treat the conversation as a draft process, not a vending machine.

Ask it to think before it answers. For tasks that involve reasoning — math problems, multi-step explanations, analyzing a passage — ask the system to walk through its thinking before giving the final answer. "Before you answer, list the steps you would take." This produces better answers and also lets you see where the reasoning went wrong if it goes wrong.

Push back when it's wrong. These systems make mistakes. They make confident mistakes. They also accept correction gracefully. If something looks off, say so: "The date you gave is wrong — check it." "That definition contradicts what you said earlier — which is right?" "I don't think this example works because…" The system will reconsider and often correct itself. Treating it as a colleague you can argue with produces better work than treating it as an oracle you must accept.

Watch for hallucination. The systems sometimes invent facts, sources, statistics, and quotations that sound entirely plausible. They do this most often when asked for specifics they do not actually have — citations to particular studies, biographical details about lesser-known figures, exact numbers. Verify anything that matters. The system will produce a confidently wrong citation as readily as a correct one, and it will not flag the difference.

Use it for the parts of the work that are mechanical, not the parts that are yours. Drafting a worksheet from a passage you have chosen, generating practice questions at varying difficulty levels, producing five different versions of the same problem with different numbers, rewriting a text at three different reading levels, translating a parent-teacher communication into Spanish or Mandarin or Bengali — these are tasks the system can do quickly and well, and they are tasks that consume teacher time disproportionate to their value. Use it for those. The judgment about what to teach, how to teach it, what your particular students need, how a particular learner is doing — that is yours, and the system cannot do it. It does not know your students. You do.

A few specific applications worth practicing in the workshop:

For ESL: producing simplified versions of authentic texts; generating dialogue practice scenarios at specified levels; creating vocabulary exercises around themes drawn from students' actual lives and work; producing parallel passages in students' first languages for comparative reading.

For GED: generating practice questions in any of the four subject areas, calibrated to the actual GED format; producing multiple worked examples of math problems showing different solution approaches; drafting mini-lectures on social studies topics; producing extended response prompts with the kind of paired stimulus passages the test uses.

For computer skills and college prep: drafting cover letters, college essays, and professional emails as starting points students can revise; explaining unfamiliar interfaces and software workflows; producing practice scenarios for job interviews; walking through the steps of a financial aid application or a course registration system.

For your own work: drafting lesson plans, generating differentiated versions of activities, producing rubrics, summarizing articles you don't have time to read, drafting communications to administrators or to parents, generating ideas when you are stuck. The system is a useful collaborator on the parts of teaching that are not the teaching itself.

A note on what not to do. Do not paste student work containing identifying information into a system whose data practices you have not reviewed. Do not use these tools to write things that should be written by you — a personal note to a student in distress, a memo to a supervisor, anything where the writing is the relationship rather than a task to complete. Do not present AI-generated material to students as your own work, especially if you are also asking them not to do the same. The honest version is more useful pedagogically anyway: "I asked the AI to draft this, then I revised it; here is what I changed and why." That models what you want from them.

The point of prompting practice is not to become a virtuoso of the prompt. It is to become someone who can put these tools to work on the parts of your job that benefit from them, while keeping the parts that should remain yours firmly in your own hands. The teachers who do this well are not the ones who learn the cleverest tricks. They are the ones who think clearly about what they need and then ask for it.

In the workshop we will do this together, with real tasks, on the screen, in real time. Bring something you are working on.

This is the piece about money, and about the small everyday questions that turn out to be the political ones. How much does it cost to use these tools. Who pays. What you get for free and what you don't. What "free" actually means when the company is one of the most valuable in the world. Whether AI access belongs in a contract negotiation. Whether it already does, whether the locals know it, whether members are paying out of pocket for tools their employers benefit from.

We will start with the practical mechanics and move to the questions underneath them.

