This month I attended a brilliant an insightful Green Web Foundation presentation by Chris Priest, Professor of Sustainability and Computer Systems and Bristol University.

In his presentation he cites a UN report titled The Environmental Cost of AI’s Energy Use – Carbon, Water and Land Footprints. It contains some truly staggering data.

As generative AI usage increases it’s worth knowing what the cost is. I’ve included some of the headline figures relating to model training, usage and data centers below:

Training

• Training models like GPT-5 is estimated to have a carbon footprint of 42,000 tonnes of CO2e, requiring 700,000 tree seedlings (about equal to the number of trees in 40 Central Parks in New York or 155 times the trees in Toronto’s High Park over 10 years to offset).
• The water footprint of GPT-5 training is estimated at 1 billion liters, enough to meet the annual domestic water needs of more than 135,000 people in Sub-Saharan Africa.
• The land footprints of training GPT-4 and GPT-5 are estimated at roughly 0.9 km² (126 football fields) and 1.5 km² (210 football fields), respectively.

Usage

• While training is highly resource-intensive, the continuous inference phase used to generate responses for billions of interactions is estimated to account for 80% to 90% of total AI energy use.
• A typical ChatGPT-style text query is about 200 times more energy-intensive than text classification (such as spam filtering).
• Generating a typical AI image requires 2.9 Wh,making it 60 times more demanding than a short text answer and 1,450 times that of text classification.
• Video generation represents the most energy-intensive frontier with high-resolution long clips on large models drawing more than 415 Wh per clip, meaning a single short AI video can draw as much electricity as 200,000 spam classifications.

Data Centers

• The energy consumed by data centers in 2025 was enough to supply the annual residential electricity needs of the entire population of Sub-Saharan Africa, 1.3 billion people, for 2.6 years.
• Data centers’ electricity use in 2025 had a carbon footprint of 189 million tonnes of CO2e, which would require 3.2 billion tree seedlings grown over 10 years to offset, roughly the total number of trees in the entire United Kingdom.
• The land footprint of 2025 data centers’ electricity demand was 6,900 km², nearly 4.5 times the size of Greater London.
• The physical lifecycle of AI hardware presents a growing crisis. AI infrastructure could generate up to 2.5 million metric tons of e-waste annually by 2030, equivalent to discarding nearly 250 Eiffel Towers every year.

He also shared Hannah Richies’ analysis on How Much Electricity Does AI Consume. The article contains useful stats for task based electricity usage.

Tasks that require reasoning or agentic input use vastly more electricity than simple prompts, but simple prompts from billions of users quickly add up too.

What I found most interesting in Hannah’s piece what is missing from our knowledge at the moment. She estimates that text queries only made up around 2% of AI based electricity usage in 2025.

What is using the other 98%?

We can only speculate because companies aren’t releasing the data. It’s likely that video and image generation is a small to medium sized part of it. We don’t have accurate estimates of energy use per task or scenario here. The bigger piece looks likely to come from thing like:

• AI search summaries.
• Advertising and consumer behaviour analysis
• AI driven algorithms on Social Media and other media platforms
• Content Moderation
• Translation

To me, some of these uses feel more useful to society than others. But until we get transparency from the largest players, we can’t know what uses the most electricity and we can’t make a value judgement on what we use it for.

Accessibility is often an after thought when it comes to website and apps. Making a site as easy to access for the maximum amount of users is the right thing to do but there are some very good reasons to do so as well.

Lucy Collins of Web Usability knows this much better than me. Her view is that sites with high levels of digital accessibility have a competitive advantage. She justifies this stance with her CARE framework:

• Compliance – Inaccessible digital services can constitute a breach of the UK Equality Act
• Altruism – Accessibility empowers independence and shapes how people manage their finances, access healthcare, connect with services, and live their lives, which makes it a basic human right, not a nice-to-have
• Return on Investment – the Click-Away Pound survey found that around 69% of disabled users will simply leave a website they find difficult to use, a loss of £17 billion to the UK economy
• Experience – Accessibility is not just for people who currently identify as disabled. Features designed for people with disabilities frequently end up improving the experience for everyone.

There’s much more information and wisdom in the article, so give it a read.

Accessible sites are more visible to AI crawlers and good for GEO

Generative Engine Optimisation (GEO) is a hot button topic at the moment. Many a “lunch and learn” has been spent talking about how to make your site AI friendly. One effective way might be to make it more your site accessible for all users. This article from Forbes argues that because AI crawlers read websites in a remarkably similar way to screen reading technology.

Things like heading hierarchy, semantic HTML and well structured content all make web pages easier to read for users and bots. So accessible sites are more likely to appear in chatbot output as a result.
If accessibility offers a competitive advantage in discoverability as well as usability what are you waiting for?

Check your site on an accessibility tool like WebAim’s Wave or get in touch with an accessibility agency like Web Usability or As It Should Be and take some action.

One of Kier Starmers’ final actions as Prime Minister was announcing that under 16s will be banned from using social media in the UK from Spring 2027. An exhaustive list of apps and platforms included in the ban is yet to be released but initial reports say Snapchat, TikTok, YouTube, Instagram, Facebook and X will all be included.

As the parent of a pre-teen I’ve been very interested in the implications of the ban and have mixed feelings

On the one side it’s clear that social media can be a harmful and toxic influence on young and old users alike. LGBTQ advocacy group GLAAD details some of the harms in it’s annual social media safety report. It focusses on LGBTQ+ specific indicators but the risks it highlights are reflected across demographics.

This year they report that protections for LGBTQ+ people across Meta platforms, YouTube and X are worse than last year. It’s probably no coincidence that all these firms are based in the US. Only TikTok maintained its 2025 score (an unimpressive 56 out of 100).

So from that point of view, the ban seems to offer a level of protection. Yet others think that a blanket ban will do more harm than good. Professor Sander van der Linden, a Cambridge social psychologist argues that legislating the social media companies to be “safe by design” is the more effective option. “A ban on social media would act as a blunt tool, and would carry many unintended consequences”. Those most vulnerable and alone will suffer disproportionately from being able to access community and connection than those with stronger support networks.

If there is one thing for sure, it’s that young people everywhere need safe spaces to explore and express themselves. In post austerity and post COVID UK, there is a lack of community spaces and that without them, the social media ban risks exacerbating the situation. The government has pledged £132.5m to fund activities offline in our communities but it’s unclear if this will meet the needs of young people across the country.

Finally it’s worth remembering what this ban means for the general population and not just young people. The Online Safety Act requires everyone in the UK to prove that they are old enough to access restricted platforms. But this article points out that there are many risks associated with identity tracking as standard. The Electronic Frontier Foundation argues that there is no solution that doesn’t have serious privacy, accuracy, or security problems.

This policy has real implications for the open web and it’s something I’ll be keeping an eye on, both as a parent and a curious digital professional.

Don’t forget, we want to hear more from you, the Curiously Green community. If you’re heard or read something that may be of interest, please share any links, and your thoughts with us.

Even better, we’d love to know what you’re working on. If you have any case studies or projects you’d like to share, or new approaches you’ve tried that may be of interest, this is a great way to share with like minded folk so please head over to our submission form and tell us all about it.

We can’t wait to hear from more of you. 💚

Use the Digital Declutter toolkit to help you measure and reduce your digital emissions.

Test your web pages with Website Carbon.