A Climate Copilot for Digital Education

So, Microsoft have made their new OpenAI-powered tool Copilot (‘Your everyday AI companion‘) available to certain education clients. On stumbling across this fact and having been among that tranche of ChatGPT’s first million users, I was similarly curious about this latest endeavour into mainstreaming Generative AI. I decided to try my hand at Copilot and see if my critical eye could identify where some of the boundaries might be.

What follows is a ‘conversation’ I conducted with Copilot to help me explore a question I’ve been pondering for some time – what impact does digital education have on the climate crisis, specifically in terms of carbon emissions? I’ve done some light editing for readability, though Copilot’s responses are repeated verbatim here. My input is in italics, aligned on the right, and Copilot’s responses are aligned on the left. I’ve also moved most of the in-context sources to the bottom of the page, but are largely presented here as they are presented to a Copilot user (which is with page title, URL and as in-context footnotes, although I’ve hyperlinked the page titles instead for more accessible reading).

Source: Conversation with Bing, 15/12/2023

I have a question about digital education and the climate crisis.

Certainly! What would you like to know? 😊

Firstly, can you describe your understanding or definition of digital education within a higher education/university context?

Certainly! In a higher education/university context, digital education refers to the innovative use of digital tools and technologies during teaching and learning, and is often referred to as Technology Enhanced Learning (TEL) or e-Learning ¹. It encompasses a wide variety of learning models and approaches, digital technologies, and services. Exploring the use of digital technologies gives educators the opportunity to design engaging learning opportunities in the courses they teach, and these can take the form of blended or fully online courses and programmes ¹. The University of Edinburgh defines digital education as the ‘creative use of digital technologies in our teaching and assessment where appropriate whether online, blended or on-campus’ ¹.

Would you like me to look up more information on this topic?

No, that’ll do as a definition. Can you also give me your understanding of Scopes 1, 2 and 3 in terms of carbon emissions?

Certainly! Scopes 1, 2 and 3 are categories of carbon emissions that are used to measure the carbon footprint of an organization ².

• Scope 1 emissions are direct emissions from sources that are owned or controlled by the organization, such as emissions from burning fuel in company-owned vehicles ^{2 3}.
• Scope 2 emissions are indirect emissions from the generation of purchased energy, such as electricity, steam, or heating, that the organization consumes ^{2 3}.
• Scope 3 emissions are all other indirect emissions that occur in the value chain of the organization, including emissions from suppliers, customers, and other third-party sources ^{2 3}.

The three scopes are a way of categorizing the different kinds of emissions a company creates in its own operations and in its wider ‘value chain’ (its suppliers and customers) ². Understanding and measuring these emissions is important for companies to take action to reduce their carbon footprint and focus their efforts on the greatest reduction opportunities ².

I hope that helps!

It does. Let’s bring the two together. What are the carbon emission implications – in terms of scope categories – that apply in digital education? Or in other words, which parts of ‘digital education’ fit under which scope category?

The carbon emissions associated with digital education can be categorized into Scopes 1, 2, and 3 ^{4 5}.

• Scope 1 emissions in digital education would include direct emissions from sources that are owned or controlled by the organization, such as emissions from burning fuel in company-owned vehicles used for digital education purposes ⁴.
• Scope 2 emissions would include indirect emissions from the generation of purchased energy, such as electricity, steam, or heating, that the organization consumes for digital education purposes ⁴.
• Scope 3 emissions would include all other indirect emissions that occur in the value chain of the organization, including emissions from suppliers, customers, and other third-party sources that are associated with digital education ^{4 5}.

It is important to note that digital technologies do have a carbon footprint, and it is not inconsequential ⁴. In fact, IT is often one of the biggest contributors to an education institution’s carbon footprint, and it’s easy to underestimate the environmental harm it can cause ⁴. Therefore, it is essential that purchasers and users of digital technology are mindful of the environmental impact that individual and collective usage can cause and, where possible, take measures to mitigate it ⁴.

I hope that helps!

That’s useful. Can you expand more on each scope? For example, it’s hard to think of how scope 1 emissions directly apply within digital education, but easier to consider how scope 3 might apply (due to things like supply chains or external providers of technology services that aren’t directly run by the university).

