Real Bullshit vs Mathematical Bullshit

Let us begin with the serious question, delivered with appropriate gravity.

Is AI bullshit more environmentally destructive than actual bull shit?

This is, unambiguously, a question that needed to be asked. And the answer is both reassuring and deeply unsettling, depending on which part you focus on.

First, the honest answer

If you mean overall environmental damage today — actual measurable greenhouse gas emissions, right now, on this planet — then no. Bullshit from AI is not currently more eco-destructive than bullshit from bulls.

Livestock has a much larger established greenhouse-gas footprint. Cattle produce methane. The wider livestock system produces substantial methane, nitrous oxide and CO₂. It has been doing so for a very long time, at enormous scale, and the numbers are not comfortable.

AI's environmental cost comes primarily from the electricity and infrastructure required to train and run models. Data-centre electricity demand is growing rapidly — particularly because of AI — but the global footprint is still smaller than livestock's.

So: real bullshit currently has the bigger environmental footprint.

Mathematical bullshit has the more rapidly scaling one.

But here is the really uncomfortable question

Value per emission.

A cow produces greenhouse gases while producing food. Something tangible. Something a human can eat. The cow participates in a system that we have, for better or worse, decided has a genuine output worth having.

An AI can produce greenhouse gases while producing something nobody asked for, nobody needs, or nobody can verify.

We already ask agriculture to justify its environmental cost by what it produces. We have entire industries, certification schemes and policy frameworks dedicated to that question. Nobody lets a farm emit freely without some expectation of output.

We have not yet started asking the same of AI.

Bullshit from bulls is a climate problem. Bullshit from AI could become a climate problem without even producing anything worth having.

That is the editorial. Not that cows are greener than AI. But that the standard we apply to agriculture — justify your emissions — has not yet been applied to the other system currently scaling fastest.

And now, the maths

Since we are already in this conversation, we might as well go the rest of the way.

What if we took the metaphor completely literally? What if every AI response was physically equivalent to one real, cow-sized pile of fresh manure?

OpenAI reported more than 2.5 billion ChatGPT messages per day in July 2025. A recent measurement of cattle found roughly 11–16 kg of manure per animal per day. Let us be generous and use 15 kg per response.

Now. Earth's land surface is approximately 149 million km². To cover all of it in a single centimetre of AI-generated physical bullshit, you would need approximately 1.49 trillion cubic metres of material. At roughly one tonne per cubic metre, that is about 1.49 quadrillion kilograms.

If you want a 10 cm layer: roughly 1,090 years. A full metre: approximately 10,900 years.

Which sounds almost reassuring, until you remember that the 2.5-billion-message figure is already old relative to today's AI expansion. The IEA says global data-centre electricity consumption was about 415 TWh in 2024 and is projected to reach around 945 TWh by 2030, with AI as a major driver.

The thought experiment is not really: how long would AI take to cover the Earth in bullshit?

It is: how quickly are we approaching an industrial system capable of producing bullshit at planetary scale?

The Editorial Reality Check: Britain, specifically

But here is where the thought experiment gets genuinely useful. Rather than using 15 kg as a physical mass equivalence, let us use the actual carbon footprint of real manure as our unit of comparison.

A cow's daily enteric fermentation and waste management cycle produces approximately 10 kg of CO₂e. A single high-quality LLM inference — one ChatGPT-4 level response — produces roughly 0.00008 kg of CO₂e. That means to match the carbon footprint of one cow-day of real bullshit, you need approximately 125,000 AI responses.

Now scale that globally. In Q1 2026, research puts daily global LLM interactions somewhere between 1.5 and 3 billion prompts across all providers. Using a midpoint of 2.25 billion:

For context, the global livestock sector produces approximately 19.45 million tonnes of CO₂e every single day. That means the entire global AI inference industry currently produces roughly one hundred times less carbon than agriculture does. Every day.

So the cows are still winning. Comfortably.

But now the UK calculation. If we return to our physical metaphor — 15 kg of manure per response, because we are committed to this — the UK's land area of 242,509 km² requires approximately 2.425 billion tonnes to cover in a 1 cm layer. Global AI at 2.25 billion responses per day produces 33.75 million tonnes of equivalent material:

About ten weeks. To bury Britain beneath a centimetre of it. A 10 cm layer: roughly 1.8 years. A full metre: roughly 18 years.

At which point you would struggle to locate the country under it, which feels like an appropriate metaphor for several things currently happening.

And unlike cow shit, mathematical bullshit doesn't need to be physically transported to spread. It doesn't decompose. It doesn't fertilise the land. A cow's manure eventually breaks down and feeds the soil. Mathematical bullshit gets indexed, cached, summarised, fed back into the machine as training data, and used to generate more of itself.

We aren't just covering the UK in waste. We are building a digital crust that never goes away.

Real bullshit covers fields. Mathematical bullshit covers the information environment. And the information environment has no edges.

The verdict

Real bullshit: larger current environmental footprint. Bounded by the number of cows. Regulated, more or less. Expected to produce food in exchange for its emissions.

Mathematical bullshit: smaller current environmental footprint. Scaling rapidly. Not yet subject to a value-per-emission standard. And unlike its agricultural predecessor, it can reproduce itself.

The headline writes itself:

Which is more destructive — the shit from bulls, or the bullshit from AI? One covers fields. The other covers the information environment. And we have been much better at regulating the one that smells.

We made agriculture justify its emissions. We built frameworks, targets and policy around it. We continue to argue about whether it does enough.

We have not yet asked AI to answer the same question. Not: how much do you emit? But: what are you producing that is worth it?

That question is coming.

Probably in about 65 days.