The basic structure. The major AI companies all offer their products on roughly the same model. There is a free tier that gives you limited access to the better models, a consumer subscription in the range of twenty dollars a month that gives you fuller access and higher usage limits, a higher consumer tier (around two hundred dollars a month at the time of this writing) for heavy users, and a set of business and enterprise plans for organizations. The free tier is genuinely useful and many people will not need to pay. The twenty-dollar tier is where most individual professional users settle, and it is meaningfully more capable for sustained work. The two-hundred-dollar tier removes most usage friction and is calibrated for people doing heavy daily work — researchers, writers, programmers, teachers preparing material at scale.

Usage limits. Even on paid plans, there are limits. The systems meter how much you can do in a given window — typically a rolling five-hour or daily window, sometimes weekly — and when you hit the limit you wait. The limits are real and they bite for the kind of sustained work that produces a packet like this one. Heavy users on the lower paid tier regularly hit walls. The companies adjust these limits without notice, sometimes tightening them, sometimes loosening, with the announced reasons rarely matching the pattern users observe. Worth knowing because it shapes what you can plan for. If you are going to build curriculum on the assumption that AI is available when you sit down to work, you want to know the tool will be there.

Context windows. The "context window" is how much text the system can hold in its working memory during a single conversation. A conversation that exceeds the window starts losing the early material — the system forgets the beginning of what you've been working on. Current frontier models hold the equivalent of several hundred pages of text at once, which is enough for nearly any teaching task and most research tasks. But long projects — drafting a curriculum, working through a textbook, sustained writing — push against this limit. The tools handle long context unevenly: they can hold a lot but they don't always reason equally well across all of it. Practical implication for teachers: for big projects, plan to break work into segments rather than feed everything in at once.

Tokens. When the companies talk about consumption they talk about tokens. A token is a chunk of text — roughly three-quarters of a word on average, though it varies. Subscriptions are denominated in tokens behind the scenes, and the costs you pay correspond to tokens the system reads (your input) and tokens it produces (its output). For consumer subscriptions you don't track this directly; the company handles the math and gives you tier limits. For institutional and API access the token cost is the cost. Worth knowing because it explains why some operations feel "expensive" — uploading a long PDF and asking for a detailed analysis consumes far more than asking a single question, and the system's usage limits track accordingly.

Near future. The price of the underlying capability has been falling steadily. The cost of running a model with a given level of capability has dropped by roughly an order of magnitude per year for several years running, and the trend is likely to continue for some time. This means the capability available at a given price point is rising fast. The model you get for twenty dollars a month next year will outperform the model you get for two hundred dollars today. Whether that translates into lower prices for users is a different question. The companies have so far chosen to use falling costs to offer more capable models at the same price points rather than to lower prices, and there is no particular reason to expect this to change unless competition forces it.

Far future. Here the picture is more contested. The optimistic version says capability will keep rising and costs will keep falling until AI access becomes effectively a utility — cheap, abundant, ambient. The pessimistic version says the most capable systems will become more, not less, expensive, because the marginal value of frontier capability rises as the systems become more useful, and the companies will price accordingly. The realistic version is probably both at once. Basic AI access will likely become very cheap and ubiquitous. Frontier capability — the kind that does real work for serious users — will likely remain expensive, gated, and concentrated in the hands of those who can pay. This is the pattern of every previous technology that started as a luxury and ended up everywhere. Some of it spreads. The most capable version stays at the top.

Now to the harder questions.

Individual versus institutional accounts. Most teachers reading this are using AI on individual accounts they pay for themselves, or on free tiers that limit what they can do. This is anomalous. The work that AI assists with is work the teacher is paid to do, on behalf of an employer who benefits from the productivity gain. The pattern is familiar. The teacher who buys their own classroom supplies, the social worker who uses their personal phone for client calls, the home health aide who drives their own car between appointments — workers paying for the tools the employer profits from is a structural feature of low-wage and credentialed labor in this country. AI access is the newest entry on that list.

