I've Automated Boring Jobs for 30 Years. Here's the Anti-AI Argument I Take Seriously.
The carbon math favors AI up to 1,500 to 1. The best objection to that math comes from the study's own authors.
For 30 years, my job has been deleting work. Repetitive work with no added value, the kind people redo by hand because nobody took the time to automate it. Depending on where you stand, that's called freeing people from dumb tasks, or cutting jobs. The debate is as old as the tool, and AI is just its latest episode.
In that debate, I thought I had a solid argument: carbon. A study published in Scientific Reports in 2024 calculates that a page of text produced by AI emits 130 to 1,500 times less CO2e than the same page written by a human. And the human number includes everything: the commute, the house, the vacations.
That's exactly where it gets stuck. When my page disappears, I don't. My commute, my heating and my vacations keep going, and the machine's emissions land on top of them instead of replacing them. The study's authors partly admit it in their own discussion. So the real question is whether that argument wrecks my position as a builder, or only half of it.
30 Years of Killing Boring Work
My bet hasn't moved much in 3 decades: when a task is repetitive and adds nothing, a machine should do it, and the task moves instead of disappearing. The typing goes away, and someone now has to decide what the output means. The critics' version of the same sentence is shorter: automation takes jobs. Both readings can describe the exact same project, depending on whether you sit next to the person who got their afternoon back or the person whose contract wasn't renewed.
AI made that argument louder, because the tool now writes, draws and answers customers, and the electricity to run it shows up on national grid statistics. If I'm going to keep defending what I do, my best evidence and the strongest objection to it belong on the same page.
The Carbon Math: AI vs. the Human
In February 2024, Scientific Reports published a paper by Tomlinson, Black, Patterson and Torrance comparing the footprint of an AI and a human producing the same output. Per page of text, the AI emitted 130 to 1,500 times less CO2e (CO2 equivalent, the standard unit for all greenhouse gases combined). Per image, 310 to 2,900 times less.
The method is what makes the number interesting. The authors take a person's average annual footprint (about 15 tonnes of CO2e for an American, roughly 1.7 kg per hour) and prorate it over the time spent writing. That gives about 1,400 g of CO2e for a page written by an American, against about 2.2 g for a ChatGPT query under 2023 assumptions.
I'd suspected this before reading the paper: the human side isn't a laptop and a desk lamp. It's the person's whole life, sliced by the hour. Car, housing, flights, all of it is already baked into the 1,400 g.
In plain terms, the study makes a narrower claim than "AI is clean": a person's hourly share of a full life weighs far more than a model's share of a data center. For a builder who automates writing-type tasks, that's the strongest piece of evidence I've found.
It comes with limits worth stating right here, not at the bottom. The per-query assumptions date from 2023, and the study only covers writing and illustration. Since then, according to the International Energy Agency (as of September 2026), the energy per AI task has dropped by at least an order of magnitude a year, but heavy uses like video, reasoning and agents consume hundreds to thousands of times more per query than a basic one. So the ratio depends a lot on what you ask the machine to do. Also, one of the co-authors declares holding Nvidia shares. Worth knowing, not disqualifying.
Klarna and the Routine Tier
If you want to see my thesis tested at scale, Klarna is the obvious case.
In February 2024, the company announced that its AI assistant had handled 2.3 million conversations in its first month, 2 out of every 3 customer service chats, with resolution times under 2 minutes. Those are Klarna's own figures.
In May 2025, its CEO told Bloomberg that cost had been too dominant a factor and that quality suffered for it, as CX Dive reported. Klarna started recruiting humans again so a customer could always reach one. (Turns out people with a payment problem would rather not argue with HAL 9000.) The AI still handled about 2 out of 3 inquiries, and the company reported response times 82% faster and 25% fewer repeat issues since launch. CX Dive also points back to the job cuts and the hiring freeze Klarna went through in 2024.
Klarna didn't love the "reversal" headline. In a response reported by Forbes, the company said it was still investing in AI, that its assistant did the work of more than 800 roles (up from the 700 it first announced), that it had never removed human support, and that the human pilot covered 2 new agents. That's a company with a stake in its own story, so I read it as such. As of July 2026, trade coverage from CX Today describes the same blended setup: AI on the high-volume repetitive work, people for the complex or sensitive cases.
That's the pattern I've been betting on. The routine tier goes to the machine, and judgment stays with a human.
The case backs me on the routine tier. But it ends on a CEO's admission and a company correcting the press about itself, and it covers one support queue. The addition problem sits a level higher, where the question isn't who answers the chat.
Where the Critics Land a Real Hit
The objection goes like this. The emissions attached to a person don't vanish when their task does. Replace a human's page with an AI's page, and you don't swap one footprint for another. You add the machine's emissions to the human's, who is still commuting, heating a home and taking vacations.
The study's authors partly concede this in their discussion. They write that the human time freed up by AI could itself generate new environmental costs. They also say their analysis leaves out job displacement, legal questions and rebound effects. The same objection shows up in the reader comments under the paper, which is a reader's comment, not a peer review, but it's the right comment.
That's the hole: deleting a task doesn't delete the person who used to do it. What's still unmeasured is what the workers whose task got automated actually do with their time elsewhere, and what that costs.