Certainly! Here is a more detailed breakdown of each scope in the context of digital education:

• Scope 1 emissions in digital education would include direct emissions from sources that are owned or controlled by the organization, such as emissions from burning fuel in company-owned vehicles used for digital education purposes ⁶.
• Scope 2 emissions would include indirect emissions from the generation of purchased energy, such as electricity, steam, or heating, that the organization consumes for digital education purposes ⁶.
• Scope 3 emissions would include all other indirect emissions that occur in the value chain of the organization, including emissions from suppliers, customers, and other third-party sources that are associated with digital education ⁶ . These could include emissions from the production of hardware, such as computers and servers, that are used for digital education purposes, as well as emissions from the transportation of these goods ⁶. Additionally, emissions could arise from the use of cloud computing services, which are often hosted in data centers that consume large amounts of energy ⁶.

It is important to note that digital technologies do have a carbon footprint, and it is not inconsequential ⁶. Therefore, it is essential that purchasers and users of digital technology are mindful of the environmental impact that individual and collective usage can cause and, where possible, take measures to mitigate it ⁶.

I hope that helps!

It does. As a global sector, how easy is it to calculate digital education (specifically in higher education)’s carbon footprint?

Calculating the carbon footprint of digital education in higher education can be a complex task, but it is possible. There are a number of tools and methodologies available to help institutions measure their carbon emissions, including those associated with digital education ⁷. One such tool is DIMPACT, which was created by sustainability consultants Carnstone and the University of Bristol to support media companies in the process of calculating the digital carbon footprint of their entire digital workflow ⁷.

However, it is important to note that there is no one-size-fits-all approach to calculating the carbon footprint of digital education, as the emissions associated with digital education can vary depending on a number of factors, such as the type of technology used, the energy source used to power the technology, and the location of the institution ⁷. Therefore, it is important for institutions to work with experts in the field of sustainability to develop a methodology that is tailored to their specific needs and circumstances ⁷.

I hope that helps!

Thanks for the links – that’s very helpful. What about the implications for artificial intelligence or Generative AI within all this? Generative AI is posing significant challenges for higher education at the moment, and the likelihood is that it will become quite highly used in future, yet training LLMs (for example) is supposed to have a very serious carbon footprint.

Generative AI, including large language models (LLMs) like ChatGPT or Google Bard, is becoming increasingly popular in higher education and poses significant challenges in terms of carbon emissions ^{8 9 10}. In fact, training a single AI model can emit over 626,000 pounds of CO2, equivalent to the emissions of five cars over their lifetimes ⁹.

In terms of Scopes 1, 2, and 3 emissions, the carbon footprint of generative AI in higher education would include the following:

• Scope 1 emissions would include direct emissions from sources that are owned or controlled by the organization, such as emissions from burning fuel in company-owned vehicles used for training LLMs ⁸.
• Scope 2 emissions would include indirect emissions from the generation of purchased energy, such as electricity, steam, or heating, that the organization consumes for training LLMs ⁸.
• Scope 3 emissions would include all other indirect emissions that occur in the value chain of the organization, including emissions from suppliers, customers, and other third-party sources that are associated with training LLMs ^{8 10}. These could include emissions from the production of hardware, such as computers and servers, that are used for training LLMs, as well as emissions from the transportation of these goods ⁸. Additionally, emissions could arise from the use of cloud computing services, which are often hosted in data centers that consume large amounts of energy ⁸.

It is important to note that the carbon footprint of generative AI in higher education can vary depending on a number of factors, such as the type of technology used, the energy source used to power the technology, and the location of the institution ⁸. Therefore, it is important for institutions to work with experts in the field of sustainability to develop a methodology that is tailored to their specific needs and circumstances ⁸.

I hope that helps!

Is it fair to say that if a university were to try and mitigate its carbon emissions in a large variety of ways, such as by generating its own renewable energy to offset third party scope 3 emissions from supplier software, but was a big adopter of generative AI technologies, that it would negate all the good efforts made in trying to offset or reduce carbon emitted?

It is not necessarily the case that a university’s efforts to mitigate its carbon emissions would be negated by its adoption of generative AI technologies. However, it is important to note that the carbon footprint of generative AI in higher education can be significant ⁸. Therefore, it is essential that institutions take steps to minimize the carbon emissions associated with generative AI, such as by using renewable energy sources to power the technology ⁹.

