AI vs. Human Energy Efficiency
A Thought Experiment on Generating Product Requirement Documents
These days, conversations about the environmental impact of AI are everywhere, especially when it comes to the massive energy consumption of large language models (LLMs). The general consensus? AI is incredibly resource-intensive. But what if we take a step back and ask: Is this really the whole picture? In this post, I conducted a quick thought experiment to explore whether, in some cases, AI might actually be more energy-efficient than traditional human-driven processes.
The Setup: AI vs. Human
This idea came to me as I was considering the energy implications of generating a product requirement document (PRD). Typically, writing a PRD involves either spending hours of human labour on research or, more recently, generating one through an AI like ChatGPT in a matter of seconds. But which method is actually more energy-efficient when you consider everything, from the food humans need to the energy required for AI computation?
To dig into this, I used ChatGPT to help me with the calculations and even to draft the first version of this blog post. In doing so, I could quickly get the concept out there with minimal environmental output—exactly the point I wanted to illustrate.
But keep in mind, this isn’t about definitive results; it’s about starting a discussion. Here’s how the numbers shake out in this thought experiment.
Energy Breakdown
AI-Generated PRD:
When you ask an AI to create a document, it relies on energy-hungry data centres with vast computational resources. However, those centres are optimized for efficiency.
Compute Energy: The energy needed for a single AI prompt is roughly 250-500 watt-hours (Wh).
Data Center Overhead: Factoring in cooling systems and other infrastructure, the total energy comes to about 375-750 Wh.
Human-Generated PRD:
Now let’s consider the traditional approach: a person who is unfamiliar with the topic researches and writes the document.
Laptop Usage: Over 3-5 hours of research and writing, a typical laptop would consume around 250 Wh.
Food and Caloric Intake: Mental work for this task might consume around 300-500 calories, translating to 350-580 Wh.
Food Production Footprint: Producing the food to fuel this work requires 2.5 to 4.5 kWh (2,500-4,500 Wh) and uses about 300-500 litres of water.
The Verdict: A Quick Comparison
When you add it all up, here’s what you get:
AI-Generated PRD: 375-750 Wh
Human-Generated PRD: 3,100-5,300 Wh
The results suggest that using AI could be 6-14 times more energy-efficient than relying on a human, at least in this specific scenario. Additionally, the water usage involved in food production for human work vastly exceeds what’s needed to cool data centres for a single AI prompt.
But again, these calculations are simplified. They don’t take into account the significant energy used in training LLMs or the full life-cycle environmental impact of the data centres. The point here is not to declare a winner, but to question assumptions and spark a broader conversation.
A Different Perspective on AI’s Environmental Impact
This analysis challenges the common narrative that AI always leads to significant environmental harm. Of course, training large language models does consume vast amounts of energy, but once those models are deployed, they can be incredibly efficient at performing certain tasks. For routine business documents like a PRD, AI could potentially be the greener option.
This post is an invitation to think critically about the nuances of AI’s environmental footprint. We’re often quick to label AI as unsustainable, but perhaps in some cases, it’s not as clear-cut as we think.
Closing Thoughts: Let’s Discuss
My goal with this post is to encourage discussion. These results aren’t definitive, and this thought experiment isn’t meant to be a rigorous study. But if we start questioning broad assumptions and dig into specific use cases, we can have more informed debates about AI’s role in our world.
So, what do you think? Is AI really the environmental villain it’s often made out to be, or are there scenarios where it’s the more efficient choice?