Ambar presses her hand to her forehead, nose crinkled in concentration as she considers the question on her screen: how many sevens in 91? The ten-year-old has been grappling with it for about a minute when she smiles: “13!”.
Her tutor responds by posting a large smiley cat picture on her screen – the virtual equivalent of a pat on the back. He is sitting on the other side of the world in an online tutoring centre in India.
Ambar, who attends Pakeman primary school in north London, is one of nearly 4,000 primary school children in Britain signed up for weekly one-to-one maths sessions with tutors based in India and Sri Lanka. The lessons, provided by a company called Third Space Learning, are targeted at pupils struggling with maths – particularly those from disadvantaged backgrounds.
Together with scientists at University College London (UCL), the company has analysed around 100,000 hours of audio and written data from its tutorials, with the goal of identifying what makes a good teacher and a successful lesson.
Tom Hooper, the company’s CEO, said: “We’re looking to optimise lessons based on the knowledge we gain. We’ve recorded every lesson that we’ve ever done. By using the data, we’ve been trying to introduce AI to augment the teaching”.
Initially, the company’s 300 tutors will receive real-time, automated interventions from the teaching software when it detects that a lesson may be veering off-course.
Pupils on the programme have a 45-minute session with the same tutor each week. They communicate through a headset and a shared “whiteboard” (they can’t see each other). The lessons at Pakeman school are tailored to the individual, including visual rewards linked to the child’s interests. Premier League strikers for nearby Arsenal, cute animals and pink, iced doughnuts flash up on the screens of Ambar’s classmates.
In addition to the raw audio data, each lesson has various success metrics attached: how many problems completed, how useful the pupil found the session, how the tutor rated it. Using machine learning algorithms to sift through the dataset, the UCL team has started to look for patterns.
An early analysis found, perhaps unsurprisingly, that when tutors speak too quickly, the pupil is more likely to lose interest. Leaving sufficient time for the child to respond or pose their own questions was also found to be a factor in the lesson’s success, according to Hooper. These observations are likely to form the basis of the initial prompts that the tutors will receive, probably in the form of messages flashing up on their screen.
“We’re going to be drip-feeding it in in relatively simple ways to start with,” said Hooper.
As the technology evolves, the interventions could become more sophisticated and the software might play a more active role in teaching, raising questions about the extent to which intelligent software could replace human teachers.
Rose Luckin, a professor of learner centred design at University College London, who is collaborating with Third Space Learning on the project, said: “What we are very interested in is the right blend of human and artificial intelligence in the classroom – identifying that sweet spot.”
According to Luckin, AI provides a unique opportunity to assess which teaching strategies are working and to individualise teaching.
“It would be able to say, for this child at the moment, Jolly Phonics is working well,” she said. “You would be able to look back over their reading process and see which interventions worked. The potential for the use of AI to make education tractable and visible is huge.”
However, she predicts that the insights gleaned from AI will often be applied by human teachers. “What I’m really concerned about is that people don’t run away with the idea that kids have to be plugged into the computer,” she said. “It’s about so much more than that.”
Hooper agreed that the aim is not to replace teachers with robots. “There’s a slightly dubious conversation about how AI will make humans irrelevant, but it’s not at all about replacing humans,” he said. “Our whole belief is that for children disengaged with the subject, who are lacking in confidence, people is what matter. An algorithm can’t provide that.”
He said he does not expect his tutors, most of whom are science graduates, will be concerned about the automated feedback. “We’ll need to be considerate about it,” he said, adding that it would not be “a bossy algorithm barking orders at people”.
Shazli Mahroof, 27, a tutor team leader based in Colombo, Sri Lanka, said he was not worried about being replaced by a teaching robot in the near future. “It’s not the computer who is going to teach,” he said.
The tutors already have one lesson each week assessed by supervisors, and it is fairly obvious, subjectively, when things are progressing well, according to Hooper.
“We’re asking ‘how do we promote those teaching events at scale?’” he said.
Companies entering this sphere also need to convince parents and teachers that the data being collected is both secure and will ultimately benefit pupils. A previous data analytics project in New York state schools, run by the company InBloom, collapsed in 2014 after becoming embroiled in privacy concerns.
“The whole thing became toxic,” said Luckin. “It’s really important that we do it right.”
At Pakeman primary, it is the last maths session before the Christmas holidays. An infant class in the school hall is rehearsing a performance and the school has an end-of-term feeling in the air, but the atmosphere in the online learning session is one of hushed focus.
After finishing the lesson, Ambar said that maths used to make her anxious, but since starting the weekly tutorials in Year 5, she has started enjoying it. “When they give you horrible sums, they help you,” she said. “I was scared to do it, but it was actually fun.”
guardian.co.uk © Guardian News and Media Limited 2010
Have something to tell us about this article?