AI’s Carbon Footprint: Are We Chasing the Wrong Metric?

Its a Thursday evening in the City. Office lights on full across near empty floors, the HVAC hums, and a junior lawyer scrolls through Clause 17.1 yet again on a chunky Dell.

Meanwhile, in a data‑centre less than a few miles away, a large language model could finish that review in seconds, sipping less power than the lawyer’s desk lamp. Headlines still slam “AI’s energy hunger”, but the bigger carbon bill is racking up right there in the office.

Focusing in on a single, everyday legal task and the numbers flip: inference emissions are tiny compared with the carbon pouring out of buildings, laptops and, yes, people simply breathing. Below is a like‑for‑like look at one focused review (using typical 2024 UK grid data and an industry‑average data‑centre PUE of 1.58)

Scenario 1 — Fully Automated AI

A local 7 B model on a small box would be lower again, drifting into milligrams.

Scenario 2 — Hybrid Lawyer, Efficient Laptop

Scenario 3 — Office‑Based Lawyer, Standard Laptop

The AI’s share, even with a state‑of‑the‑art model, is trivial. The big swing factors are where the lawyer sits and how the building is run, importantly I'm not suggesting lawyers hold their breath, but do you want to be the person adding 79g of carbon to the process if not needed?

Why This Matters

Focusing only on the silicon misses the real culprits: buildings, bodies, and clunky hardware. If the goal is to cut carbon, three moves pay off faster than squeezing inference cycles:

1. Automate the low‑complexity stuff. Full automation chops emissions to fractions of a gram.
2. Ship better kit. An ultra‑efficient laptop halves a user’s device footprint across the year.
3. Rethink location. Hybrid arrangements can massively reduce office overheads by hundreds of grams an hour and save on commuting too.

Transparent tagging makes those trade‑offs obvious, with a deliverable explicitly labelled:

• Automated AI review: 0.06 g CO₂e
• Human‑assisted review, office worker: > 80 g CO₂e

Now it’s clear where the impact lies, move the work, modernise the hardware, rethink the space and start questioning what else we’ve missed.

Beyond CO₂: The Environmental Cost We’re Not Tracking

Most legal ESG efforts stop at carbon, but AI’s footprint stretches further. Water usage is a major one. Training and running large models, especially in traditional data centres, requires intensive cooling, often using hundreds of thousands of litres of water. That doesn't show up on a carbon dashboard, but it should.

Then there’s the physical kit. Bigger models drive demand for more powerful GPUs, shorter upgrade cycles, and more e-waste. It’s easy to overlook in legal tech, where the tools feel virtual. But the hardware stack behind those tools has a real cost like rare-earth minerals, emissions-heavy manufacturing, and disposal logistics that most ESG reporting ignores.

If you’re claiming a shift to AI is sustainable, it’s worth asking: are you just moving the problem?

Imperfect Numbers, Better Questions

Yes, grid intensity swings hourly. Data centre PUEs vary. We haven’t even touched embodied emissions or water usage in depth. The numbers will never be perfect — but waiting for perfect is how nothing changes.

Publish the best estimates you have. Invite scrutiny. Refine the model. Repeat.

Carbon is just one angle, but are we designing systems that make the most of AI, while using the least of everything else?

Progress beats paralysis every time.

Sources

• Knight Frank (2023) on hybrid vs office energy splits
• Notebookcheck power‑draw tests for MacBook Air M2 and Dell Latitude 5420
• Epoch AI estimate for GPT‑4o inference energy
• Uptime Institute 2023 survey on average data‑centre PUE
• Carbon Brief: 2024 UK grid intensity figures
• ScienceDirect studies on average human CO₂ exhalation