When I talk with districts about generative AI, the pace of adoption reminds me of another material that arrived as a miracle: plastic. Plastics were invented early in the twentieth century and took off after the Second World War because they were cheap, light, and could be shaped into almost anything. Every one of those selling points has an echo in how AI gets pitched to schools, which promises new products, lower costs, more efficiency, and the ability to adapt to any subject, any grade, and any context.
Then there is what happened next. Geyer, Jambeck, and Law’s 2017 accounting in Science Advances estimated that 8,300 million metric tons of virgin plastic had been produced by that point, and that of the roughly 6,300 million metric tons of waste generated by 2015, about 9 percent had been recycled, 12 percent incinerated, and 79 percent had accumulated in landfills or the natural environment. The material that solved so many problems became a problem that does not go away.
I do not think the analogy is perfect, and I will get to where it strains. It has stayed with me because it names two costs at once. One is physical: the energy and water behind every model that gets trained and every prompt that gets answered. The other is the flood. A material that is cheap to produce gets produced in volume, and the volume becomes the waste. Low-cost, low-quality AI text is already piling up in the places students go to look things up, and it is getting harder to find the teacher-made, human-checked material underneath it.
My own relationship with these tools is complicated, and I want to be honest about that before offering any advice. I remain concerned about models trained on writing and images collected without consent or compensation. At the same time I use generative AI regularly, in my own work and in the tools I build for districts, and I find parts of it useful. I am not making a case for abstaining here, only trying to work out what using less, and using it more deliberately, looks like once you have decided not to walk away.
What is the footprint?
Large language models need substantial computing power to train and to run, and that translates into energy and emissions wherever the electricity is not renewable. The training figures that circulate, such as a single model emitting as much carbon over its training as several cars do over their lifetimes, come from a small number of studies of specific models and are better read as illustrations of scale than as numbers for any particular product. What is easier to say with confidence is that everyday use has become the larger and faster-growing share of the cost. A single query is small. Millions of people making them all day are not, and the pattern of use is the part any of us can actually change.
Which model should you use?
Picking a lighter model is the most direct way to shrink the footprint of a task without giving up the task. Smaller models handle a surprising share of routine work, drafting, summarizing, and reformatting, at a fraction of the compute of the largest ones. Models built for a specific job often outperform general-purpose models on that job while using far less. And when a small model can run on the device in front of you, through something like LM Studio, the cloud infrastructure and the round trips to a data center drop out of the equation entirely. None of this is exotic. It is closer to choosing the right size of vehicle for a trip than to any technical feat, and at the scale of an organization the difference adds up.
Should you start at all?
The most environmentally friendly AI interaction is the one that never starts. Before I open a chat window I try to ask whether the tool is adding something or whether I am reaching for it out of habit. A factual question that a search engine can answer does not need a generated paragraph around it. An image that exists to decorate a slide for one meeting is a real cost for a momentary purpose. When I have built AI into a workflow, the useful version has been narrow: a few places where the model clearly improves the outcome, rather than a layer across everything. In the tools I build for districts I try to route requests through plain keyword matching before anything reaches a model, so the expensive step only runs when it has to. The FAQ Chatbot is built that way.
How you interact matters too, though less than whether you interact at all. A complete, specific prompt tends to get a usable answer in one pass, where a vague one produces a chain of clarifications that each cost compute. Asking for the level of detail you need rather than an exhaustive treatment shortens the response. Grouping similar tasks into one request instead of opening a fresh session for each is a small saving that repeats. These are habits rather than rules, and individually they are minor, but they are the habits that scale across a staff.
What can organizations do?
Organizations that adopt these tools widely have a different kind of responsibility, because their defaults set the habits of everyone under them. The practical moves are unglamorous. Tiered access can steer routine use toward efficient models and reserve the heaviest ones for the tasks that need them. Tracking AI-related energy use as an actual metric, even roughly, makes the cost visible in a way that a line item for licenses does not. Carbon offsetting can balance emissions that cannot be avoided, though I would not want anyone to mistake an offset for a reduction.
Beyond the organization, the changes that would matter most are upstream: data centers on renewable power, efficiency treated as a research priority rather than an afterthought, disclosure standards so that a district can compare the footprint of two providers, and regulation that puts environmental cost on the table alongside privacy and safety. Individuals do not control any of that. Districts, cooperatives, and state agencies have more leverage than they usually exercise, mostly through what they ask for in procurement.
Where does the analogy strain?
I should say plainly that this argument is easy to turn back on me. Schools are already wasteful in ways nobody writes essays about. Forests’ worth of worksheets get printed daily. HVAC systems keep empty gyms at an even temperature all summer. I generate cover images for articles that do not strictly need them, in the same way I order things from Amazon that I could pick up in town. I am not using the plastic comparison to say generative AI is uniquely sinful, only that we have watched this pattern before and have some idea of how it ends.
The comparison strains in one other way. Plastic did not become a global problem because individuals chose the wrong bag at the grocery store, and it is not going to be solved by better bag choices either. The scale of the problem was set upstream, in what was cheap to make and profitable to sell. I suspect the same is true here. The choices I have described are worth making, and I make them, but I hold them next to the knowledge that the footprint of this technology is being decided somewhere I do not have a seat. What I can do is keep my own use deliberate, keep it visible to the educators I train, and keep asking the questions in procurement conversations where someone might actually have to answer them.
If you have found ways to make this visible in your own district, or if you think the plastic comparison breaks down somewhere I have not noticed, I would like to hear it. You can reach me at licht.education@gmail.com, and there are more tools, articles, and resources at bradylicht.com.
