The science of reading and AI-written stories
What the science of reading says about phonics, decodable and leveled text, how AI-written stories can follow it or break it, and how Peabody limits each story.
The science of reading says children learn to read by connecting letters to sounds, taught in a planned order, and then practising on text they can actually sound out; comprehension then rests on language and knowledge. An AI can write a story at a reading level, but a 2025 test found no model hit a 4th-grade target on its own, and readability formulas say nothing about phonics. Peabody (peabody.kids), free through the end of 2026, measures every story in code and limits early stories by phonics step.
| Finding | Number | Source |
|---|---|---|
| Systematic phonics vs other instruction | d = 0.41 overall; 0.55 begun in K or 1st, 0.27 later | Ehri, Nunes, Stahl and Willows 2001 (National Reading Panel data) |
| The same data, weighted equally | d = 0.24 | Camilli et al. 2003 (size contested, not direction) |
| Parents listening to a child read | d = 0.52 | Sénéchal and Young 2008, 16 studies |
| Parents reading to a child | d = 0.18, not significant | Sénéchal and Young 2008 |
| Language models simplifying text to grade 4 | None of the model and prompt combinations hit the target | Day et al. 2025 |
What "the science of reading" means
The phrase covers a body of research, not a product. Its core is the Simple View of Reading from Philip Gough and William Tunmer (1986): reading comprehension is decoding multiplied by language comprehension. If either is near zero, so is reading. Anne Castles, Kathleen Rastle and Kate Nation's 2018 review, "Ending the reading wars", sums up the consensus: systematic phonics is necessary and well supported for learning to decode an alphabetic language, and it is not sufficient, because understanding what you read depends on vocabulary, language and knowledge of the world.
The phonics evidence comes largely from the National Reading Panel's analysis, published in full by Linnea Ehri and colleagues in the Review of Educational Research in 2001: 66 comparisons from 38 experiments, an effect of 0.41 overall, larger when phonics starts in kindergarten or first grade (0.55) than after (0.27). A 2003 reanalysis that weighted studies differently found a smaller 0.24. The size is argued over; the direction is not.
How a word becomes a sight word
Ehri's phases describe what a beginning reader does with a word: guessing from a picture or the first letter, then sounding out letter by letter, then reading in chunks. Her theory of orthographic mapping says a word becomes instantly recognised when its spelling, sound and meaning are bonded, and that even irregular words are mostly regular. David Share's self-teaching hypothesis (1995) adds the practical rule: each time a child successfully sounds out a new word on their own, that is a learning event for that word. Studies put the number of exposures needed to learn a new word form at about four to six, more for weaker readers.
The design rule follows: a beginning reader's text should be full of words they can sound out with the patterns they have been taught, so every successful decode counts.
Decodable text, leveled text, and the argument between them
| Decodable readers | Leveled readers | |
|---|---|---|
| Built around | The phonics patterns taught so far | Overall difficulty: sentence length, word frequency, picture support |
| What it trains | Sounding words out | Reading fluently at a comfortable difficulty |
| The risk | Stilted, odd sentences; little story | Predictable text invites guessing from pictures and context |
| The evidence | Decodability raises decoding-strategy use and accuracy (Cheatham and Allor 2012); overall evidence called inconclusive | Widely used; the guessing it invites is the strategy Ehri's model says a reader must outgrow |
Jill Cheatham and Jill Allor's 2012 review in Reading and Writing found that decodable text increases the use of decoding strategies and accuracy, and argued for using several criteria, not decodability alone. The University of North Carolina's literacy institute calls the evidence on decodable books inconclusive and recommends them as a temporary scaffold. The common ground is that decodable text is for the first months of reading, and leveled, real-language text comes after.
Where an AI-written story can go wrong
A language model writes fluent, appealing stories. It does not count phonics patterns, and it tends to write harder than it is asked.
The clearest test is from the Readability Consortium. Day and colleagues (2025) asked GPT-4 Turbo, Claude 3 and Mixtral to simplify sixty 12th-grade passages to grades 8, 6 and 4, with four prompting methods and a multi-agent pipeline. Several combinations hit grades 8 and 6. At grade 4, none did. The research summary behind this post found nothing published on controlling a model to the phonics patterns of a kindergarten or first-grade reader, which is below where readability formulas even apply. And a formula like Flesch–Kincaid counts only sentence length and syllables, so a model can hit the number with short sentences of rare words.
