3 Educational Challenges for Personalized AI Tutors

Generative AI has taken all industries by storm, including educational technology (EdTech). Generative AI is a game-changer for children’s EdTech because of its potential to provide uniquely personalized learning experiences. Unlike previous technologies that adjusted learning based on an average model of a child, generative AI trains its models on personal data directly contributed through individual children's interactions. EdTech developers predict that such precise and dynamic personalization will revolutionise kids’ learning in various forms, particularly in relation to personalized AI tutors.

Personalized learning with AI

In a recent TED Talk, Sal Khan, founder of Khan Academy, outlined his vision for transforming education with AI-driven personalized tutors. The academy already offers personalized learning in that students can choose content and learn at their own pace. With generative AI, the guidance students receive can be personalized in the form of AI tutors asking questions and scaffolding students’ answers in the style of Socratic Dialogue. Imagine avatars posing thought-provoking questions in children’s own language and at a pace and level adjusted to the child’s needs or preferences.

It is hard to tell whether such personalized AI tutoring will lead to better learning and better citizens. To start, academic research is not keeping up with AI innovation. Collaborations between EdTech industry and academia are often lacking and, in the case of generative AI, industry's leap ahead of academia was particularly evident. At this stage, as confirmed in a systematic review of studies, there is not enough evidence to conclude that even the use of chatbots is beneficial for language learning.

However, to anticipate effects in plausible scenarios with personalized AI tutors, we can reflect on these findings and consider the extent to which they might apply in different contexts. Current research concerned with speech technologies and conversational agents trained to interpret human language provides some pointers for possible effects.

Conversational agents. Studies on children’s interactions with digital interfaces – images, TV shows, or books –scaffolded by digital tutors typically compare the difference or added value of “real” and digital tutors. Those that adjust responses to the child — for example, in the form of customized responses of cartoon characters — show some promising learning benefits. For example, when 3-to-6-year-olds read stories with an automated conversational agent (digital tutor), they understood the story as much as they did when an adult read the story to them. What was different was the nature of children’s responses: When they read with the digital tutor, they attempted to answer clearly to be understood by the automatic speech processing. But when they read with the adult, their answers were more diverse and relevant.

One might anticipate that children’s answers will get more aligned with natural conversations as the technologies’ usability advances. What remains a challenge is addressing the pitfalls of digital personalized education. Two challenges are particularly salient: children’s agency and the personalization-pluralization balance.

Children’s agency with personalized AI tutors. Children’s agency, the possibility to control and make a choice, is often neglected with adult-made technologies. Paradoxically, although personalized learning is sold under the banner of child-centred education, automated recommendations and learning adjustments happen automatically, with direct use of children’s data but minimal children’s agency.

Generative AI was not developed for children but it is increasingly applied to education contexts. The question of whether the benefits really flow back to the child is especially salient with personalized AI tutors developed by commercial companies. With generative AI EdTech, children’s speech is used to train models embedded in tools that personalize their interactions without children being able to actively give informed consent.

While we wait for these ethical challenges to be addressed by AI bills and their enforced implementation, it is worth remembering the promise and pitfalls of personalised learning with new EdTech. The promises have been highlighted in terms of learning motivation, engagement, and meeting children where they are. The pitfalls relate to the use of personalization as a sole means to boost children's education and its digital design.

Education Essential Reads

The Bad News About Teen Sex Education Reflects Good Trends

The Rise of Homeschooling: The Confessions of an Educator

Personalized persuasive design is one of the manipulative features used in digital media to keep children's attention on the screen for as long as possible. With such a design, meaningless games or questions pop up on the screen but because they are adjusted to what users engaged with before, they catch their attention. While personalized nudges in the form of recommendation algorithms are unique to digital personalization, the broader problem of personalization is its connection to pluralization, or diversity of experience.

Will personalized AI tutors teach children to help and serve those whom they would not typically think of as their ‘friends’? Or to slow down the pace of learning to practice patience, and introduce difficult and at times frustrating tasks to practice executive functioning? Or to challenge children’s thinking with content that is not personalized and different from what they liked and engaged with before? Socrates would ask whether children will be pushed toward particulars rather than universals. The more a design leans towards the pole of personalization, the more it removes children from a collective reality — a reality that is challenging and complex because of its diversity, not its homogeneity.