The Day an AI invented books that don't exist
Early in building Transforming Access, our first-grade science curriculum, someone stepped outside our normal process and asked a general-purpose AI chatbot for a list of recommended books to pair with a lesson. What came back looked great. Real-sounding titles, warm little descriptions, and authors who were real people, some of them people we knew.
Some of those books did not exist. They weren't out of print or hard to track down. The AI invented them and attached them to real authors, because that's what these tools do when they don't have the answer. They fill the gap with something that sounds right, and they do it with total confidence.
We caught it because our process requires a person to check every single item, and the process caught the shortcut that went around it. But I keep thinking about the versions of that story where nobody checks. A parent prints the list and spends an afternoon at the library hunting for a book no one ever wrote. A teacher builds a unit around a source that isn't real. And now imagine that same confident invention happening inside the science content itself, in a lesson you're teaching tomorrow.
If you've seen the posts that say "I just gave my curriculum topics to ChatGPT and my planning is done," this is the part they leave out.
Why this happens, in plain language
The industry word for it is hallucination. I'd just call it making things up. General-purpose AI tools generate plausible text rather than checked facts, and there's no warning label on the made-up parts. A fabricated book title looks identical to a real one. A wrong fact about the water cycle gets plunked down in the same confident sentence as a right one. And the longer a session runs, the less you can trust it. A chatbot can start out working well and then drift back to its defaults a dozen prompts in. What does that look like? The lesson you generate on Tuesday afternoon may be sloppier than the one from Tuesday morning. You might feel like you have a good groove in a particular thread, but your chatbot drifts off and loses the plot.
That risk lands hardest on the learners we serve. Nonspeaking and partially speaking students who use augmentative and alternative communication (AAC) often have fewer ways to flag confusion in the moment, and they've spent enough of their educational lives being underestimated. They deserve content that's actually true.
A tool, like every tool before it
I want to be clear that I'm not against these tools. The preparation burden on teachers and families is real, and AI can lift part of it. We use AI ourselves.
Think about how we've navigated over the years. First paper maps and AAA TripTiks, those spiral-bound booklets with your route highlighted page by page. Then MapQuest printouts riding shotgun. Then Google Maps on a phone, and now Waze rerouting you around an accident in real time. Each step handed more of the work to the technology, and nobody (except your luddite uncle) misses tracing a route by hand with a highlighter. But two things never changed. You still decide where you're going, and you still have to watch the road. If you've seen The Office, you already know where this goes: Michael Scott drives straight into a lake, with the water in plain view and Dwight shouting from the passenger seat, because "the machine knows!"
When we work smarter, not harder, the tool plans the route, the person drives.
AI in education is the same kind of shift, and the same rules apply. It can take over the route-planning layer of teaching, the formatting, the drafting, the first pass. It cannot take over the destination or the driving: deciding what a learner should reach, and watching what's actually in front of you. The danger comes from a growing habit of letting it do both, from sitting back and saying, "the machine knows!" while our car careens into the lake. We need to be mindful of how we're using these tools, especially in education, where the person on the receiving end is a child who has no way to know the lesson was never checked. Use it with thought, not as a replacement for thought.
How a small team built 67 lessons
The team behind this project was basically four people, and we built 67 complete lessons, each with scripts, slides, and three worksheet versions, in six to eight months. The tool that made that possible is called Adaptiverse. Full transparency: I co-founded it, and it's a separate company from REV, so you should know I have a stake in both.
Here's why we didn't just use a general chatbot, and why I'd ask you to think twice before you do.
First, off-the-shelf AI doesn't share our assumptions. Ask it to "adapt" a lesson and it does what the field has always done: it lowers the reading level, cuts the vocabulary, and swaps reasoning questions for recall. It reproduces the deficit model, the old assumption that the learner is the problem to be worked around, because that's what it learned from. You can see this in research from the disability community itself: a 2026 study with CommunicationFIRST and other advocacy organizations found generative AI defaulting to outdated and offensive language about disability, which participants attributed to disabled people being absent from how these tools are built.
Second, chatbots hallucinate, which is how we ended up with a lovely list of books nobody wrote.
That's why Adaptiverse works differently. It's built on decades of REV's educator-written materials, and error checking is built into every layer of the process. The AI checks its own work against rules we wrote into every request it handles, and in some cases a second, separate AI model reviews the output, so one system is catching the other's mistakes. Names, titles, and places are checked against named sources, and those sources are tracked through the whole generation process, which is why you get clickable citations you can trace back instead of taking the machine's word for it. It's also constrained to hold grade-level intellectual demand and to draft all four Scaffolding Framework question types, instead of drifting back toward the deficit model.
And even then, the tool never has the last word. Every lesson, script, and worksheet in Transforming Access was checked and refined by an experienced educator. We verified the science, confirmed every recommended book is real and worth reading, rewrote questions that didn't sit right, and adjusted language to match how we actually talk to learners. Nothing reached a family without a person reading it closely first.
The AI gave us speed. The verification is what makes it a curriculum. It's also why this took months instead of weeks. The human finishing is the slow part, and it's the part that makes the result trustworthy. That's the model I trust: the tool drafts, people verify, and the combination is stronger than either one alone. You can see the result in the free lessons from Transforming Access, 30 first-grade science lessons open to everyone thanks to grant funding from The Arc Maryland and the Maryland Developmental Disabilities Council.
