Julie Highsmith remembers how hard it used to be to find stories her kindergarteners could use for reading practice.
The phonics program she uses at her school in Florida’s Palm Beach County district only provides two decodable texts per lesson—short passages designed to align with the specific letter-sound patterns students are learning. Highsmith would hunt for extras from other sources, but they didn’t always match the curriculum’s sequence of phonics skills.
But last school year, she found an AI tool called Project Read. Highsmith, who now teaches 1st grade, selects her curriculum and lesson, and the program crafts endless decodable texts her students can use for practice.
Now, she estimates that 75-80% of the decodables her students read are AI-generated.
“It’s so easy to differentiate,” Highsmith said. “It’s only upsides. It’s really so simple.”
As AI use continues to grow in classrooms across the country, more schools and early reading teachers are now asking the technology to write passages for their students to read.
At least 100,000 individual educators use Project Read, according to the company. Other tools that offer passages beyond decodable text are popular too. Over the past year, more than 180,000 educators have used Text Rewriter, a tool from the education AI company MagicSchool that adapts texts to different reading levels, according to the company.
AI-generated stories seem to offer a solution to a longstanding need teachers have in the classroom: Young kids learning to read require lots of practice and exposure to text; AI can provide endless options, tailored to each student’s learning needs.
But some in the field suggest that there’s reason to be cautious about what these tools create.
Crafting the right teaching text is tricky, researchers say, and requires bringing together many different sources of information: a curriculum’s scope and sequence, scales for rating text complexity, an understanding of children’s language development, and an individual student’s specific challenges, among them.
AI writing also has its own patterns and quirks, said Thomas Juzek, an assistant professor of computational linguistics at Florida State University.
Kids who get a steady steam of these texts could experience less variety in their reading diet. And, while there is still a lot to learn about the best mix of reading materials, there is consensus among reading researchers that kids need a variety of text types, both to acquire background knowledge and use their reading muscles.
“It’s a big question,” Juzek said, of the long-term effects of regularly reading AI text. “What does that do to us?”
AI writing shows unique hallmarks
Studies of AI writing in general show some well-established differences from human writing. AI prefers certain vocabulary, phrases, and sentence structures. AI writing and editing also default toward a neutral, middle ground, shying away from taking a strong position on either side, said Juzek.
Importantly, he said, this doesn’t just happen when AI models are asked to write a new text. It happens when they’re asked to rewrite text, too. “Even when we think we’re doing a grammar check, the models shift the meaning of what we write,” he said.
Applied to reading practice, it’s possible that these features of AI text could shape children’s reading and writing abilities, said Maryellen MacDonald, a psycholinguist and emerita professor of psychology at the University of Wisconsin-Madison.
“If, for example, some AI decided to use lots of passive sentences like, ‘Joey got scared,’ ‘The book got torn,’ and kids read a lot of them, they would get better at reading passive sentences, and they would be missing out on getting better with other kinds of sentences,” she said.
But that would likely depend on reading a heavy volume of AI text, MacDonald said.
Minimal research details how well AI meets the demands of writing text for classroom use.
One 2025 study tested six publicly available large language models on the task. Researchers asked several versions of ChatGPT, Google’s Gemini, Meta’s Llama, and Diffit—the only one of these specifically designed for educators—to create two kinds of passages: decodable text, and general passages written at a specific grade level.
For the leveled text, the technology reliably wrote passages at a higher grade level than the one specified, as measured by Lexile levels and the Coh-Metrix, another measure of text complexity. The decodable texts generated did include the letter-sound patterns the researchers asked for, but the passages often sounded stilted or unnatural, and included multisyllabic words that could be challenging for beginning readers.
In the year since the study was published, AI tools have gotten better at some of these tasks, said Linling Shen, a doctoral student at the University of Texas at Austin, and the lead author on the study. In her current work, Shen has created more detailed text-creation prompts that have led to more accurate text leveling.
