Byte-ing the Planet
Olga Kostiuk, S10, takes a look at the environmental impact of artificial intelligence to ask "is it worth it?"
Olga Kostiuk
Senior 10
Photography by Ruby Hellier
Art Director, 2026
So far, the 21st century has been defined by scientific breakthroughs - AI, quantum computers, genetic editing techniques, “reusable rockets”, the development of “attribution science” and many more. They all come with undeniable benefits: AI and quantum computers allow faster and more advanced research, increasing the possibilities of further breakthroughs; genetic editing helps with combatting genetic diseases, saving lives; “reusable rockets” revolutionised space exploration by making it cheaper, and “attribution science” recognises climate change as a consequence of human actions, creating ways we can act to solve it.
However far away some of these benefits may seem to you, the harms of these technologies – AI in particular - are closer than you might think. In 2015, Stephen Hawking gave a memorable speech, in which he called AI
the best or the worst thing ever.
Overall, he believed that AI was capable of doing more good than harm with proper management. Eleven years later, the world sees Al management in the hands of greedy companies doing more harm than good.
DeepLearning and artificial intelligence first emerged during the early 2010s as a tool for researchers to analyse large sets of data in a shorter time and with greater accuracy than human data scientists can. The first models were not perfect - there was always a trade-off between speed, accuracy and the volume of data analysed. Over time, the Large Language Models have become better at balancing all three “efficiency factors”, and there are a number of scientific discoveries that can be attributed to AI, such as the aforementioned genetic editing techniques.
As AI got better and better, there came the natural question of “what now?”, and AI progressed from being a ‘lab tool’ to a ‘global’ mainstream one, becoming the fastest growing industry in the world.
Part 1:
The environmental costs
In 2022, OpenAI released a LLM that redefined mainstream generative AI: ChatGPT. For the first time, the efficiency of the software put a strain on our environmental resources. LLMs the size of those like ChatGPT, Gemini, Copilot and others “live” in massive server farms. Their high-performance chips (GPUs) require massive amounts of energy to run, and cooling systems evaporate litres of freshwater to keep the servers running cool and prevent hardware failure.
There are two main parts of an AI model that use up resources: the training and the inference. On balance, roughly 80-90% of AI’s total energy footprint comes from the “inference” - the users having conversations with AI compared to the 10-20% that the initial training requires. To put this into perspective, the training of OpenAI’s GPT-4 model alone used “50 gigawatt- hours of energy, enough to power San Francisco for three days”, according to MIT’s Technology Review. Every query uses 10 to 15 times more electricity than a Google search, and every conversation (a series of 20-30 queries) “drinks” hundreds of millilitres of water by evaporating it in the cooling towers. Multiply this by hundreds of millions of daily users worldwide and you get a global resource crisis. So much so, that as of 20th January 2026, the UN have announced the world going from a freshwater water crisis to
a state of global water bankruptcy.
This declaration of ‘water bankruptcy’ signifies a final point of no return, not just a temporary shortage. If before many people argued that, in time, AI would be able to generate a cooling system for itself that doesn’t require water and move towards a utopian model that doesn’t use unreasonable amounts of resources, now we are already at the point of resource of bankruptcy and we do not have the capacity or the time to wait for AI to solve all of these issues - the crisis is here and now and the utopian resource-efficient AI model still remains but optimistic hope for a distant future.
What is worse, is that tech giants keep going on using our resources and trying to keep it secret: Apple announced plans to spend $500 billion on manufacturing data centres in the US over the next 4 years, while Microsoft covered their latest development in Wisconsin with NDAs and labelling its energy and land use a “trade secret”.
This is another hidden cost land mass. Server farms are huge and require a lot of land to be built on - 200 acres minimum. These have to be flat, level ravines within proximity to water and energy supplies.
Sadly, this land comes from deforestation.
Previously, we destroyed the tropical rainforests for farmland for cattle. Now, we do it for AI data centres.
Part 2:
It is not worth it:
AI is integrated into our lives beyond chatbots - it is everywhere: in apps that we use every day to organise our time, do our shopping, book flights and such. The main negative impact of this over-reliance is “cognitive atrophy” - a gradual weakening of skills through disuse. “digital amnesia” was first defined as the decline in memory, due to reliance on search engines.
With AI now serving as a search engine that synthesises information for us, saving us having to read through websites and articles and do the critical thinking ourselves, neuroscientists fear the amplification of this effect with this dependency. This is because the human brain is neuroplastic - it strengthens connections that are used and weakens those that aren’t. AI’s frequent use disables the critical analysis and memory pathways, and we end up trading the very skills that got us into this age of progress for a few seconds of convenience.
There is also the issue of ethical use of AI; when we use it for just about anything and everything: is a recipe worth it, when you can use 10 times fewer resources on a simple Google search?
Part 3:
So what?
AI should stay in labs, confined to research institutes, where it can actually do more good than harm. Mainstream AI is not worth it - it harms, the planet, society, and the AI models themselves, and is more beneficial to the companies that own it than anyone else.
When training large LLMs the size of ChatGPT, data sets are often enormous and uncontrolled - there is no human “sorting” of data, simply because of how much data there is, causing a lot more machine bias and hallucinations than there should be.
To add insult to injury, the more we train mainstream generative AI, the more it degrades; the data sets used get flooded with AI generated data, creating a negative feedback loop and leading to “Model Collapse” - a point beyond which the training of a model doesn’t better it, but does the opposite.
If we don’t bridle AI back to having a clear scientific purpose in labs, we will succumb to it. With the current resource consumption, we will reach a global “day 0”, before AI has the opportunity to develop “a will of its own” and enslave humanity.
So, yes, the threat of an AI takeover is very real, but it will take over our resources before it takes over our governments and the human race as a whole. Today’s world isn’t an age of scientific enlightenment; it is an age of dystopian crises incurred by the very “scientific enlightenment” that was there to solve them.