The institutional alternative is for the employer — the school district, the CWE, the union local — to provide AI access as part of the conditions of work. This can take several forms. Bulk licensing arrangements with major providers, which lower the per-seat cost and give the institution control over data and privacy terms. Hosted instances of open-weight models, where the institution runs the infrastructure and pays no per-seat fee at all. Negotiated access through CBA language that treats AI tools the way prior contracts have treated computers, internet access, professional development, and reimbursement for work-related expenses.

The bargaining question. AI tools are increasingly necessary to do the job at the productivity level employers expect. Once a tool becomes necessary, it becomes a working condition. Once it becomes a working condition, it is a legitimate subject of bargaining. The question is whether the locals representing adult educators — and the locals representing every other category of worker now using these tools — start to write AI access into contracts before the cost shifts permanently to the worker, or after.

Some unions are already doing this. Writers Guild language on AI use in screenwriting was a landmark in the 2023 strike. SAG-AFTRA secured related language on voice and likeness. The question for adult educators is whether the next round of CBA negotiations includes language on AI access — who pays, what tools, what data protections — or whether the cost continues to be quietly absorbed by individual teachers buying their own subscriptions.

Where TPS concerns enter. The cost-and-access conversation, taken seriously, leads in two directions. The first is bargaining, which is the immediate and concrete move. Get language into contracts. Get employer-provided access. Get institutional licensing instead of individual out-of-pocket payment. This is arguably the work of the next contract cycle if not the present one.

The second is longer. Even if every union local in the country secured employer-paid AI access tomorrow, the underlying ownership structure would not change. The tools would still be controlled by five corporations whose business model is to capture rent from a technology built on collective work. Bargaining over the cost of access is necessary and not sufficient. The deeper question is whether workers — through their unions, through cooperatives, through public institutions, through whatever vehicles they organize — start to claim ownership of the AI infrastructure itself, the way prior generations of workers claimed ownership of housing, banking, and food production through cooperatives that survive to this day.

That is the long arc. It does not replace the bargaining work. It frames it. The contract gets you access to the tool this year. The ownership question is whether your students' children, who will use these tools daily for everything, are renting the tools or holding the deed.

I have placed this piece late, because the question gets clearer when the earlier pieces are in place. The question, in its anxious form, is: how do I assess writing now that students can have an AI write it for them. The question, asked better, is: what is assessment for, what does it tell me, and what should it tell me, in a world where the technology my students use to do their work has changed.

The anxious question has a short answer, which is that you cannot reliably detect AI-generated writing, the detection tools do not work, and trying to catch students will turn your classroom into a surveillance operation that fails at its stated purpose and corrodes the relationships that make adult education work in the first place. We can stop there on the detection question. It is a dead end. The energy spent there is energy not spent on the better question.

The better question is the one adult educators are already in a good position to answer, because adult education at its best has never been a pure credentialing operation. The GED is a credential, yes, and our students need it. But the work in our classrooms is broader, and often more urgent than test preparation. We teach speaking and listening for survival. We teach reading and writing as capacities for a life, not just as performance for an examiner. We teach mathematics as a way of thinking, not just as a sequence of procedures to reproduce on a Pearson form. We teach civics and science and history because the people in our rooms are citizens and workers and parents and neighbors who deserve the tools their society has accumulated, not because the test demands it. The credentialing is one outcome of the work. The work itself is wider.

This wider frame is the one that survives the arrival of AI. The narrower frame — assessment as detection of authorial labor — does not. So the practical question becomes: how do we assess in a way that actually tells us what we need to know, in a classroom where some students will use AI well, some will use it badly, and some will not use it at all.

A few principles that have emerged in conversation with teachers thinking seriously about this:

Move what you can to the room. Writing done in your presence, in real time, on paper or on a school device with monitored access, tells you what a student can do without assistance. This does not have to be all the writing. It can be one piece per unit, calibrated to take the temperature of where a student actually is. The point is not to surveil. The point is to have a baseline that you and the student can both see. Many adult students will welcome this if it is framed honestly: we are going to do some writing together in class so that I know what you can do on your own, because that is what the GED will ask of you on test day.

