Today (22-Sep-2026) I sent the following email to the principle investigators of the scientific collaborations of which I’m a member in the place I work.
Subject: Use of Claude
Hello [redacted]:
This is a hard letter for me to write.
At the past couple of [redacted] group meetings, the use of Claude has become casually discussed. The subject of whether its use was ethical was never mentioned. I didn’t interrupt the meeting to bring it up because I felt it wasn’t my place.
I’m not going to detail my reservations about its use here either. I assume you can guess what they are.
Obviously, I have no say whatsoever in the analysis tools used by the [redacted] groups. Nor do I wish to have such a say.
However, in an ideal world, I’d like my personal ethics to align with the ethics of the science my name is attached to, even though my name has negligible meaning in the academic world.
I therefore request that, for all future [redacted] papers which include the direct or indirect use of Claude or any other form of commercial large-language model, my name not be included in the author list.
I recognize that this is unimportant and almost certainly won’t make any difference. But sometimes enough small trivial gestures add up to important change.
I chose to omit the reasons for my reservation in my email. However, I have more space in this blog. Here are problems I have Claude, OpenAI, ChatGPT, and similar large-language models (LLMs).
-
Energy consumption of data centers.
An unbelievable number of data centers are being constructed to train and implement the LLMs. The central tech components for these centers are “graphics processing unit” (GPU) cards. Such cards require a lot of power. The increased power consumption has caused energy prices to rise in the areas that these data centers are constructed.
Data centers are often being built in rural areas. The net effect is that the people who live in the rural areas have to deal with higher electric bills for the benefit of LLM users who live in urban areas.
It’s also not clear to me what supplies this increased energy consumption. Is it coming from renewable sources? Or is it adding to global warming and other ecological problems caused by the use of fossil fuels?
-
Water consumption of data centers.
One of the things we learn in physics is that energy, sooner or later, turns into heat. Anyone with a computer knows that it needs some way to cool it down. Systems with high-end GPU cards need more sophisticated cooling systems.
Data centers require an enormous amount of cooling. In practice, this means using water. This has caused water utilities to increase prices as well.
As with the price of electricity, the cost of this can be borne by rural communities that can ill-afford the increase.
-
The cost of GPU cards.
As more data centers are built, they become the major customers of graphics cards. The demand goes up, and therefore so does the price. Anyone who’s purchased a graphics card recently knows that the cost of a recent-model GPU card has increased dramatically.
This doesn’t just affect people who want to build high-end desktops to play Call of Duty. In my area of physics research, there has been increasing interest in using GPU cards for pattern recognition of event topologies within particle detectors. This involves pattern recognition or “machine learning” (ML), which is related mathematically to LLMs but is different functionally and implemented differently.
As the price of GPU cards goes up for the use of handling LLM queries, the GPU cards used for scientific research have become more expensive as well. Research institutions don’t have the budget of major corporations. We’re trading scientific advancement for the sake of making natural-language queries.
-
The cost of memory chips.
Along with the cost increase of GPU cards, memory chips have become more expensive too. This trickles down to any computer, not just those with high-end graphics cards.
In effect, the cost of these data centers is being transferred to the general consumer, and to anyone else (e.g., hospitals) who makes use of equipment with a memory chip inside.
-
The economic bubble being created out of capitalistic greed.
A lot of companies are building data centers when there’s no clear path to profit for them. I don’t pretend to understand the full economics of it, but I gather that businesses are investing trillions and no one is making a real profit off it except for the manufacturer of most of the GPU chips, NVIDIA.
There is a strong possibility of an economic collapse due to over-investment in data centers. They’re being built, causing the environment and economic problems I outline above, with no clear benefit except a rising amount of corporate debt.
Sooner or later that loan will become due. We could see a repeat of the 2008 collapse of the real-estate market. Only the amounts associated with data-center speculation are at least an order of magnitude higher.
-
The ethics of training LLMs.
At their core, the LLMs are being trained by scouring the web, scanning copyrighted works, and also scanning personal information that people have unthinkingly dumped into the servers of big companies.
Why do you think Gmail is free? Because you supplied Google with marketing data. Now those same emails, Google docs, and everything else that seemed free are being fed into the maw of LLM training.
The LLMs are creating responses that are sourced without consent or attribution. In scientific research, this is called “fraud.” When one of my fellow researchers asks an LLM for help in phrasing a paper, how do they know they’re not plagiarizing content from another scientist?
LLMs work by making predictions about the next word in a sentence based on what it’s seen in the masses of words that were scanned to train the model. How many different alternatives would have fed the model to follow the words “pseudoscalar mesons”? If there are relatively few sources, the odds of stealing from someone else’s Google doc is high.
-
Hallucinations.
LLMs create results based on probabilities that a given phrase sounds “reasonable.” That means the results look correct, but doesn’t mean that they are correct.
We’ve already seen many cases where an LLM simply returns a wrong answer. Lawyers have been censured for using LLMs to write legal documents with citations that were structured correctly, and were completely false. LLMs have given incorrect diagnoses for medical conditions.
Scientific research is not immune to this. Scientists are human. We’re just as likely to accept argument from authority as anyone else. Are we just as likely to accept an hallucination as those lawyers were?
Exam question: Look at this blog post. How many instances can you find where I’ve simply accepted the word of others on the basis that they spoke in an authoritative voice?
-
Blind acceptance and an erosion of critical thinking.
I’ve already seen this in my own work. A couple of years ago, a student showed me a piece of code and asked me to review it. I pointed out a problem with it, and asked them to fix it. They couldn’t. It turned out that they’d used ChatGPT to create the code, and had no idea how it worked or how to correct the problem.
Perhaps this objection is in the same category as those made in my youth, when pocket calculators were introduced in the classroom.
However, I feel there’s a fundamental difference. Back then, what students lost was the ability to extract a square root, to use a slide rule, or do linear interpolation from a log table. That’s qualitatively different from having an LLM compose answers for you that you don’t fully understand.
Those are the reasons I’ve taken the stance that I have.
The most likely outcome is that my name will be removed from some future scientific papers. This would have no effect on my career, since I’m long past the point where citations would matter to my non-existent academic reputation.
A potential outcome is that might affect my job.
If behaving ethically were easy, then we wouldn’t need philosophers.
All I can do is stick to what I feel is right, and accept the consequences.
I heard from one of the principle investigators the next day. It turns out that the group had debated the ethics of AI, just not at the group meetings. They acknowledged my concerns, and invited me to attend a separate meeting in which such issues could be discussed. So it turned out better than I hoped, at least in the short term.