At a time when artificial intelligence (AI) systems and tools based on large language models (LLMs) are being rapidly introduced into industries and daily life, the basic definition of ‘thinking’ and the essential skills we teach the next generation are being called into question.

In this interview, Dr Shuchi Grover, a leading voice in computing education who has recently become our Director of Research and Impact, shares how her work in computational thinking is evolving.

Can you share the story of your path in computer science (CS) education?

Most people in the education and CS education world know me from my research in computational thinking and K–12 CS education over the last 15 years. What is less known, perhaps, is that I started my career as a software engineer after completing my undergraduate and graduate studies in CS. About 25 years ago, I made a concerted shift to education, completing a Masters in Education from Harvard University in 2003, and then after a gap earning a PhD in the learning sciences (with a focus on K–12 CS education) from Stanford University in 2014.

Over these last two and a half decades, I have trained my efforts on helping young learners and school-aged children develop 21st-century competencies in computer science, data science, AI, and cybersecurity; as well as on STEM and non-STEM learning experiences that integrate computational thinking, AI, CS, and data science. My research has also attended to promoting interest and a sense of belonging in CS among learners from historically underrepresented groups.

I recently joined the Raspberry Pi Foundation as Director of Research and Impact. I feel very fortunate, as this role builds on all the work I have done over the course of my professional life and also affords me an unparalleled opportunity on a global scale to continue this work I’ve been so passionate about in both formal and non-formal learning settings.

You are well-known for your work on computational thinking. Since the development of LLMs, how has the definition of ‘thinking’ been changing?

This question is deep and thorny, and I’m not sure we have a complete answer to it yet. I believe that thinking as a human endeavour continues to be valid and means what it always has meant: a cognitive process that involves making new connections and creating meaning. In the education literature, thinking is often equated to problem solving. So teaching students ‘thinking skills’ has meant teaching them logic and ways to solve problems — typically in the context of a domain. In the context of K–12 CS education, computational thinking essentially means computational problem solving.

What changes with LLMs is not the definition of thinking itself, but rather what thinking skills students need most urgently. For students, the idea of ‘critical thinking’ has become much more critical (no pun intended) in an era when LLM-based tools offer quick and easy ways to produce answers. Students need to be equipped with the skills to evaluate AI outputs, and to follow up in deliberate and mindful ways to ensure that the AI-generated answer they ultimately take away is factually accurate, unbiased (to the extent that it can be), and valid for their context. They should also have the ability to recognise when an output is not suitable for their purposes, and when they would be better off approaching a problem or project as they would have in the pre-LLM era. These kinds of metacognition and evaluation skills must be crucial elements of AI literacy training.

How has data changed AI, and how has it impacted CS education?

Over the past 5 to 10 years, the scope, pervasiveness, and complexity of computing applications have grown substantially. This growth has been propelled by developments in AI and machine learning (ML). Many of the ML methods that underpin these developments have been in existence for much longer, but two key ingredients were still needed: large quantities of data, and the requisite computational power to process those quantities of data efficiently. Around 10 years ago, these became a reality. Combining so-called ‘big data’ captured from the countless human activities on the World Wide Web with new, powerful graphics processing units (GPUs) enabled AI scientists to build powerful prediction, classification and, most recently, generative AI models. Thus these scientists ushered in a new paradigm of computing that is data-driven. 

This has expanded the scope of what we need to teach students as part of CS education. In the context of AI and ML, you now have traditional programs that follow the algorithmic, deterministic paradigm of programming, but also ML applications that follow a data-driven, non-deterministic/probabilistic paradigm. CS curricula must help students develop an understanding of both. And data and data science are the crucial connective tissue between CS and AI/ML, so data literacy (which also captures elements of data agency and data equity) is critical to CS and AI learning experiences. 

Ethical issues in the context of data and AI have become more heightened and pertinent: issues of data privacy, safety, bias, responsible and explainable AI, and most importantly, impacts of AI systems on society. Understanding of these issues — what we can call ‘sociotechnical literacy’ — needs to be much more central to CS education now.

Considering the advances in AI and LLMs, what computing-related skills that we are used to teaching as part of CS are still relevant for young learners?

Let me begin by saying that there is no AI without CS. So understanding CS is important and foundational even in this age of AI and LLMs. The rationale for teaching CS and coding to learners aged 5 to 18 has always been primarily about (a) preparing the next generation to understand, and thrive in, a world where countless aspects of day-to-day life are driven by computing, and (b) providing them with the tools and skills for problem solving and creative expression. That goal has not changed. Foundational coding skills are still important and relevant for learners.

However, there is the new reality we must contend with: it is now easy to produce accurate code using LLM-based tools. We need good research on what this means in terms of how we teach coding. There are many questions related to this issue for which we need empirical evidence: What are the foundational skills for programming effectively with AI tools? What CS topics, skills, and concepts must we emphasise or de-emphasise? Could teachers be supported by generative AI tools in teaching coding, and if so, how? Will use of AI tools result in poor learning for students? How might students leverage LLM tools in ways that don’t harm their foundational understanding of coding concepts, and at what age and stage? What kinds of LLM tools are safe and suitable, and what preparation must students have before they use them? What bigger, more sophisticated projects might students create with the help of an LLM tool? How might LLM tools aid student learning through formative feedback? Can LLM tools aid in metacognition by prompting reflection at the right moments in a project? These are just some of the many, many questions we need to answer to shape CS education over the coming years.

---

A version of this interview also appears in issue 29 of Hello World, available as a free download. Subscribe to the magazine to never miss an upcoming issue.