Three things follow for any AI story tool:
- Measure the output, don't trust the prompt. The level has to be checked in the text that was written.
- For beginners, check phonics, not a formula. A story for a child on short vowels needs words with short vowels.
- Don't confuse a story with an assessment. A child tapping through a story says nothing reliable about their decoding.
How Peabody writes and limits a story
AI writes Peabody's stories; code measures and limits them.
- Thirteen levels, measured in the text. Each story is measured in code: Flesch–Kincaid with dictionary syllable counts, sentence length, words per screen, the share of words off the Dale–Chall familiar list, and for early readers the share of sight words and of words the child can sound out. A short text's level is reported as a range, because any formula wanders on a few hundred words.
- The writer aims low. The prompt asks for the easy half of the level's range, because a model writes harder than it is asked.
- Too hard means repair. A draft over its level's limits is repaired, a few screens at a time, or rewritten. If it still misses after the repairs, it is kept and marked as having missed rather than thrown away, so a few stories run above their level.
- The first level is limited by phonics step. For level 1, Peabody follows a phonics order from short-vowel words through blends, digraphs, silent e, endings, two-syllable words, r-controlled vowels and vowel teams. A word the child cannot yet sound out at their step is counted, and a level-1 story may hold at most six of them; over that, the repair swaps words. From level 2 up the phonics check retires and the reading-level measures take over.
- Peabody's own library is held to more. Level 0 library books must be at least 90% sight words or words a beginner can sound out, and level 1 at least 85%.
- A word counts as known slowly. A practice word is known only after three right first reads, without help, on three different days.
- Read aloud first. A new story is read aloud by default the first time, so a child hears the language while the decoding is still hard.
There is one place Peabody departs from strict decodable practice. Every page at the first levels carries an emoji picture, with the word it shows coloured to match. Pictures can invite guessing, which is the habit decodable text tries to break; Peabody chose them because its youngest readers are four and five. It is a trade-off, stated so you can judge it.
A sample of stories is also read by a second, stronger model for writing quality: each child's first five, then about one in four. That review scores Peabody's writer, never the child.
What Peabody does not do
- It does not listen to a child read. It cannot measure fluency or decoding, and it does not claim to.
- It doesn't measure Lexile or give any norm-referenced score. A Peabody level is a place in Peabody, not a grade equivalent.
- It does not teach phonics lessons. The stories and word practice give a child text to read at their step; explicit phonics teaching still belongs to a teacher or a parent.
- It has not been tested in a trial. Nothing here is a claim about what it does to a child's reading.
The research on parents is the most useful part for tonight: listening to your child read (d = 0.52 in Sénéchal and Young's 2008 meta-analysis) did much more than reading to them (0.18). Pick a story at their level, let them read the first page to you, and help with the sounds when they stop. Start free or open the library.
Learn more elsewhere
- Reading Rockets: Reading 101: a free course on phonics, fluency, vocabulary and comprehension.
- Reading Universe: a free map of reading skills with classroom videos and phonics checks.
- IES guide: foundational skills for K-3 reading: four recommendations rated for evidence, with take-home tip sheets for parents.
- Shanahan on Literacy: a careful blog by a former member of the National Reading Panel.
For about seventy more, checked and sorted by who they're for, see where to learn about AI.
Questions parents ask
What is the science of reading in simple terms?
The research showing that children learn to read by linking letters to sounds in a planned order, practising on words they can sound out, and building the vocabulary and knowledge needed to understand what they read.
Are AI-written stories decodable?
Not on their own. Models write fluent text but do not count phonics patterns, and in a 2025 test none hit a 4th-grade readability target. A tool has to measure and limit each story in code, as Peabody does for its first level.
Should my beginning reader read decodable or leveled books?
Both, in order. Decodable text is a scaffold for the first months, so each successful decode teaches a word. Leveled, real-language stories come as the child's phonics grows.
Does Peabody replace phonics instruction?
No. It gives a child stories and words to read at their step and measures each story's level, but it does not teach phonics lessons or listen to a child read.
Checked against how Peabody works on 18 October 2026. Written with help from AI. Peabody is a reading app, not a reading assessment: if you're worried about your child's reading, talk to their teacher.