Questions to ask before you trust an AI tool with your learner's education
Whether you're a parent evaluating an app or a teacher whose district just bought a license, these questions will tell you most of what you need to know.
Where does the content come from, and can I trace it? If the tool can't show sources, you're trusting the model's memory, and the model's memory includes things that never happened.
Who reviews the output before a learner sees it? If the answer is "no one," then the answer is you. That's workable, but only if you know it's your job.
What happens when the tool doesn't know something? Tools with guardrails say so or cite out. Tools without them fill the silence with fiction.
Was it built by educators, or just trained on the internet? A tool built around an instructional model behaves differently from a chatbot doing an impression of a teacher.
Does it keep expectations at grade level? Some tools "simplify" content for disabled learners by lowering it and calling it support. Look for tools that change the support, not the expectations.
Can it invent a citation, a book, or a fact without flagging it? Ask the company directly. If they say it can't happen, be more careful, not less. Every AI system can be wrong; the good ones are built to make the wrongness catchable.
And one question for yourself: do I have the time to review what this tool produces? If the answer is no, then a curriculum that was already reviewed by educators will serve your learner better than raw AI output, no matter how fast the AI is.
Frequently asked questions
Can AI tools like ChatGPT create lesson plans?
Yes, and quickly. But general-purpose chatbots generate plausible text rather than verified facts, so lesson plans made this way can include invented sources, fabricated book titles, and factual errors with no indication anything is wrong. They're drafts that need human review, not finished lessons.
What are AI hallucinations in education?
Hallucination is the industry term for an AI presenting made-up information as fact. In education this can mean fake citations, books that don't exist, or incorrect content taught to students as true. It matters more in education than in most fields because the audience often can't spot the errors.
I heard some AI models hallucinate less. Doesn't that solve the problem?
Some models do make things up less often than others, and the differences are real. None of them are at zero, and no company building these tools claims otherwise. A lower error rate changes how often you'll catch something, not whether you need to look. The useful question isn't "which model is best?" but "what happens between the AI's output and my learner?" A stronger model with no review will still eventually put a made-up fact in front of a child. A decent model with an educator reading behind it won't.
Is AI-generated curriculum safe for students with disabilities?
It can be, when AI drafting is combined with human review by experienced educators and the tool keeps content at grade level while adjusting support. Unreviewed AI output carries extra risk for students who have fewer ways to flag confusing or incorrect material.
What does the disability community say about AI?
The most useful recent snapshot is a June 2026 white paper from the Consumer Technology Association, produced with Access Living, Disability Belongs, and CommunicationFIRST, based on interviews, surveys, and focus groups with disabled people about how they actually use AI. The findings hold both truths at once. AAC users in the research were optimistic about AI, especially word prediction, which participants credited with helping them communicate more fully and keep pace in conversation. And the same research surfaced the concern this post is built on: tools developed without disabled people's input don't meet the community's needs, down to generative AI defaulting to outdated and offensive language about disability. Participants also cautioned that AI editing tools can change what a communicator is trying to say. As one AAC user put it, "I am my own voice." That's the standard any AI tool in education should be held to: it should carry the learner's thinking and the teacher's judgment, not replace either one.
What is Transforming Access?
Transforming Access is a first-grade science curriculum from Reach Every Voice, built for all learners and intentionally designed to include nonspeaking and partially speaking students who use text-based AAC. Lessons were drafted with AI assistance through Adaptiverse and then reviewed, edited, and customized by experienced educators. Thirty of its 67 lessons are free.
Isn't AI bad for the environment? I've heard about the water use.
The concern is worth putting real numbers on. Google reports that a typical text prompt to its AI uses about 0.24 watt-hours of energy and about 0.26 milliliters of water, roughly five drops, and OpenAI's stated figures are similar: about 0.34 watt-hours, or what an oven draws in roughly one second, and about a third of a milliliter of water. A full lesson draft is a longer exchange than a typical prompt, but it's the same order of magnitude: drops of water, and less electricity than the laptop it's displayed on uses in a couple of minutes. Compare that with what it replaces. An educator adapting a lesson by hand might spend two or three hours with a laptop running, a dozen browser tabs open, and a show streaming in the background, and that evening of screen time uses many times the energy of the AI call that drafts the same lesson in minutes. The environmental questions about AI are real at the scale of billions of daily queries and the data centers built to serve them, and they deserve scrutiny. But for one educator preparing one lesson, the AI-assisted version is the lighter footprint. And whatever the footprint of a single draft, the question that matters most for your learner is the same one this whole post is about: did a human check it?
Lisa Mihalich Quinn, M.A / M.Ed. / MBA / ATACP is a licensed special educator with more than 15 years of experience making academic content accessible for neurodiverse students and learners who use Augmentative and Alternative Communication (AAC). She is a former Maryland Public Schools teacher and the founder of Reach Every Voice, an organization dedicated to empowering individuals with communication access needs, and cofounder of Communication for Education, an online training program for people who support students using text-based multimodal communication in educational settings. Most recently, she has been working to lift the burden of content adaptation on parents and educators with the new transformative lesson adaptation tool, Adaptiverse App.
Lisa's passion for inclusion and equity runs deep, driving her work to help educators, learners, and families think creatively about how to reimagine systems that are historically resistant to change. She pushes folks to shift mindsets from "this is just how we do things..." and "we can't because..." to embody a spirit of "what if we tried..."
.png)
.png)
.png)
Comments