Assess process, not only product. Ask students to show you the steps. The outline. The first draft. The revision and what changed. The conversation with the AI assistant, if there was one, including the prompts and the back-and-forth. A finished piece of writing tells you less now than the trail of decisions that produced it. The trail is harder to fake, and more importantly, the trail is what you actually want to teach. A student who can show you their thinking is a student who is learning, regardless of which tools were in the room. A student who hands you a polished piece they cannot account for has not learned what the polished piece would suggest.

Make the AI use legible. Tell students that using AI is permitted, expected even, on certain assignments — and then ask them to disclose how they used it. I asked Claude to give me a topic sentence and I revised it. I had ChatGPT translate my draft from Spanish into English and then I edited the English. I used Gemini to check my grammar, and I disagreed with two of its suggestions, here is why. This kind of disclosure is the actual literacy you want students to develop. It is also unfakeable in an interesting way: a student who genuinely worked with an AI can tell you about it. A student who did not cannot.

Distinguish what the test will ask from what the life will ask. The GED, as it stands, is administered in a controlled setting. Test-day performance is what the credential measures. Your students need to be able to produce extended response writing without AI assistance because they will be asked to do so under examination conditions. This is a reasonable thing to teach toward and a reasonable thing to assess toward. Outside the testing room, life is different. Most writing your students will do for the rest of their lives — at work, with their families, in correspondence with institutions — is writing they can use any tool for. Teaching toward both contexts honestly is the work. Teaching as if the testing condition is the only condition that matters is dishonest in a different direction.

Do not retreat into oral assessment alone. Some teachers, faced with the AI problem, are moving toward all-oral assessment. This has uses, particularly for ESL, but as a wholesale substitute for written assessment it is a regression. Many of our students need the written practice, both for the test and for the life beyond it. Writing is a thinking discipline, not just a performance medium. Students should still be writing. The conditions and the framing are what change.

Use the AI to do some of the assessment. This is uncomfortable for some teachers but worth thinking about. AI assistants are good at producing rubric-based feedback on student writing, identifying patterns of error across a set of papers, generating differentiated practice problems calibrated to particular weaknesses, and producing first-pass evaluations that you can then review and adjust. None of this replaces your judgment. It can, however, free up your time for the parts of assessment that require your judgment — the sustained attention to a particular student, the reading between the lines, the relational work that no system will do because no system knows your students.

The harder thing this is asking of us. Most of what AI disrupts in assessment was already not working very well. The five-paragraph essay produced under timed conditions and graded against a rubric was a useful proxy for certain things and a distortion of others. The standardized test was always a partial measure, taken seriously for institutional reasons more than pedagogical ones. The hand-graded stack of homework was always more about completion than about learning. AI has not broken assessment. AI has stripped the cover from forms of assessment that were already doing less work than we pretended. The honest response is not to restore the old forms with more surveillance. It is to assess the things that matter, in ways that the new tools do not undermine — and many of those ways are the ones experienced adult educators have always trusted more than the bureaucratic forms anyway. Conferences with students. Portfolios. Process documentation. Real reading of real student work over time. Conversation as evidence.

The work of teaching is the relationship and the judgment. The tools change. The work does not.

Every language teacher knows that the question is this English correct? is more political than it sounds. Whose English. Correct by what standard. Correct to whom. The history of language instruction is partly a history of which English got named the standard, which speakers got told their English was deficient, and which students learned to translate their own speech into the dialect of their teachers and bosses before they could be heard. Adult ESL teachers in NYC work with this every day. The students in our rooms speak Spanish in fifteen varieties, English in five, languages whose grammars have nothing to do with European patterns, and the city dialect that gets spoken outside the classroom and corrected inside it. The question of what counts as correct is the central question of the work, and it is not a settled question.

AI systems have inherited this question in a particular form, and they have inherited it without inheriting the conversation that should accompany it.