For the systemic version of the critique, the most recent full-length reference I know of is IA, le grand enfumage (Payot, May 2026, in French) by Lou Welgryn and Théo Alves Da Costa, who co-lead Data for Good, a nonprofit of tech volunteers that presents itself as a citizen counter-power. The publisher's description frames data centers hungry for energy and water, the job market, military uses and social networks as the logical outcome of concentrated power in tech, not as bugs. It's an activist book and doesn't hide it, but the addition argument stands without it, on the study's own stated limits.
Nadella Invoked Jevons. That Is the Critics' Point.
In late January 2025, after DeepSeek shook US tech stocks, Satya Nadella posted on X that "Jevons paradox strikes again," as NPR reported. His point: as AI gets more efficient and accessible, usage will skyrocket. (The paradox, briefly: when efficiency makes something cheaper, total use can grow enough to cancel the savings. Efficiency behaves like a Gremlin: feed it, and it multiplies.)
When a CEO cites Jevons to calm investors, he's making the critics' case for free.
The mechanism is simple. A gain in efficiency per unit only lowers total impact if volume doesn't grow faster than the gain. And the IEA sees exactly that tension right now. Energy per AI task keeps falling by at least an order of magnitude a year, yet data center electricity demand grew 17% in 2025, and 50% for AI-focused data centers. The capital spending of 5 tech companies now exceeds global investment in oil and gas supply.
Locally, it stops being abstract. In Ireland, data centers went from 5% of metered electricity in 2015 to 22% in 2024 and 23% in 2025, according to the Central Statistics Office. In raw terms, that's 1,238 GWh in 2015 against 6,969 GWh in 2024. Both numbers come with caveats. The CSO counts all data centers, not just AI, and it identifies them by searching for known operator names, sector reports and large consumers, since they aren't a separate category in its source data. And most of that rise predates generative AI. The IEA also notes that some communities now oppose data center projects over electricity bills and environmental concerns.
The other side deserves its weight too. NPR's Planet Money points out that most economists find rebound effects fairly small in modern energy markets, that the extreme version of Jevons may be overstated, and that the financial payoff Nadella was counting on is far from guaranteed. At world scale, the IEA projects data centers at about 3% of electricity demand in 2030. So the global stake is modest, and the local stake is very concrete.
Still, if rebound is the industry's stated plan, the per-page math can't settle this alone. It measures the page and says nothing about how many pages get written once pages cost almost nothing.
My Own Rebound: 3 Branches, Nothing Shipped
I didn't need the IEA to see this one. I've lived it.
Before AI coding tools, every idea I had went through a silent test: is this worth 3 weeks? Most ideas failed it, and the cost did the deciding for me. Once Claude Code pushed the execution cost close to zero, every idea passed the test at the same time. I wrote up how AI broke the filter that used to kill ideas back in March, and that piece already leans on why I traded vibe coding for prompt contracts.
The short version: my accounts dashboard had its core done, then grew 3 open branches around it, and nothing shipped. One dashboard to rule them all, and 3 branches that never came back from Mordor.
That's Jevons at solo-builder scale: cheap execution widened my scope instead of shrinking my workload.
Half My Thesis Survives
On the dull, repetitive work, I'm keeping my position. Klarna has its assistant handling 2 out of 3 conversations and keeps humans reachable for anything that needs judgment. That's what I've watched happen for 30 years: the repetitive task goes to the machine, judgment stays with a person.
On volume, the critics are right. Data center electricity demand grew 17% in 2025, and the IEA projects it from about 485 TWh in 2025 to about 950 TWh in 2030. The data doesn't show that AI is climate-neutral, so I'm not claiming it.
What's still unknown comes down to 2 questions, and no published data settles them yet: whether energy per task will fall faster than the number of tasks grows, and what the people whose task got automated actually do elsewhere.
Until someone measures either one, my dashboard's 3 open branches are the only rebound data I have first-hand.
Sources
- Tomlinson, Black, Patterson, Torrance, The carbon emissions of writing and illustrating are lower for AI than for humans, Scientific Reports 14, 3732 (2024)
- IEA, Key Questions on Energy and AI, executive summary (2026)
- CSO Ireland, Data Centres Metered Electricity Consumption 2025
- CSO Ireland, Data Centres Metered Electricity Consumption 2024, Key Findings
- NPR Planet Money, Why the AI world is suddenly obsessed with Jevons paradox (February 2025)
- CX Dive, Klarna changes its AI tune and again recruits humans for customer service (May 2025)
- Forbes, Klarna Reverses AI Push, Says Customers Prefer Human Support (May 2025)
- CX Today, Klarna's AI Customer Service Roadmap: Why Human Support Still Matters (July 2026)
- ActuaLitté, IA, le grand enfumage, Lou Welgryn and Théo Alves Da Costa (Payot, May 2026)
This post may contain affiliate links. If you click them, I might earn a small commission — costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure.
The carbon argument for AI automation holds up in the data, but only if the person whose job moves actually moves too. The demo-vs-product checklist in the kit helps you measure whether your automation is shipping real impact or just adding emissions to existing ones.