Even so, she said, the passages AI returns don’t always sound natural. They sometimes use language in ways that could confuse young readers, she added, like over-relying on pronouns.
“Written language is hugely ambiguous,” said MacDonald.
Text for young readers that goes through normal publishing processes typically strips out some of this ambiguity, with “a lot of eyes” evaluating whether a child will be able to understand it, she said.
“Something that a teacher generates with these AI tools is not going to get that same kind of vetting,” she said.
How teachers say the tools stack up
If publicly available, general-purpose AI tools appear to fall short for creating reading materials, what about programs specifically built to write these texts?
Their numbers are growing. Tools like Project Read and LitLab create decodable texts, designed to be aligned to the curriculum program of the teacher’s choice. AI for multi-purpose classroom use, like MagicSchool and School AI, both have tools that teachers can use to rewrite text to meet different reading levels.
Big publishers have integrated similar features, too. HMH AI tools allow teachers to create texts and readings; Amplify’s AI capabilities can create decodable text; and Savvas’s “Studio for Literacy” can craft reading selections.
It is unclear how well these tools accomplish their intended purposes, as they have not been well studied. The companies that create them say they’ve put in place guardrails to avoid common problems.
Project Read is sensitive to the order in which different curricula introduce letter-sound patterns, and how they teach different word structures, said Vivek Ramakrishnan, the company’s CEO. Its AI learns from exemplars of well-written decodable text, he said, and a human-built assessment of decodability is used to refine the model and test output.
On MagicSchool, the Text Rewriter operates with “explicit direction on writing for a target grade level,” said Caroline Humphrey, the staff product manager on MagicSchool’s Enterprise AI Systems team, in an emailed statement.
“We regularly run internal evaluations that measure our tools’ outputs against grade-level readability benchmarks,” she said, testing the same prompts multiple times and using the results to further refine the tools.
Teachers who use custom-built AI tools to write text say they’re not perfect—but still useful.
Jean Gunderson, a Title I reading interventionist in Elkton, S.D., uses Diffit to rewrite comprehension passages at a lower level for her 4th-6th graders and generate comprehension questions.
There are some hiccups. The articles can sound stilted, and sometimes the program swaps out more complex vocabulary for simpler words that don’t exactly have the same connotation, she said. Once, she received a comprehension question asking why a character didn’t like playing soccer, when he hadn’t played soccer in the story.
But on the whole, the output is “pretty good,” she said. “As somebody who’s been up until midnight some nights writing stuff like this, I was just blown away.”
Diffit did not respond to a request for comment.
Still, Gunderson said, she relies heavily on her professional judgment, honed through years of experience, to sort through the materials Diffit generates. Newer teachers might struggle to do the same, she said.
AI-generated text could be ‘an important training ground’
As AI companies continue to refine their models, the errors that some teachers and researchers have observed in AI-generated text will likely diminish.
But more abstract questions about how the text will shape children’s understanding of language and books are harder to answer.
“The decodables have come a long way in the past 10 years. There are now more that have a wider range of characters, of topics; there are some that are really clever that I don’t think AI can do,” said Anne Murphy Karabell, a researcher at American University’s Baker School of Education who studies early childhood literacy.
Karabell is also a member of Schools Beyond Screens DC, part of a national coalition that advocates for stricter guardrails around technology in the classroom.
Holding a book and being able to turn the pages “makes the reading come alive” in a way that a single sheet printed off a computer may not be able to, she said.
But MacDonald, the psycholinguist, said that many schools don’t have stacks of decodable books as an option.
“If the AI tool makes it easier to move them beyond single word reading, and get them reading more texts, we don’t know yet, but I would be inclined to say that’s a short-term win,” she said.
“The patterns of book language are really strikingly different [than spoken language],” MacDonald continued. “They require learning a different dialect of how language works. That happens through exposure, reading books. The AI-generated texts may not be exactly like the book language humans write, but they’re going to be an important training ground anyway, if the alternative is not getting to read connected text.”