The systems are trained on enormous quantities of written text — internet content, books, articles, code. The text is not a neutral sample of human language. It is overwhelmingly weighted toward written English, particularly the kind of English produced by professionals, journalists, academics, and people with the time and access to publish online. It includes much less of the spoken English of working people, the written English of non-native speakers, the English of immigrant and diaspora communities, the English of African American Vernacular speakers and Caribbean Englishes and the regional dialects of the American South and the urban Northeast. Other languages are even more unevenly represented — Spanish reasonably well, French and German well, Mandarin variably, and most of the world's languages thinly or not at all.

What this means in practice for the ESL classroom is that the AI systems are confident speakers and confident judges of one particular kind of English, and their confidence does not always track their accuracy. They will "correct" sentences that are perfectly fine in the variety of English the student is actually trying to learn or to use. They will smooth out distinctive constructions that a careful writer might have chosen on purpose. They will treat code-switching as error rather than as a sophisticated linguistic practice. They will reproduce the assumption that one English is the English and the rest are deviations from it.

A few specific patterns worth knowing:

Accent and pronunciation in voice modes. The voice features in current AI systems are trained predominantly on speakers of standardized American and British English. They handle these accents very well and other accents less well. A student with a strong regional or non-native accent may find the system mishears them, asks them to repeat, or transcribes their speech with errors that a native English-speaking interlocutor would not have made. This is the system's failure, not the student's. The pedagogical point is to name it as such rather than to let the student conclude that their pronunciation is the problem.

Grammar correction for non-native writers. When students paste their writing into an AI system and ask for corrections, the system will produce a "corrected" version that often goes beyond fixing errors and rewrites the prose in a more standardized register. The student's voice gets flattened. Constructions that worked, that carried meaning, that reflected the student's own language patterns, get smoothed away. For a student trying to develop their own written voice in English, this is a problem. For a student trying to pass the GED extended response, it is less of a problem because the test itself rewards the standardized register. The teacher's job is to help the student understand which goal they are pursuing in a given moment and to use the tool accordingly. Sometimes you want the smoothing. Sometimes you want to keep your voice and let the system fix only what is actually broken. The skill is knowing which is which.

Translation between languages. The systems are good at translating between languages they have seen a lot of, and the quality drops noticeably for less-represented languages. Spanish-to-English and English-to-Spanish translation is reliable, and the system can handle most varieties of Spanish well. Translations involving Haitian Creole, Quechua, Bengali, Yoruba, Tagalog, and many other languages are less reliable, sometimes much less. Worth verifying with a native speaker when accuracy matters — for parent communications, for legal documents, for medical instructions.

Cultural assumptions in examples and scenarios. When you ask the system to produce a dialogue practice scenario or a reading passage, the default cultural setting it produces tends to be middle-class American with a particular set of cultural reference points — restaurants, offices, suburbs, particular kinds of family structures, particular kinds of work. This is not always wrong but it is always particular, and it is not always the world your students live in. Asking the system to set scenarios in contexts your students recognize — a worksite, a clinic waiting room, a public school parent meeting, a court appearance, the kind of work an immigrant in this city actually does — produces material that lands differently. The system will do this if you ask. It will not do it on its own.

The deeper question. What we are looking at is not a glitch. It is a structural feature of how these systems are built. The training data reflects whose voices got recorded and whose did not, whose languages got transcribed and whose did not, whose English got named correct and whose got named broken. The systems are inheriting a long history of linguistic prestige, and they are reproducing it at scale, in classrooms and offices and voice interfaces all over the world. None of the major labs has solved this. Some are working on it. Some are not. The labs that are working on it are doing so within the constraints of a business model that incentivizes performance on benchmarks tilted toward standardized varieties.

For adult ESL teachers, the practical implication is that you are now using a tool that has opinions about English. You and your students need to know what the tool's opinions are, when to take them seriously, and when to push back. This is itself a literacy worth teaching. Students who learn to evaluate the system's corrections critically — to ask is this actually wrong, or does this system just prefer it differently — are learning something more useful than acceptance of the corrections themselves. They are learning that language judgment is a contested activity, that different audiences expect different registers, and that fluent speakers move among registers deliberately. That is the actual ESL skill. The AI system can help build it, if it is used with awareness of what the system is and is not.

