Humans vs A.I. - Carbon Footprint

Generative AI uses energy and pumps out CO2 yes, but so do humans.

Training

The average person creates 5 metric tonnes of CO2 per year.

Training OpenAI's GPT-4 released 300 tonnes of CO2.

So generative AI clocks in at 60 times more carbon intensive than humans, right?

No.

Humans take far longer to train than a generative AI model. 

If you've used ChatGPT you'll know how eclectic its responses can be. To educate a human to a similar level would take at least 20 years. 

• AI - 300 tonnes of CO2; 
• Humans: 100 tonnes of CO2.

Inference/Operation

After training, the AI model is ready for 'use'.

NVIDIA estimates that 80-90% of the energy cost of neural networks lies in ongoing inference.

From estimates of power usage, K. Ludvigsen estimated ChatGPT's CO2 emissions at between 10 and 207 tons in January 2023, assuming a Swedish carbon intensity of 9g / kWh.

So ChatGPT inference emits 186 tonnes of CO2 per month at maximum estimates.

On the face of it, that blows any human carbon footprint out of the water.

However, we need to consider ChatGPT's capacity. In August 2023, ChatGPT (chat.openai.com) handled 1.43 billion visits from 180.5 million unique visitors.

Let's compare that to the closest human query-answer role: call centre operator.

Average call handle times vary by industry, but most callers would consider 10 minutes or less acceptable.

A call center operator working a solid month of twenty 37.5 hour days would handle 4500 ten-minute calls in that month.

Equating call-center voice calls with website visits we get:

• Humans: 4500 calls/416 kg CO2, or 11 calls/kg CO2.
• AI: 1.43 billion visits/186 tonnes CO2 or 7700 visits/kg CO2.

The AI would make up its excess carbon generated during training in 1.1 million calls/visits. In human time that's 20 years, but at ChatGPT capacity that's half a day.

Simplistic answer:

It produces more CO2 (but less time) to train AI than it does to raise people. After that, though, AI scales way harder.

Complex answer:

In reality, humans and machines are inseparable. Humans are heavily involved in training AI, screening both input and output. Of course, raising and educating humans also requires vast amounts of time and resources. 

We haven't counted the impact of making the hardware or feeding the engineers. Then again, on the human side, we don't tally the teachers or the discarded baby car seats.

When the AI model is unleashed, it still needs human prompts. On the other hand, call center operators are heavily supported by technology such as VoIP and their terminals. Neither humans nor their tech are fully independent.

Looking forward, AI has greater potential for carbon efficiency, much of which involves greening energy sources. Humans can't be fed directly from solar panels, respire CO2, and eat things that respire CO2. That cannot be changed without changing human nature.

However, given how intertwined humans are with their machines, making technology more efficient will also make humans more efficient.

When hearing about how disastrous new developments could be, it's important to compare them with how disastrous things are now.