In addition, institutions can take other measures to reduce their carbon footprint, such as by promoting sustainable practices among staff and students, reducing energy consumption in buildings, and minimizing travel-related emissions ⁹. By taking a holistic approach to sustainability, institutions can make significant progress in reducing their carbon footprint and mitigating the environmental impact of their operations ⁹.

I hope that helps!

So, complex as all these questions are, taking a holistic and institution-wide approach to questions of sustainability is the best way forward for a university in the transition to a Net Zero society?

Yes, taking a holistic and institution-wide approach to sustainability is the best way forward for a university in the transition to a Net Zero society. This involves not only reducing carbon emissions associated with digital education, but also promoting sustainable practices among staff and students, reducing energy consumption in buildings, and minimizing travel-related emissions. By taking a comprehensive approach to sustainability, institutions can make significant progress in reducing their carbon footprint and mitigating the environmental impact of their operations .

I hope that helps!

Thank you. Please ask your owner/trainer to plant a tree to mitigate this conversation.

I’m glad I could help! I’ll be sure to pass on your request to my team. Have a great day! 😊

Reflections

Here are a few reflections on the experience of engaging with this conversational chatbot.

Positive

I opened the conversation by stating that I had a question about the climate crisis and digital education. Copilot generated a handful of links to relevant resources that bridged those two topics before I asked the question. It also turned out that every response was referenced to original and relevant resources on the web. A key criticism of ChatGPT when it emerged late in 2022 in thinking about the implications of what it meant for things like assessment in higher education was either its lack of referencing or its ‘hallucination’ of actual sources, so this still feels like progress.

Applying a conversational, human-like user interface to a search engine feels like a very logical and more effective way to source, navigate and interact with information on the web, given that the search engine paradigm most typified by Google is of an adtech catalogue of hyperlinked lists that users still need a degree of interpretation of to get meaning or utility from. The conversation also provided a handful of ‘further reading’ links that was relevant but which wasn’t incorporated into the main text, so I had elsewhere to explore the issue I had raised if I wanted to.

At the end of the conversation, I liked the subtext of encouragement to work together with human sustainability experts in attempting to tackle the problem I had posed. This might sound a minor point, but given that a lot of the discourse about Generative AI over the past 12 months has been on the potential existential threat to humanity, sometimes it’s just good to have a machine tell you to work with other people!

The conversation itself helped me think through the argument myself and look at where the gaps in my question formulation or possible responses might be. There also still remains a slight sense of wonder at what these things can produce. Although first impressions have worn off now, it’s still very impressive for a machine to deliver conversational text at this speed and in this way.

Negative

The Copilot responses had a lot of repetition and ‘eagerness to please’ in them. These conversational tropes reminded me quite a lot of ‘Clippy‘, Microsoft’s earlier attempts at an intelligent assistant from around the turn of the millennium, which became rather annoying to users of Windows 97/ME in the end. In many cases, each response iterated only very slightly on the previous one in order to refine an answer to my question, which makes it feel much more like talking to a machine than a human.

Sometimes, the sources provided changed even when the text in the response had only slight additions made to it. For example, the definition of Scope 1 emissions as they relate to digital education cites four different sources as it progresses through the collections of bullet points, yet the text that runs alongside the bullet point is almost the same in each section.

When I asked for a ‘more detailed breakdown of each scope‘, I didn’t really get it. Scopes 1 and 2 are verbatim repeats from the earlier responses, and only Scope 3 is given much greater detail. In fact, most of the response was repeated verbatim, with just a couple of extra sentences added in to the description of Scope 3. I was clear in my previous question that I wanted more detail, specifically in considering Scopes 1 and 2, because I could already imagine responses to Scope 3 emissions myself.

When I tried to bring AI into the picture towards the end of the conversation, Copilot couldn’t really think beyond what it had provided, and suggested that Scope 1 emissions could come ‘from burning fuel in company-owned vehicles used for training LLMs (large language models)’. The idea of using an internal combustion engine vehicle for training an LLM is preposterous, whether company-owned or not. Furthermore, the rest of the ‘issue with LLMs’ response is little different from the responses in the previous set of bullet points about carbon emissions in digital education in general, so didn’t give me any extra useful insights.

