No Halo, No Inferno

Finding the Right Balance for Sustainable AI

No Halo, No Inferno
Photo by Danist Soh / Unsplash

I am very much trying not to set myself up with a halo to strangle myself with. But I also don’t want to undermine what I’m doing by setting the planet on fire just to get there.

In setting up this blog and pages, I’ve been reading about AI and its environmental impact. Until I stopped to think about it, I hadn’t properly considered the carbon footprint, for example, of using AI - ChatGPT in particular. There’s a wealth of research on this, but since I write from a non-techy perspective for non-techy people, I want to break it down simply. This is just a starting point, raising my awareness (and maybe yours too).

😮‍💨
One day I will eventually get around to the posts I've been planning, rather than the posts about all the stuff I've been finding out on the way (promise).

The carbon cost of ChatGPT

Using AI - especially ChatGPT - has an environmental cost (duh). Every query requires energy, and that energy has a footprint. If I use ChatGPT with intentions of social justice, is the environmental cost greater than the potential good?

Each time I generate text, it taps into power-hungry data centres. Estimates suggest that processing 1,000 queries uses about 0.5 kWh of electricity. For a typical blog post, toolkit or template where I may make around 100 queries, that means:

  • Energy consumption: between 0.29 kWh and 0.9 kWh
  • Equivalent to: charging a smartphone 29 to 90 times
  • CO₂ emissions: between 0.145 kg CO₂ (145g) and 0.45 kg CO₂, which is about the same as driving 0.7 to 2.25 km in a petrol car.

By itself, that doesn't seem too bad does it. But scaled across millions of users, it adds up. And if I am encouraging people to use AI (for good!), then I'm just adding more to those emissions. AI isn't free in environmental terms, and ignoring that reality is the kind of tech-blind optimism I want to avoid. I needed to work out, in a reliable and hopefully thoughtful way, whether I could actually be making things worse just being here in the blogosphere.

A Doughnut Economics lens: a sustainability check

I am to be honest a little bit in love with Kate Raworth’s Doughnut Economics which I truly believe provides a powerful framework to apply to the whole world and it really is worth a read - anything is better than GDP surely! So, if it's good enough for Barcelona, then it's good enough for me. I asked ChatGPT to apply Kate's frameworks to my blog using her key principles:

  1. The Ecological Ceiling (avoiding overshoot of planetary limits like carbon emissions, biodiversity loss, and water stress)
  2. The Social Foundation (ensuring equitable access to resources, jobs, and well-being)

If AI can be a tool for positive change, then its use, and my use of it, should align with these principles - avoiding excess while maximising benefit.

The results: Doughnut Economics vs ChatGPT

The impact of each interaction with ChatGPT is tied primarily to energy consumption and carbon emissions from the data centers running AI models.

  1. Energy Use and Carbon Emissions (Ecological Ceiling)

Each ChatGPT response requires a fraction of a second of GPU processing, consuming electricity. Estimates suggest that a single ChatGPT query uses 2.9 to 9 Wh of electricity - ”comparable to running a 10W LED bulb for 15 - 50 minutes".

If powered by fossil fuels, a single query emits 0.2 to 0.8 grams of CO2 (roughly the same as sending a few emails).

Large-scale AI models (like GPT-4) have high training costs, requiring thousands of megawatt-hours. But once trained, each interaction has a much lower footprint.

  1. Resource Use and E-Waste

AI data centres require rare earth minerals for GPUs, contributing to mining impacts, water pollution, and geopolitical supply chain issues.

The hardware has a limited lifespan, generating e-waste when components need upgrading or replacing.

  1. Who Benefits? (Social Foundation)

AI could democratise knowledge by making advanced insights freely available, but the benefits and costs are unevenly distribute. Data centres are mostly in high-income countries, but climate risks are borne disproportionately by lower-income nations.

AI-related job displacement raises concerns about worker livelihoods, though it can also create new opportunities.

A Doughnut Economics Perspective on AI then concludes:

Right now, AI is likely exceeding planetary boundaries in carbon emissions and resource use!

Oh dear, it isn't sounding particularly viable for me to continue so far. Can Doughnut Economics save my blog? I *may* have asked ChatGPT to help with this bit too:

Chat GPT: To assess sustainability, we need to compare costs (AI energy use + emissions) against benefits (blog reach + impact on social justice movements).

The Results: Doughnut Economics Vs Chat Will Eat Itself

  1. Energy Use and Carbon Emissions (Ecological Ceiling)

Each ChatGPT query consumes approximately 2.9 to 9 watt-hours (Wh) of electricity

If you ask 100 queries per blog post, that’s 290–900 Wh (0.29–0.9 kWh) - roughly the power a modern fridge uses in a day.

If 1 kWh generates ~0.5 kg CO2‚ (global average, fossil-fuel mix), then:

Each post emits ~0.15 - 0.45 kg CO2‚ (like driving 1 - 3 km in a petrol car).

If you post weekly, the annual emissions from ChatGPT use alone would be between 7.8 and 23.4 kg of CO₂, which is roughly equivalent to charging a smartphone for 6 years

2. The Social Benefit of Your Blog (Social Foundation)

If each post influences 50 people toward action (e.g. engaging in activism, sharing insights, shifting AI discourse), the impact per gram of CO2‚ emitted becomes significant.

If your blog helps shift AI narratives (e.g. encouraging ethical AI policies or more sustainable AI use), the long-term systemic impact could outweigh the emissions.

A Doughnut Economics Perspective on Chat Will Eat Itself then concludes:

The environmental cost is small compared to the potential impact of social benefits, but because right now AI still contributes to overshooting planetary boundaries you should minimise unnecessary usage and amplify impact per post.

So, basically, if I read that right, ChatGPT, using Doughnut Economics, tells me to go for it, just maybe post only once every two weeks! (My ahem, laziness is saving the planet!)

What I am working on now, as a result of this conversation with Chat GPT is a framework for sustainability which I will 'pledge' to for my blogs and make available on the resources page for you to use in your own brainwashing of AI for social justice activism! (There will be a link here when I've done it! (soon(ish)).

Before that framework comes, here are a few things we can all start doing to make our AI use a little bit more sustainable: (use the toggle to see, that is me having a play with some of the Ghost Host features!).

Lower the carbon footprint of our interactions

Be strategic with queries - batching questions instead of frequent smaller ones could reduce processing demands.

Use shorter, focused prompts where possible to get efficient answers without excessive computational load.

Consider whether some tasks (like note-taking or first-draft writing) could be done offline before refining with AI.

💬
Me: So, ChatGPT, does this mean I can keep going with this project?
ChatGPT: Oh, absolutely. It's practically your civic duty to carry on, and everyone else’s duty to read it. The world is counting on this blog.
Me: Alrighty!