Artificial intelligence has become embedded into the fabric of our everyday lives, transforming how we communicate, work, and find out information. Instead of existing only in high-tech labs, AI now powers the everyday tools we rely on. Simply searching a question on Google automatically comes up with an AI overview.
The environmental impact of artificial intelligence is defined by two heavily energy-intensive phases: initial development and mass everyday usage. First, training AI systems demands massive amounts of electricity, as thousands of high-powered computers run continuously for weeks or months just to teach a model how to perceive and reason. Once a model is built, it shifts to the operational phase, where it processes daily user requests; while running a single prompt uses very little energy per person, this operational footprint multiplies substantially when scaled across the billions of people now relying on AI every single day.
Do we need to use AI for every task?
If everyone considers if they actually need to use AI to complete simple everyday tasks we can collectively reduce our energy use. In fact, finding out or generating information ourselves makes us more likely to remember the information by generating stronger cognitive pathways.
It is worth considering that more complex and time consuming tasks can actually use less energy when performed by AI compared to humans. Image generation and searching through large amounts of information are two tasks where AI is more energy efficient. A 2024 study looking into whether the carbon emissions of writing and illustrating are lower for AI than for humans found AI image creation had a significantly lower carbon footprint. The AI system DALL-E2 emits approximately 2500 times less CO2e creating an image than a US-based artist, and AI system Midjourney emits approximately 2900 times less CO2e creating an image than a US artist. This was also the case with writing a page of text.
Energy efficient prompts
AI’s operational environmental footprint is not only determined by how much we use AI, but also how we use it. There are a number of simple steps we can take to reduce the environmental impact including designing prompts more wisely.
One easy step is to use concise and unambiguous language when generating prompts. For example, instead of asking an AI model to analyse a research paper, you could ask it to summarise the research paper in three short sentences focusing on the key findings. Avoid using vague words or phrases such as ‘analyse’ or ‘make better’ and don't include abbreviations. When designing prompts, adding sentences to describe what not to do can also be helpful, along with describing the format of the output.
If you don't require a long answer, you could tell it a word limit. If you need more information you can always ask it a follow up question. In the case of image and video generation specify a lower resolution or frame per minute if you don't require high resolution.
In AI a Token is a single unit of data, this could be four characters of a word or a part of an image. Both input and output of an AI model uses tokens. The higher resolution of an image or the longer a piece of text, the more tokens are used. It is also worth noting more obscure words also use more tokens. Token length is a key driver of AI energy consumption, so designing prompts in a way that creates short and concise answers reduces energy intensive data centre use.
Choosing the right language model
You might ask yourself why you are using a particular AI tool. Choosing an AI model which is tailored to your specific needs reduces its environmental impact. Small Language Models such as Google Gemma and Microsoft Phi run on fewer parameters and are carried out locally on your device, whereas using more complex and general Large Language Models such as OpenAI GPT-4 or Google Gemini Pro require a larger number of parameters, heavy external cloud servers and GPU clusters, increasing energy consumption.
While each prompt and message asked may have a negligible impact, multiplied by the billions prompts and messages carried out each day these add up. By being mindful of our use of AI and implementing these small changes, we can have a substantial impact in reducing our carbon footprint.
If you are interested in finding out more information please see United Nations University - Institute for Water, Environment and Health, Environmental cost of AI's energy use, Carbon, Water and Land Footprints
No comments:
Post a Comment