VIPKID: Machine Learning in Online Education

How should the new leader in online English education power human learning with machine learning?

Over the past two hundred years, the human society has been transformed through three industrial revolutions (steam, electricity, chips), yet the way we educate our workforce has remained largely unchanged. With the fourth revolution brought by the advancement in machine learning, the education industry in China is racing to catch the opportunities. VIPKID,  “Uber for online English teaching”, is a marketplace that allows native English speakers to teach English online to children in China on a one-on-one basis. VIPKID is setting the application of machine learning as a major strategy pillar.


Through providing a standardized curriculum for any native speaker to be able to teach, VIPKID has reached over $1 billion annual revenue with its commoditized English lessons. However, in order to retain students for the long run, the company needs to maximize the efficacy of teaching by incorporating personalized learning to tailor to each student’s needs and preferences. This could be aided by the use of deep learning. Meanwhile, VIPKID is a sales and services-heavy business with half of its labor dedicated to customer services facing students’ parents. With the advancement of Natural Language Processing and Natural Language Understanding, these labor-intensive services and sales tasks could be significantly automated to streamline operations.

Looking across the industry, a competitive service Liulishuo, the AI English Tutor mobile app, has recently completed an $72 million IPO at NYSE [1]. Although Liulishuo’s pure AI tutor is still no replacement for human teachers, VIPKID needs to invest into the forefront of AI technology to stay ahead of the curve.

Machine Learning in Action

To date, VIPKID has taken a series of initiatives aiming to incorporate machine learning technologies into its products. The company established the VIPKID Education Research Institute in partnership with Stanford University to conduct machine learning research in areas such as language ability assessment for student placement [2]. August 2018, the company also partnered with Microsoft to leverage its AI technology to improve the learning experience [3].

As VIPKID records all of its online sessions between students and teachers, with the help of Microsoft, it can analyze the 10 million+ minutes of video sessions, looking specifically at students’ facial reactions to the materials they learn. As interactivity is crucial in online education and each student has distinct ways of expressing feelings, VIPKID has developed a complex deep neural network to analyze students’ eye movements to assess their engagement. With this, the company can further improve the matching algorithm to pair students with teachers of best fit on a 1-1 setting.  In addition, VIPKID rolled out Chatterbox, a new AI product that listens to students’ pronunciation, scores them, and correct them, enabling students to practice speaking after class [4].

Online classroom showing a student’s facial analysis. Numbers on bottom right indicate ratings across various engagement metrics.


As the various algorithms become open source and GPU costs eventually coming down, in order to gain an edge in the machine learning era, proprietary data is key. Today, as VIPKID rolls out more education products (lessons for other formats, languages, subjects), it needs to record all students’ learning journeys, so it can possess the most comprehensive learning profile for each child and can derive deep customer insights to offer the most relevant services as these children grow up. On the other hand, to create an operationally learn corporation and support rapid growth without multiplying its sales and services staff, VIPKID needs to adopt AI customer service tools to automate substantial parts of its service responses; it should also deploy sales call analytics tools powered by machine learning to optimize the effectiveness of customer acquisition efforts. Meanwhile, to complement its technology capabilities, VIPKID should formulate the strategy of its corporate development team to focus on adding machine learning capabilities in learning context by acquiring valuable startups.

Looking Ahead

With the accumulation of good-quality interaction data between students and teachers, it is possible that in the foreseeable future VIPKID develops a pure AI teacher who can deliver the curriculum tailored to the students’ learning styles and real-time engagement. The product might not encompass the same level of care, humor, and empathy as a real life teacher, but perhaps VIPKID can offer the AI version at a fraction of the cost, making these lessons affordable to children in the poor rural areas of emerging economies. This would, however, stir controversies around how we educate children — Should students be taught by machines? How would that affect the growth of these young minds psychologically?


(Word Count: 736)



[1] Mars Woo, “China’s English learning app Liulishuo raises $72m in US IPO”. Deal Street Asia. September 28, 2018. Accessed November 11, 2018.

[2] “The internet boom in foreigners teaching China’s children online”. South China Morning Post. September 7, 2017.  Accessed November 11, 2018.

[3] “VIPKID Launches V+ Strategy: Redefining Online English Education Standards”. VIPKID Official Website. August 2, 2018.  Accessed November 12, 2018.

[4] “AI Tech, why is VIPKID ahead?” Hubei English Online. July 20th, 2017. Accessed November 12, 2018.



Medtronic – missing out on a personalized medicine opportunity, or appropriately cautious?


JP Morgan Chase & Machine Learning

2 thoughts on “VIPKID: Machine Learning in Online Education

  1. Very interesting article, I have seen the growth of VIPKid and have been astounded by its growth. I remember the CEO saying, “If the US won’t pay its teachers well, the we will”, which I found extremely powerful in leveraging technology to democratize language learning and education. I think that VIPKid has shown how strong and defensible of a company it is by providing a very customized experience and personal experience to ensure the success of its students.

    To answer your question, I think in the near future, basic level education will all be taught with Artificial Intelligence, with limited human oversight. Basic concepts in language, math, science are the same across the realm and with advancements in AI, ML, and language processing, it would be possible to almost identically replicate the student teacher relationship outside of the physical human aspect. I think there will be a lot of backlash initially about this method of teaching as we have seen historically with the advancement of even Massive Open Online Courses (MOOCs) which don’t incorporate a physical teacher. However, I think this will become normalized in the future as long as these companies and schools are able to create the right incentives to use these courses. This biggest challenge these online courses face is not necessarily hurdles in the technology or material, but its the hurdle to get users motivated and disciplined enough to take them. If they become required for grade advancement, this would be a sufficient requirement in my opinion.

  2. Coming from another Education Tech start-up prior to HBS (more focused on K-12 though), I loved reading about VIPKID and its mission to revolutionize learning. My start-up similarly had a sales/services-heavy business model, especially as we scaled. While I’m fascinated by applying machine learning to personalized learning, I’m especially intrigued by the idea to use machine learning to help sales/customer success teams serve customers even better. I know there was a lot of manual, tedious work required by our staff in order to provide the best customer experience possible, and machine learning could help innovate ways of working within the company.

    To answer your question, I strongly believe machines/robots will never (and should never) replace teachers. The teacher’s role in a classroom is not just a content/knowledge deliverer; if this were the case, they would definitely be replaceable with AI! The best teachers we saw using our product were amazing at facilitating learning, challenging students’, and mentoring them using personal relationships. This could never be replicated by a machine as it requires empathy and a very human approach. However, I think this argument only holds for really exceptional teachers. For the ones who are just mediocre, or even bad-quality, I’m wondering if they can easily be replaced. I thought of a question off of your question: Is it possible that schools that are under-resourced and with lower-performing teachers (those don’t always happen simultaneously, but they can) will be more likely to get their teachers replaced with machines, while this doesn’t happen on the other end of the spectrum? While this revolution in machine learning in education has the potential to decrease inequities and even the playing field, I also worry that it can further the divide.

    Great article, and really thought-provoking topics. Great job!

Leave a comment