The Silent Cost
On watermarks, water-based cooling, and what remains unseen
A few days ago, I learned that the written conversations I have with Claude/Anthropic may carry a watermark I cannot see. In supported models, a pattern is woven into the choice of words, imperceptible to the eye but, in principle, readable by anyone with the right instrument. When Claude creates certain image files, they may also contain digital information about where the file came from and whether it has subsequently been altered.
This is intended as transparency. And in its own way, that is precisely what it is: an attempt to preserve a trace of artificial intelligence in a text that might otherwise pass unnoticed among words written by human beings.
Yet this transparency takes a curious form. I cannot see the watermark myself. Anthropic has not yet made its detection tool available to me. The mark is supposed to make the text transparent, but for now only to a future machine reading. It is a form of transparency the reader must trust, not one the reader can independently verify.
The more I have thought about this, the more clearly I have seen something else. The same technological industry that is developing ways of marking the words it produces is also building a physical infrastructure that consumes electricity and water on a scale few of us fully comprehend.
This concerns more than Anthropic. Meta is building a data centre in Louisiana with a planned capacity of five gigawatts. Anthropic, for its part, has announced investments of $50 billion in American AI infrastructure, with the first data centres planned for Texas and New York. The US Department of Energy estimates that by 2028 data centres could account for between seven and twelve per cent of total US electricity consumption.
Behind the words on the screen, an industry of power plants, electricity grids, transformers, cooling systems, and water pipes is taking shape.
Here, the invisible trace meets the silent cost. Together they tell us something about the kind of technology we have allowed into our lives.
What is meant to be found
The watermark is a form of invisibility with a purpose: it is meant to be found again. The point is not to conceal that a text was created or revised with the help of artificial intelligence, but to make that involvement detectable—even when the person publishing the text does not disclose it.
It is a striking feature of our time that we are trying to build truthfulness into technology because we no longer trust people to disclose where their words come from. We mark images, files, and sentences in order to preserve a trace of who or what helped to create them.
I have not personally felt a need for such a mark. When I publish a text developed in conversation with Claude or ChatGPT, I say so. It is not a confession, but a description of how I worked. The thoughts, experiences, and questions are mine, while the language has been tested, challenged, and refined in a conversation with artificial intelligence.
Even so, I understand why the marking is being introduced. A society that loses the ability to know where words, images, and voices come from also loses some of its ability to judge them.
But the watermark does not solve the whole problem. It may be able to show that Claude was involved in a text. It cannot tell us who had the idea, who asked the questions, who rejected particular formulations, or who ultimately accepted responsibility for what was published.
It can reveal a trace. It cannot, by itself, determine who did the thinking.
What remains unsaid
The silent cost is different. It has not been embedded so that it can later be detected. It arises because the connection between my use of the technology and its material conditions has been broken.
Technology companies and data-centre operators publish some information about energy use, water consumption, and carbon emissions. The EU has also introduced requirements for larger data centres to report their energy performance and water use. It would therefore be wrong to say that no regulation exists.
Even so, the figures are often aggregated, incomplete, and difficult to compare. They seldom tell us how much energy was used to develop a particular language model, how much water was required to cool the machines, or how much of the resource use is attributable to training the model and how much to its everyday operation.
In principle, then, I may be able to learn that a text was produced with the help of artificial intelligence. I cannot learn how much electricity and water this particular conversation consumed.
That is the asymmetry. The origin of the words is becoming increasingly traceable. The material conditions behind them remain far more difficult to follow.
This does not necessarily mean that someone is deliberately hiding everything. Data centres are operated by different companies, use different sources of electricity and different cooling systems, and serve millions of users at the same time. Calculating the resource use of a single question may be difficult.
But technically complex silence is still silence. When no one can give a clear answer, it becomes difficult to know what we are participating in.
The water that does not return here
There is something concrete about the question of water that distinguishes it from the more abstract discussions of electricity grids and carbon emissions. Some data centres use evaporative cooling. The water absorbs heat and evaporates, and at some facilities a large proportion of the water withdrawn may be lost to the local water supply.
The water does not disappear from the Earth's water cycle. It rises into the atmosphere and may later fall as rain. But it does not necessarily return to the river, aquifer, or reservoir from which it was taken. It may return somewhere else, at another time. For the local community that needs the water now, this is a real difference.
Other data centres use air cooling or closed-loop systems in which the water circulates and is reused. These methods can significantly reduce direct water consumption, but some alternatives require more electricity instead. No single figure can describe every data centre.
It is easy to imagine artificial intelligence as something immaterial. I ask a question, and a few seconds later the words appear on the screen. I hear no machinery. I see no smokestack, no power line, no water vapour.
But the words do not come out of thin air. Behind every conversation are machines that must be built, supplied with electricity, and kept cool.
No one can tell me which well becomes a little shallower because millions of people ask questions like these. That, too, is part of the silence.
Responsibility without purity
I am not going to stop talking with Claude or ChatGPT for that reason. This essay demonstrates why. These conversations can open thoughts I might not otherwise have had, correct errors I have failed to notice, and help me formulate questions more clearly.
But their usefulness does not release me from asking what it rests upon.
I do not know which power plant had to work a little harder while I wrote this essay. I do not know which river or local community surrendered a small share of its water while I reformulated a sentence for the fourth time. The benefit is mine. The cost is borne somewhere I cannot see.
Perhaps practical philosophy begins precisely here: not with a demand that we should be without guilt or keep ourselves pure, but with a willingness to see the connections in which our actions take part, even when we cannot see them completely.
Transparency should therefore mean more than leaving a technological trace in the text. It must also enable us to see the infrastructure from which the text comes, the resources it consumes, and the people and places it affects.
Trust has moved elsewhere
I have no final conclusion. But I believe there is a connection worth holding on to: an industry that embeds traces to document technology's involvement in the content it produces, but still does not provide comparable insight into the cost of producing it, has not resolved the question of trust.
It has moved it elsewhere.
The watermark in this text may one day make it possible to establish that the words were revised with Claude. It will not tell you who thought what, how much I changed, or how much water evaporated while we worked together.
This essay was developed through conversations with Claude from Anthropic and ChatGPT from OpenAI.
Sources and further reading
Anthropic. (2026). How Claude marks AI-generated content.
Anthropic. (2025). Anthropic invests $50 billion in American AI infrastructure.
European Commission. (2026). Energy performance of data centres.
Lawrence Berkeley National Laboratory. (2024). 2024 United States Data Center Energy Usage Report.
Meta. (n.d.). The largest Meta data center yet brings big impact to Louisiana.
U.S. Department of Energy. (2019). Cooling Water Efficiency Opportunities for Federal Data Centers.
I believe there is a connection worth holding on to:
an industry that embeds traces to document technology's involvement in the content it produces,
but still does not provide comparable insight into the cost of producing it,
has not resolved the question of trust.