You can hear the child reading as you walk down the corridor. The child stops occasionally and discusses an unknown word or a mispronunciation with the teacher, and then carries on with the story.

Everything seems to be going well, so you glance into the classroom as you pass, rather than popping in and interrupting. But as you do, you stop in confusion: the child is alone at the desk by the door. The teacher is working with the small groups scattered around the room. 

On closer inspection, the “person” the child seems to be reading to - and discussing pronunciation with - is not actually a person at all: it’s a tablet computer running AI software that seems not just to be “hearing” what the child is saying but understanding and providing feedback, too.

Is it too much of a stretch to see AI being used in this way - as a primary tool for teaching reading - in schools? While AI-powered virtual assistant products do a decent job of interpreting human speech for menial tasks, they are yet to reach a point of 100 per cent accuracy of interpretation, so just understanding a child’s speech amid
a busy classroom - let alone being able to interpret it and provide feedback - would be, you would think, beyond it. 

And yet low-level reading aid products powered with AI are already available, and a new breed of products are being introduced to the market that attempt to be that main cog in the reading education wheel. So, are we closer to the scenario above than we may like to think?

First, let’s be clear: teaching children to read is the most important step in education, and yet it is fiendishly difficult to get right. That humans have managed to work out a process to ensure that the vast majority of pupils hit key stage 2 able to read is something to be celebrated, and many teachers will be wary of risking that success by handing over the process to a machine.

However, a number of products that have emerged over the past few years are part of a concerted effort to allay those fears. Powered by AI, and with well-optimised user experiences and detailed data feedback tools, these products are aimed at being a key part of reading instruction. And during the pandemic, they have found a window of opportunity: children working remotely and social distancing restrictions make teaching reading digitally an appealing concept. 

One such product is Fonetti, the result of a collaboration between the University of Edinburgh’s School of Informatics and the company Auris Tech. It’s an app recommended by the Department for Education as part
of its Hungry Little Minds initiative. In layman’s terms, the technology listens to what someone is saying and then compares it with what someone was expected to say using machine-learning processes.

“A child selects a book and as they read [out loud], the text colour [on the screen] changes to green [if they say the word correctly],” explains Auris Tech’s chief commercial officer, Colin Tankard. “If they get stuck on
a word that they don’t really understand, they can double-tap the word and it says out loud to them what the word is. If they get a word wrong, it doesn’t change colour [to green], but importantly, we don’t stop the child reading at that point and say, ‘You got that word wrong and you can’t move forward until you get the word right.’

“We allow them to carry on reading because it’s really important for the child’s psychology that they keep progressing and moving on, and there are no barriers stopping them. But when they get to the end of the book, they’re told, ‘You got a few words wrong. Why not go back and try again?’”

Children are rewarded for their performance with fun animations, fanfares and badges to further encourage them. Meanwhile, the AI crunches the data and gives teachers information on how the child is performing and where they stand compared with other children. “[It can show a teacher] how that person is doing in relation to maybe other children of the same age in the same class and what sort of books they should be moving on to,” says Tankard.

He argues that it’s not just within the current Covid restrictions where the tech is useful; he believes that it is something that can play a big role in education in more normal times. “Some children struggle with reading and sometimes you find that if their reading isn’t very good, they feel very embarrassed when they’re reading out loud, and so they almost want a safe place to read. But they also need somebody to help them out and to tell them that they’re getting a word wrong,” Tankard explains. 

Another speech tool developed for children that relies on AI technology is Buddy, a “virtual English tutor” that allows pupils to practise their vocabulary and pronunciation. Dmitry Stavisky, co-founder and chief operating officer at MyBuddy.ai, creator of the Buddy app, concedes that one of the major challenges that developers of speech-tech tools for children face is that children’s speech is very different from that of adults. 

“Their vocal apparatus is still developing, [and] the pitch and pattern of children’s voices are very different from adults’,” says Stavisky. “A kid’s voice varies more than
one of an adult and changes over time as the kid grows up. Moreover, the challenge is not limited to the acoustic models of kids’ speech - it’s less regular, presenting additional difficulty for both automatic speech recognition and natural language-understanding algorithms.”

To get around this problem, MyBuddy.ai had to develop a tech solution optimised for children’s voices. Having now done that, Stavisky also thinks AI is the way forward for teaching reading. 

“When you have a computer system that can ‘understand’ such speech, you can take the next step and build it up to teach children proper pronunciation, [and] make their speech more intelligible,” he says. “This requires a lot of practice, which is hard and expensive to provide with human tutors. Therefore, it’s important to build AI-based tools that can do that affordably and conveniently.”