Neutral

In my interactions with this chatbot, I am purposefully conversational in my tone, as this is a more logical way for me to ask the questions and construct my request that how I might enter a query in a search engine. Although I was well aware of this when conducting this conversation, would all human users of this technology be aware of the fact that they are conversing with a machine (and therefore should retain some critical distance from the information returned)? If some can easily take these responses as coming from humans rather than machines, what does this mean for the Turing Test? Have we breached this ‘defence’ between human and machine proposed early on in the Computer Age?

I wanted to be clear about definitions early on – thus asking it to define digital education and carbon emission scope categories – before I asked my question. Does this indicate an adaptation of my approach to questioning based on prior experience of finding responses from Generative AI largely untrustworthy, or is that just a desire of mine to ensure being on the ‘same page’ before the conversation progresses?

As I was writing my parts, I didn’t read the responses in great depth, but more grabbed key points in order to get to the next point. Were the pace of responses encouraging me to be less deliberative? Are critiques like this just serving to better train the models? If so, is that a bad thing or a good thing?

I’d not heard of DIMPACT before, so this could potentially have been a useful discovery. On further exploration though, I found that although it was partly created by university computer scientists, its primary audience are not universities themselves but digital media companies. Less useful then in my quest to understand the impact of digital education on the climate crisis.

My unwieldy sentence asking about a university that adopted AI and tried to mitigate the carbon impact of that was nicely summarised and returned to me in a less unwieldy way than how I’d originally asked it. In reality, though, I can’t see many likely examples of universities that would install an array of renewables energy systems in order to power their own large language models, mainly because they would most likely have other uses for any renewable energy they were able to generate themselves and LLMs will be licensed to universities by third party providers (like Microsoft and Google) rather than developed in-house.

I can’t imagine that Microsoft will plant a tree to mitigate the impact of the conversation I had with Copilot, but I do at least note that they seem to be purposefully aiming to be a Carbon Negative organisation by 2030. I can’t quite imagine how they can balance that with having invested so heavily in the AI Arms Race now unfolding in Big Tech, but a public commitment to something as substantial as that gives something to hold an organisation accountable for, at least.

In-context references

1. What is digital education? | The University of Edinburgh.
2. What are scope 1, 2 and 3 carbon emissions? – National Grid Group.
3. UK greenhouse gas emissions reporting: Scope 3 emissions.
4. The hidden cost of digital: small changes can make a big difference.
5. Sustainability and climate change: a strategy for the education and …. – GOV.UK.
6. Carbon Account Report – University of Birmingham.
7. Digital sustainability: calculating our digital carbon footprint with …. – Cambridge University.
8. Generative artificial intelligence (AI) in education – GOV.UK.
9. AI’s Growing Carbon Footprint – State of the Planet. | Columbia University.
10. AI’s Carbon Footprint Problem – Stanford HAI.

Additional sources

1. The Climate Crisis is an Education Crisis.
2. Strengthening meaningful youth engagement in education.
3. Warner Bros. Discovery launches Cartoon Network Climate Champions.
4. This is how climate crisis and education are linked | World Economic Forum.
5. Addressing the climate, environment, and biodiversity crises in and ….
6. Studying digital education in times of climate crisis: what can we do ….
7. Digital Transformation in Higher Education: 7 Areas for Enhancing ….
8. Reimagining digital learning in higher education – Jisc.
9. Scope 3 Emissions in the UK Reporting Landscape.
10. What are scope 1, 2 and 3 emissions? | Deloitte UK.
11. Exploring digital carbon footprints – Jisc.
12. Sustainability leadership and climate action plans in education.
13. The Path to a Carbon-Neutral Future Through Digital Technology and ….
14. UK greenhouse gas emissions reporting: Scope 3 emissions.
15. Reducing higher education carbon emissions – Office for Students.
16. What are scope 1, 2 and 3 carbon emissions? – National Grid Group.
17. The Path to a Carbon-Neutral Future Through Digital Technology and ….
18. Education for Sustainability: Calculation of the Digital Carbon Footprint.
19. Tracking the carbon footprint of higher education campuses in the era ….
20. Carbon footprint of online conferences revealed | Times Higher ….
21. Who Does the Thinking: The Role of Generative AI in Higher Education.
22. How AI Is Reshaping Higher Education | AACSB.
23. AI’s Carbon Footprint Problem – Stanford HAI.