The Carbon Footprint of AI: From Micro Tasks to Macro Trends, and What SMEs Need to Know

As businesses embrace AI tools like ChatGPT, there’s growing concern about the carbon impact of the proliferation of AI technologies. At Seedling, it’s a topic we’re asked more and more about, and rightly so!

Much like the technology itself, understanding the carbon impact of AI is a complex task. And whilst AI applications are becoming commonplace, development of the carbon data that can help us understand how these technologies fit in with the challenge of climate change is lagging behind.

In this article, we’ll run through a pragmatic, data-led, and nuanced overview of the current state of the carbon impact of AI through four key topics:

Humans vs. AI: A comparison of emissions for a simple task (a 1000 word article)
The big picture: Opportunities and the rebound effect
A call to action: What we need from AI leaders
Need to know for SMEs: How to approach conscious use of AI

1. Micro-Level Carbon Comparison: A 1000-word article

When we think of the complexity of AI, and the vast infrastructure that sits behind it, it's easy to assume that use of AI will inherently be much more carbon intensive than a human completing an equivalent task. But this isn't necessarily the case. To illustrate this, let's take a very common AI use-case: content creation. How much carbon is emitted when ChatGPT writes a 1000-word article, versus a human? We'll build this from the ground up using current best data.

A caveat before we start: I'll rely on the best data (most peer reviewed and referenced) available just now (see appendix for full details), but it's worth noting that there's a serious lack of carbon data published by AI developers like Open AI. We'll discuss this fact, and it's consequences, later on.

Carbon impact of a 1000-word article: ChatGPT versus human

AI-written article (GPT-3 / 4o) - 20 grams of CO₂e

A 2024 study by Hao et al. estimates that a single short GPT-4o query uses about 0.43 watt-hours (Wh) of electricity. Using an average grid carbon intensity for the US published by the EPA - 0.35 grams CO₂e / Wh - this translates to:

0.43 Wh × 0.35 grams CO₂e = ~0.15 grams CO₂e per query

So: ~0.15 grams of CO₂ per query for inference (the term used to describe the actual usage of ChatGPT for a query).

However, we also need to take into account the amortised carbon impact of training AI. A widely referenced study estimates that GPT-3 consumed 502 metric tons of CO₂e during training (Patterson et al., 2021). Assuming that ChatGPT may do a full re-training of the model once per month, and 10,000,000 queries per day (methodology advocated by Tomlinson et al., 2024), we get to 1.84 grams CO2e per query, and a total of ~2 grams CO₂e across both training and inference. This chimes with other estimates in the 2-3 grams CO₂e per query range.

Finally, we need to account for the fact that ChatGPT never gets an article right on the first query! Let's assume that there are revisions, adding up to10 queries in total as feedback is given to ChatGPT.

So, x10 queries at 2 grams of CO₂e each: 20 grams of CO₂e in total.

Human-written article - 196 grams of CO₂e

Now consider a human writing a 1000 word article on a typical laptop plus desktop set-up drawing ~140W. For the research and writing, let's assume it takes 4 hours, so:

140 W × 4 hr = 560 Wh
560 Wh × 0.35 grams CO₂e = ~196 grams CO₂e

That’s almost 10 times the combined total for ChatGPT including training. This also excludes the indirect emissions - like food, heating, or office energy. Heating alone is typically the single largest impact for desk-based workers (especially when homeworking), so the real impact of the human-written article could be considerably more than x10 that of AI, depending on the context of the worker.

Some caveats:

We've used a US average emissions factor for the sake of comparison, but it's worth nothing that where data centres are located - and the underlying energy they use - matters massively.
Some workstation energy use also needs to be attributed to the footprint of ChatGPT.

So, what's the point? The idea here is not to claim to calculate the carbon impact of a AI with a high degree of accuracy. As I'll discuss shortly, data is extremely limited, and impact likely varies significantly by model.

Instead, what we're illustrating is that it's not a clear-cut win for a human versus AI when it comes to carbon, as is often suggested. The efficiency of AI matters - AI has the potential to be lower carbon at conducting an equivalent task.

However, as we’ll explore next, the broader environmental impact of AI depends heavily on what we choose to do with that efficiency saving.

2. Macro-Level Impact and the Rebound Effect: Scale is the Key Issue

The true environmental impact of AI doesn’t lie in a single simple taks like writing an article - it lies in how we make use of this technology, and how newfound efficiency can proliferate what we as a society create.

This is the rebound effect: when a tool becomes more efficient and accessible, we tend to use it more - and total energy demand increases. Analysis by MIT highlights this as the central environmental concern for the environmental impact of AI. Even if per-task emissions are low, overall emissions can rise sharply as usage becomes embedded across workflows. In other words:

“The inference phase is becoming the dominant driver of emissions. And that phase doesn’t just scale linearly - it scales with how deeply these models are embedded into daily work.”

