We know that many children are failing to learn to read in our nation’s classrooms. So it’s tempting to hope that tutoring by artificial intelligence could provide a solution. But how close are we to this for our earliest readers? And if AI isn’t yet the solution, are we making full use of digital tools that can help?
We decided to find out about AI. We recently asked ChatGPT, OpenAI’s chatbot, to play the role of a world-class reading teacher. And then we asked what it would do to help a 1st grade reader who was stumped by a particular word. We gave ChatGPT a few examples of words that might appear in a book for 1st graders: “night,” “name,” “bike,” and “hamburger.” There has been no shortage of reports and research on the importance of teaching young children how the letters in a word map onto the sounds in a word—also known as phonics. So we thought ChatGPT might be good at this task that is so critical to early reading.
We were wrong. ChatGPT had some knowledge of phonics and instructional routines, but it also made egregious errors that no good teacher or tutor would make. It mentioned calling the child’s attention to the sound of the letter “g” in “night.” It suggested telling the child to think of words with a similar sound to the letter “a” in “name” and then gave as an example the word “apple.” ChatGPT also provided “insect” as an example of a word that has a sound similar to the letter “i” in “bike.” In describing the word “name,” the bot said that the word “does not follow regular phonetic patterns,” apparently unaware of the vowel-consonant-silent-e pattern. And ChatGPT suggested that “hamburger” might be a sight word that children should memorize from flashcards.
We are optimistic about artificial intelligence. And we agree with entrepreneur and educator Sal Khan, one of the leaders exploring AI tutoring, that we need to “fight like hell for the positive-use cases” of the technology. However, it’s clear from ChatGPT’s errors that AI is not yet ready to tutor early readers.
This means that AI is also not yet ready to address the very real and debilitating consequences of exiting 1st grade without learning to read and write simple words and sentences. Renowned education researcher Robert Slavin likened this failure to letting children fall off a cliff at the edge of a playground and stationing ambulances (in this case, remedial reading programs) at the bottom. How much better, he pointed out, to build a fence that prevented any children from falling in the first place.
No serious reading researcher advocates a “magic bullet” in teaching children everything they need to know about phonics, let alone all the critical reading skills. The recent calls by journalist Emily Hanford and others to improve the quality of the nation’s training for teachers on phonics should be heeded as a key component in preventing 1st grade reading failure. So should appeals by people such as writer Natalie Wexler and education professor Susan Neuman for more knowledge-building curricula, starting in pre-K, so that children don’t fall off cliffs of reading failure in later grades because of insufficient vocabulary, knowledge, syntax, and oral-language skills.
We asked ourselves what it would take for all screen-based devices to offer beginning and struggling readers a phonics superpower.
We can’t help wondering, though: What if everyone had a technology tool that was just magic enough to fence off the cliff for 1st grade children who aren’t reached by current school improvement efforts? Who fail early because of inadequate phonics support?
In August 2022, just before ChatGPT burst onto the national scene, we held a conference at the Florida Center for Reading Research at Florida State University to discuss this vision. We asked ourselves what it would take for all screen-based devices to offer beginning and struggling readers a phonics superpower: the ability to touch or click on any word on the screen in any app or website or email or game and immediately see and hear how the letters and sounds in a word go together.
Aren’t there apps and software packages that do this already? Only to a limited extent. Text-to-speech dictionaries can translate words into their phonetic components but often not the way that a teacher would do for words that are too long to sound out sound by sound. Our free app, Reading Machine, provides customized, teacher-inspired decoding support if children type in a word they can’t read but only for the 5,000 words in its dictionary. Other reading-software packages—usually not free—will often provide this support for the words in their walled garden of reading material.
But why doesn’t our current technology already have this support built in for anything that children—or struggling adults—want to read on screen? We concluded at our conference that one reason is that there is no free database that software and hardware makers can use to build this support into their devices and programs. Building that kind of database, informed by phonics and orthographic-mapping pedagogy, is a daunting task. But we argue that such an audacious project is warranted, and Big Tech should help. With AI, this vision may be more feasible now than ever.
So maybe instead of rushing to build AI tutors, we all need to take a step back and collaboratively use artificial intelligence to help us first create a shareable word-decoding database, one that can provide every beginning and struggling reader a shame-free way to touch or click on any word in any language on any screen to see and hear that word sounded out. Perhaps we could even use that database to teach AI tutors the phonics knowledge they don’t yet have.
If we succeed, will we finally have the kind of fence needed to eliminate casualties from the cliff of 1st grade reading failure, which is primarily determined by phonics skills? We don’t yet know. But to borrow Sal Khan’s words, we think it’s worth fighting like hell to find out.