The students in our classrooms bring with them, often, lives that have reasons to stay partially out of view. Immigration status. Custody battles. Records that follow people. Domestic situations that require care about who knows what. Medical conditions that affect work and learning. Histories with employers, landlords, courts, and agencies that have not always treated them well. Adult education in this city has always been, among other things, a place where these complications are held with appropriate discretion. The teacher who teaches an undocumented student a verb conjugation does not need to log that student's status into a system. The teacher who helps a parent with a custody case write a letter to the court does not need to put the child's name into a corporate database.

AI systems complicate this in specific ways that are worth understanding. Most teachers using these tools have not thought carefully about what happens to the text they paste in, where it goes, who can see it, how long it is retained, and what the company does with it. The companies' answers to these questions vary, and they change. This piece is a working primer on what to know and what to do.

What happens when you paste something into an AI system. Your input travels to the company's servers, where it is processed and a response is generated. The input is logged. Depending on the company and the plan, the input may also be used to train future models, retained for safety review, or accessed by company employees for quality monitoring. The default settings for consumer accounts tend to retain conversations and may use them for training. The settings for paid and enterprise accounts vary, with stronger protections on enterprise plans. The free tier of most major systems offers the weakest privacy protections. Worth noting that "weakest" here does not mean "no protections" — the companies have policies, some of them substantial — but that the worker using the free tier has the least leverage and the least information about what is happening to their data.

What this means in practice for adult education. Do not paste student work containing identifying information into a free-tier consumer account. Identifying information includes: full names, addresses, dates of birth, ID numbers, immigration documents, court documents, medical records, anything that identifies a particular student's situation. A student's draft essay about their experience with the immigration system, their parent's deportation, their child's custody case, or their medical diagnosis is exactly the kind of material that should not enter a corporate training corpus, even with good intentions, even for the purpose of helping the student.

The harder version of this principle is that it applies even when the information feels routine. A student's writing about their job, their employer, their union, their family situation, their neighborhood — none of this is dramatic, none of this seems like the kind of data anyone would care about, and all of it is exactly the kind of texture that aggregates into a profile of a person's life. The aggregation is the problem. Any single piece of writing seems harmless. The accumulated record is not.

Practical workflow adjustments. Several options exist, depending on what you are doing.

The chapter's specific position. The chapter knows its membership and its students in ways that no admin-level or system-level policy can capture. A blanket institutional policy on AI use, drafted by people not in the room, will reflect the institutional risk calculus — protect the institution from liability — rather than the worker-and-student calculus — protect the people whose lives are in this writing. The chapter has done this kind of analytical work before, on other terrain, and it should do it on this one. What does AI use look like in classrooms where students may be undocumented? Where students are custody-affected parents? Where students have been harmed by the very institutions whose data they would be feeding? These are not abstract questions. They are the chapter's questions, and the chapter's answers will be more careful than any answer that comes from above.

A note on reassurances. The companies will tell you their systems are safe, their data practices are strong, and that worry is unwarranted. Some of this is true. Some of it is marketing. The honest position is that current data practices are reasonably good for most consumer purposes and not necessarily good enough for the specific populations many of our members teach. The companies are not designing for adult learners with vulnerable status. They are designing for the general consumer market and adjusting at the margins for enterprise customers who have leverage. Adult and worker education programs do not currently have that leverage, individually. Collectively — through chapters, locals, CBOs, the CWE network and adjacent ones across the country — there is more leverage than is usually exercised. The first step is to know what you are looking at.

One more thing. The same caution applies to the teacher's own writing. When a teacher uses AI to draft a sensitive communication — a note about a student to a colleague, a record of a difficult conversation, a draft of a complaint or a grievance — that text goes through the same systems with the same data practices. The chapter has fought, on member behalf, against many things that ended up in records. It is worth being careful about what we put into new ones.