Human nature

Of course, when you are selling such products, you want to talk up their potential. The reality, according to many teachers, is that these products are currently only likely to be used for particular contexts and at particular times. The variances around what a child needs when they are learning to read, and the complexity around when and how errors are made, often requires a holistic view of the child, not a diagnostic one, says Rhiannon Dowson, a literacy lead in a primary school in the South of England. 

That’s not to say that AI won’t develop to fulfil that role, just that it isn’t doing that - or is not doing that as well as a human - yet, in her view. So where could AI be of use? 

Dowson thinks that it could work for intervention tasks or extension exercises with carefully chosen pupils, enabling teachers to “reach” more children than they would be able to otherwise. However, she stresses that this would still involve the teacher being the predominant influence on the child’s reading tuition. 

Some language experts think that assessment and evaluation, where reading difficulties arise, may be areas where AI tools could really come into their own in the future, too. Courtenay Norbury, professor of developmental language and communication disorders at University College London, explains that when you are trying to diagnose a potential language difficulty, you get the child to tell or retell a story.

“We know that’s something that kids who have problems with language find really challenging,” she says. “But it’s a really time-consuming task to do because you get them to tell you a story, then you have to transcribe that story, then you are looking to code it for the kinds of language that they use and the information that they convey. So even though it’s a really useful little task to do, people don’t do it as much because it’s pretty labour-intensive. And so there’s definitely interest [in the use of AI for this type of task].”

Jessie Ricketts, who runs the Language and Reading Acquisition Lab at Royal Holloway, University of London, also thinks AI could potentially perform this type of assessment function and “then feed into human judgement”. 

However, she says “there is a question of whether children will behave ‘normally’ with an AI system and therefore give a true reflection of their ability”.

There are other potential issues that could negatively impact assessments if AI were used, too, says Kathleen McCarthy, lecturer in linguistics and director of the QMUL Language Acquisition Lab at Queen Mary University of London.

“Speech recognition programmes have been shown to have biases, and not be great at accounting for regional varieties/accented speech, which I guess could potentially limit its use in diverse school settings, where there are potentially a range of accents, that, importantly, have nothing to do with an underlying speech and language difficulty,” she explains.

McCarthy is also not sure how precise AI could be in identifying some aspects of “clinical” speech production. “For example, it might be good at recognising that some aspect of the speech sequence is atypical - ie, correct versus incorrect - but it might not be able to identify very atypical speech sound sequences - ie, the exact sound sequences produced - and possibly small differences in the articulation of a speech sound,” she says.

“This is something that could be picked up in a face-to-face assessment with a speech and language therapist, who would see the child’s face and use a combination of visual and auditory cues - and, crucially, their knowledge of clinical phonetics - to assess the child.”

That said, McCarthy and some of her speech and language therapy collaborators believe AI technology could still potentially play a useful role in assessments, especially for “more general assessments that do not require a fine-grained analysis of the child’s speech and communication - for example, a vocabulary assessment, possibly the initial screens. This would free up more of their time to focus on the aspects of assessment and intervention that require face-to-face [interactions]”.

Rise of the machines

The important thing to note here is that the development of AI technology is accelerating fast: what today’s systems are capable of would have been unthinkable five years ago within this timescale. Also, research into the application of the technology is obviously going to lag behind the development of the technology itself. 

Thankfully, though, that research is starting to happen. In the UK, a two-year study, financed by the National Institute for Health Research, intends to gather the recordings of 600 children reading stories and then use these sound clips to help build an automated language sample transcription and analysis software platform that can be used by speech and language therapists as a decision-making support tool.

At present, it can take speech and language therapists as long as 30 minutes within a session and up to 90 minutes after a session to manually transcribe and analyse a child’s language. This is where it is hoped that the new automated assessment tool will make a difference. The tool could radically speed up this process.

Industry experts say more research around the technology’s use for phonics is also under way, with potentially huge time savings for EYFS and KS1 teachers in the pipeline. 

As useful as these tools may eventually be, though, is it likely that artificial intelligence will ever fully take the job of teaching reading away from a teacher? Norbury says that, regardless of how intelligent AI systems get, they “can’t replace good teaching practice”. She thinks that you can’t get rid of the interpersonal interaction between teachers and students, because this is where language really prospers. 

Simon Creasey is a freelance journalist

This article originally appeared in the 1 January 2021 issue under the headline “Can we trust the robots to teach children to read?”