Building Quizizz AI: A Leap Forward in Educational Technology

Quizizz has been at the forefront of creating innovative tools for teachers worldwide. From supporting different question types – to meeting learner needs – to creating various advanced strategies to help teachers engage their students, Quizizz has been empowering teachers worldwide.

With OpenAI and other AI research labs creating advanced AI models like GPT-3.5, GPT-4, Claude, etc, there is a huge shift in what is possible through technology. The shift is similar to the creation of the internet and smartphones in many ways.

With these advancements, we started thinking about how to give our teachers these superpowers.

The Inception of Quizizz AI

The development of Quizizz AI was driven by a clear objective: to alleviate the administrative load of educators. Our focus was creating capabilities to simplify and speed up the content creation process, making teaching more efficient and effective. A small team of engineers, a PM, and a designer was formed to make it possible.

Quizizz has a rich library of 400M+ quizzes created and publicly shared by teachers. Most teachers on our platform extensively use the public library to host activities for their students. While the quizzes are great, teachers try to personalize them to increase engagement and differentiate them for students with different needs before they host an activity. We piloted ‘Enhance AI’ by providing teachers with a lot of enhancement capabilities that they could do to make the quiz more accessible and/or engaging. Teachers loved it!

From building a pilot to building a Co-pilot

Even though the teachers loved it, we didn’t want to stop there. We spent a lot of time talking to teachers and looking at our analytics data to answer these questions:

• Which of the existing capabilities are most useful and which aren’t?
• Which new capabilities should we add that would be helpful for the teachers?
• How can we improve our AI to make the most used capabilities better?

We spent a few weeks iterating on Enhance AI. Everyone on the team was talking/listening to our users, looking at data, and whiteboarding product and design ideas. After several iterations, we reached a stage where the product was beneficial.

With these changes, we were able to help teachers who make effective use of our public library. The Quizizz Co-pilot would be incomplete without the teachers who like to create content. We went back to the whiteboard to figure out the different ways teachers might want to create quizzes.

We figured out the following use cases:

• Generating quizzes for a given topic
• Generating quizzes from learning content (documents, slides, videos, web articles, etc.)
• Extracting quizzes from existing worksheets, test papers, etc

We quickly built a prototype for these use cases and started gathering teacher feedback. One of the tweets related to generating quizzes from YouTube went viral.

Based on how teachers were using the product and looking at the data, we figured out ways to improve the generated quizzes. We iterated on different AI experiences across the product in the past few months, including generating answer choices across different question types, adding an autocomplete experience, generating lessons using AI, a Chrome extension to let users create quizzes from whatever they read, etc.

We have been working hard on making Quizizz a co-pilot for the teachers. While many experiments have been successful, many failed, and we navigated many challenges in building Quizizz AI.

Biggest challenges in building Quizizz AI

Content Quality

The major challenge we have faced as part of building an AI-first experience has been the quality of the generated content. While the quality of the content generated using AI is something we would have never believed a year ago, we would not want to limit ourselves.

We are trying out many UX and technical experiments to improve the quality. One of our experiments was asking for clarifications to make the quizzes more specific to teacher expectations. This helped increase the acceptance rate of the generated quizzes.

Some technical experiments we are trying are moving to bigger models for quiz creation, fine-tuning the models to be better aligned with the best quizzes, using Retrieval Augmented Generation to enrich the prompts and other techniques, etc.

Latency

While AI helps reduce the time to create a quiz by more than 1/10th, in our initial iterations the time it used to take to generate a quiz was still enough to make the teacher not wait on the loader screen.

We figured out ways to make most of the processes fast enough to make it look instant without compromising quality. Though we successfully reduced overall latency, we could not move to bigger models as that would increase the latency by a lot.

One of the experiments that we are trying out where we can use bigger models and simultaneously reduce the “perceived” latency is by doing what ChatGPT does to solve perceived latency: “streaming.” We are also figuring out ways to reduce perceived latency in all the flows as we move to bigger models across different flows.

Conclusion

Quizizz AI demonstrates how targeted AI applications can significantly impact the education sector. As we continue to refine and expand its capabilities, our focus remains on delivering capabilities that are not just technologically advanced but can also deeply adapt to the needs of educators and students.