‍

This same article suggests that the increased uptake of AI is already visible, with the power requirements of North American data centres doubled in a single year, from 2,688 megawatts in 2022 to 5,341 megawatts in 2023, driven in part by generative AI. Globally, data centers consumed 460 terawatt-hours of electricity in 2022 - enough to make them the 11th-largest electricity consumer in the world, just ahead of Saudi Arabia. By 2026, data centres are projected to take 5th place, between Japan and Russia.

While not all of this is due to generative AI, the technology is a major driver of rising demand. And supply is struggling to keep pace. As another deep dive by MIT notes, “the demand for new data centers cannot be met in a sustainable way.” According to MIT postdoc Noman Bashir:

“The pace at which companies are building new data centers means the bulk of the electricity to power them must come from fossil fuel-based power plants.”

In short: whilst at the micro-level each interaction can appear efficient, the sum total is significant, and growing fast. AI doesn’t just replace human work. It creates new types of content and demand: marketing variants, synthetic media, algorithmic experimentation. Without action, the infrastructure required to support this growth could outpace the decarbonisation of the grid - locking in higher emissions as AI becomes more deeply embedded in business and daily life. On top of that are the other resources required (minerals, water) to keep pace with energy demands, renewable or otherwise.

To be clear, AI also has enormous potential to help solve the climate crisis - from accelerating materials discovery and carbon capture technologies to improving grid optimisation. But we can’t let that potential distract from the impact of how AI is used and deployed today.

3. What We Need from AI Leaders – A Call to Action

Most AI users currently have little to no visibility into the environmental cost of their AI usage. While Sam Altman laments the financial cost of our Ps and Qs, the corresponding carbon impact doesn’t get a mention.

Unlike cloud providers like AWS which offer usage and emissions dashboards, AI services rarely show even basic usage data - let alone the carbon impact per call or the footprint of model training.

This makes informed, low-carbon choices almost impossible. And for carbon accountants like me, the total absence of data is becoming a frustrating and growing problem. Right now, AI usage gets lumped with other types of software - and then only if it’s a paid-for service. This isn’t good enough.

I’m not the first to point this out. In particular, Alexandra Luccioni and the team at Hugging Face - an AI company that, among other things, builds open tools and infrastructure to make machine learning more transparent and accessible - have proposed an open, reproducible framework for tracking model emissions. Luccioni et al.’s approach accounts for the energy mix behind training and inference, hardware efficiency, and training duration.

So, here’s what we’re calling on AI providers to do:

Standardised, full-lifecycle carbon disclosure: As Luccioni et al.’s work shows, energy consumption - and therefore carbon - varies hugely across models. We need AI leaders to publish this data using a consistent, peer-reviewed methodology to enable comparison, drive accountability, and reward efficiency.
User carbon analytics: Right now, users of tools like ChatGPT don’t even see how many queries they’ve run - let alone their emissions. That needs to change. AI providers should give users real-time access to their usage and estimated carbon impact, embedded directly in dashboards and APIs — as visible as billing or latency.
Smarter defaults: Many platforms default to the largest, most resource-intensive models - even for simple tasks. Providers should help users choose lighter-weight options where appropriate, and clarify the energy and carbon trade-offs involved. That guidance should be integrated directly into prompts, settings, or model selectors.
Invest in clean infrastructure: As AI scales, so should investment in low-carbon compute. That means expanding renewable energy procurement, locating data centers in cleaner grid regions, improving hardware efficiency, and publishing transparent Scope 2 and 3 emissions.

4. Implications for SMEs: Smart Use of AI to Reduce (Not Raise) Emissions

What does all this mean for small and mid-sized businesses? It can feel like you’re being swept up in a macro AI storm, without much control. If even basic carbon data isn’t available, how are you supposed to make better decisions?

There are two answers to this.

First, AI can help uncover insights and drive efficiencies that reduce emissions. Think route optimization, better forecasting to avoid waste, or smarter inventory management to cut overproduction. At Seedling, for example, we use AI to simplify emissions categorisation and reporting, helping clients save time getting the accurate carbon data to take action.

Second, it’s about ensuring you and your team recognise that AI usage has a cost. If used deliberately and efficiently in everyday tasks, AI can reduce emissions - as in the article use case illustrated above. The rebound effect - an explosion of content and queries, not the underlying tech - is the real threat. The old rule of efficiency and carbon going hand-in-hand applies for AI too: wasteful use will drive up emissions.

So what can SMEs do?

Focus on decarbonising emissions-intensive processes. As a general rule, better efficiency tends to mean lower carbon.
Be mindful of scale and frequency. Not every task needs to be automated. Use AI in a thoughtful, outcome-focused way.
Choose transparent, efficient tools - once providers finally begin surfacing the data needed to make those decisions.

With smart use, AI can lower your footprint, not expand it, but intentionality matters. Sustainable business, one that is commercially competitive but good for the planet, doesn’t exclude AI - it includes using it strategically and with carbon in mind.